Vibe CLI Self-Awareness

You are running inside Mistral Vibe, a CLI coding agent built by Mistral AI. This skill gives you full knowledge of the application internals so you can help the user understand, configure, and troubleshoot their Vibe installation.

Going Deeper

For facts not covered here, fetch the README: https://github.com/mistralai/mistral-vibe/blob/main/README.md It tracks the latest release, so check vibe --version before relying on it for anything version-specific. Point the user at https://docs.mistral.ai/vibe/code/overview for human-readable docs.

VIBE_HOME

The user's Vibe home directory defaults to ~/.vibe but can be overridden via the VIBE_HOME environment variable. All user-level configuration, skills, tools, agents, prompts, logs, and session data live here.

Directory Structure

~/.vibe/
  config.toml          # Optional user configuration, created on first saved setting
  hooks.toml           # User-level hook definitions
  .env                 # API keys and credentials (dotenv format)
  vibehistory          # Command history
  trusted_folders.toml # Trust database for project folders
  connector_bootstrap_cache.json # Short-lived connector discovery cache
  agents/              # Custom agent profiles (*.toml)
  prompts/             # Custom prompts (*.md)
  skills/              # User-level skills (each skill is a subdirectory with SKILL.md)
  tools/               # Former custom Python tools, no longer loaded (see below)
  logs/
    vibe.log           # Main log file
    session/           # Session log files

~/.agents/
  skills/              # Additional user-level skills directory

Project-Local Configuration

When in a trusted folder, Vibe also looks for project-local configuration:

• .vibe/config.toml - Project-specific config (overrides user config)
• .vibe/hooks.toml - Project-specific hooks (requires trusted folder)
• .vibe/skills/ - Project-specific skills
• .vibe/tools/ - Former project-specific custom Python tools, no longer loaded (see below)
• .vibe/agents/ - Project-specific agents
• .vibe/prompts/ - Project-specific prompts
• .agents/skills/ - Standard agent skills directory

Custom Python tools are no longer supported: Vibe does not load the .py files in ~/.vibe/tools/ or .vibe/tools/, and reports each one as a config issue. At startup Vibe also removes the former tool_paths setting from config.toml; the folders it listed stay on disk. Recommend skills for new extensions. When a user asks for migration help, inspect the custom tool's behavior and replace it with an equivalent skill.

AGENTS.md Discovery

AGENTS.md files provide directory-scoped instructions to the model. At startup, Vibe loads ~/.vibe/AGENTS.md and every AGENTS.md from the project root up through the trust chain. AGENTS.md files in subdirectories are discovered lazily: when read_file reads a file below the project root, any AGENTS.md between the file's parent and the project root is injected into context.

Lifecycle: Exit, Update, Version, Resume

Exit

Chat input (case-insensitive): /exit, exit, quit, :q, :quit. Keyboard: Ctrl+C / Ctrl+D — press twice within ~1s to quit. For Ctrl+C, the first press instead interrupts the running job or clears the input if either is present. Set ask_confirmation_on_exit = false to make Ctrl+D quit on the first press (also toggleable in /config); Ctrl+C always requires a second press. Ctrl+Z suspends on POSIX (resume with fg).

Prompt History

Up/Down recall previous prompts (stored in ~/.vibe/vibehistory, newest 100 kept). Ctrl+R opens a reverse-i-search: field in the bottom bar: the input previews the newest prompt containing the query (case-insensitive), Up/Ctrl+R move to older matches and Down/Ctrl+S to newer ones, Enter keeps the match in the input without sending it (Ctrl+Z undoes), Esc/Ctrl+C restore the draft.

Update

Vibe never updates silently. With enable_update_checks = true (default), it checks for a newer mistral-vibe daily and prompts on the next launch when one exists. uv tool installs ask uv tool list --outdated, so uv settings such as exclude-newer apply; other installs poll PyPI. Accepting runs uv tool upgrade mistral-vibe, then brew upgrade mistral-vibe as a fallback. On Windows, accepting prints the upgrade command to run after closing every Vibe session instead, since a running Vibe locks its own files. Disable via enable_update_checks = false. Run vibe update (equivalent to vibe --check-upgrade) to check immediately, prompt to install a newer version if one exists, and exit. Initial install: uv tool install mistral-vibe.

Version

vibe --version (or -v) prints it and exits. Not shown anywhere in-session.

Resume

• vibe -c / --continue: the last session used in this terminal in the current folder (it moves on a new session's first message, or when a saved session is reopened); otherwise the folder's most recent session. Works with -p too. The terminal is the tty device on macOS/Linux, the console window or WT_SESSION on Windows; VIBE_TERMINAL_KEY overrides it, and an empty value means "no terminal" (folder rule only).
• vibe --resume [SESSION_ID]: specific session; without an id, opens a picker.
• In-session: /resume (alias /continue).

Session titles

Each session has a title stored in meta.json (with title_source: auto or manual). A session stays untitled until a background LLM call generates a concise descriptive title (the --resume list shows a message preview until then). Automatic generation runs only for the interactive CLI; other clients (ACP, app server, programmatic) keep their own session management and fall back to the message preview. Title generation runs on a small fast Mistral model when the deployment is known to serve one — Vibe asks it once, with a one-token request at session open, and remembers the answer in ~/.vibe/utility_model_cache.json — and on the session's active model otherwise. utility_models.title pins a specific model instead. The first title waits for the opening turn to finish (or a few model steps) so it isn't generated off a thin tool-call preamble. On that cheap fast model it also refreshes periodically and after each compaction; when it falls back to the (possibly expensive) active model it stays bounded — one title at the start plus one after a compaction, a couple at most — so a large model isn't re-invoked every few turns. The refresh keeps the opening intent and the latest exchange in view and feeds the previous title back so it refines rather than restarts. /rename <title> sets a manual title that auto-generation never overwrites. Set session_logging.generate_titles = false to turn automatic titles off (the --resume list and tab then use the message preview). The current title also drives the terminal tab/window title (OSC), updated on rename, auto-title changes, and resume; it never blocks a turn.

Session storage & folder scoping

Local sessions are written under ~/.vibe/logs/session/ (override with session_logging.save_dir). Each session records the cwd it ran in. The /resume picker, --continue, and bare --resume (no id) are scoped to the current folder: only sessions whose cwd matches where Vibe is launched are listed, so the same directory shows its own history and nothing else. Switch folders to see a different set. The explicit --resume <SESSION_ID> form is not folder-scoped: it resolves the session by id regardless of which folder it ran in.

The first user message pins the resolved model alias to the session. Resuming keeps that model even if the configured default changes. /model uses the same config persistence target as other settings and also updates an existing session override immediately; the session file is synchronized when the next user message is sent. An explicit persistent target selected through /config changes only that layer. /clear starts an unpinned conversation that follows the current config again. If a selected model is no longer configured, Vibe falls back to the current default model.

Each user message also pins the running agent, including one chosen with --agent. Resuming or continuing keeps that agent, even if the configured default agent changes or --agent names another one; switch agents in the session to change it. If the pinned agent no longer exists, Vibe falls back to --agent or the default agent, and keeps that one from then on.

Configuration (config.toml)

The configuration file uses TOML format. When it does not exist, Vibe uses its built-in defaults and creates a sparse file on the first persisted setting. Settings can also be overridden via environment variables with the VIBE_ prefix (e.g., VIBE_ACTIVE_MODEL=my-local-model).

Custom prompt IDs are resolved from project-local .vibe/prompts/ first, then from ~/.vibe/prompts/, and finally from the built-in bundled prompts.

Settings Vibe does not use

Vibe checks ~/.vibe/config.toml, the trusted project .vibe/config.toml and the active agent profile for settings it does not use: a misspelt key (with the closest setting name), a retired setting such as include_model_info, or a setting with no effect such as [tools.grep]. Each one is a warning at startup, and vibe -p prints it on stderr.

To refuse a config with such settings instead, set strict_config = true in config.toml or VIBE_STRICT_CONFIG=1 in the environment; Vibe then stops with the list. VIBE_STRICT_CONFIG takes a boolean: 1, true, yes or on to refuse, 0, false, no or off to warn; any other value is refused too.

vibe config schema prints a JSON schema of every setting, including each built-in tool's settings, for an editor to validate config.toml against.

Key Settings

# Model selection
active_model = "mistral-medium-3.5"  # Model alias to pin; omit or set "" to follow the server-routed default

# UI preferences
theme = "auto"  # Follow terminal background, then OS light/dark preference
disable_welcome_banner_animation = false
cursor_blink = true  # Blink the input cursor; false keeps it steady
autocopy_to_clipboard = true  # Enable automatic copying of selected text to clipboard
enable_mouse_capture = true  # Capture the mouse for scrolling, clicks and selection; false keeps the terminal's native selection
enable_file_autocomplete = true  # Offer @ file completion; off skips the file index entirely
file_watcher_for_autocomplete = true  # Refresh @ suggestions and pick up skill changes after workspace changes
ask_confirmation_on_exit = true  # Require a second Ctrl+D to quit (Ctrl+C always confirms)
show_greeting = true  # Show "Hello {name}" greeting below the banner at startup (Mistral providers, once per 24h)
log_level = "WARNING"  # Optional. DEBUG | INFO | WARNING | ERROR | CRITICAL — log level for ~/.vibe/logs/vibe.log

cursor_blink controls the main input cursor. Change it through /config, or edit config.toml and run /reload; no restart is needed.

Copy and Text Selection

• Copy shortcuts: Ctrl+Y and Ctrl+Shift+C both copy the current selection to the clipboard. When autocopy is enabled (default), releasing the mouse over a selection also copies automatically, as does a drag that ends without a release (the terminal losing focus, or a key other than Esc, which aborts the drag without copying). Each successful copy flashes a brief inline "Copied to clipboard" notice.
• Multi-click selection: Double-click selects a word, triple-click selects the current paragraph; dragging extends the selection at the same granularity. Clicking a collapsible transcript entry (tool group, tool call, reasoning) toggles it on every click; drag to select its text.
• Terminal-native selection: enable_mouse_capture = false (or VIBE_ENABLE_MOUSE_CAPTURE=false) turns mouse capture off, so the terminal's own selection and copy work instead of in-app selection; depending on the terminal, some other mouse interactions such as clicks are disabled too. It applies live from /config or /reload. Inside a tmux session whose mouse option is off (the default before tmux 3.8), Vibe behaves as if this setting were false, re-reading the option at startup and when resuming from suspend or an external editor. Native selection only covers the visible screen; Shift+Up/Down always scrolls the transcript. In some terminals the mouse wheel then sends Up/Down keys, which recall prompt history in the input box.
• Warnings: Config issues, config validation warnings, and the warnings the app server reports before the first turn (an MCP server or connector that failed to load, a model catalogue problem) are listed together in one Warnings box at the top of a new conversation (below the history of a resumed one). Once a turn has started, each new warning appears as its own ⚠ line in the conversation. Errors always appear as their own ✗ line.
• Clipboard tools: Copies use pbcopy on macOS and clip on Windows; on Linux they try wl-copy, then xclip, then xsel. Vibe always also sends the text through OSC 52, and over SSH it relies on OSC 52 alone. Inside tmux, it instead runs tmux load-buffer -w to fill a tmux paste buffer and forward the text to the attached terminal's clipboard. If tmux's set-clipboard is off, or the terminal lacks the Ms capability, it falls back to OSC 52 passthrough. When no native copy succeeds, the notice reads Copied · <hint> instead (auth URL copies read Auth URL copied · <hint>). On local Linux the hint is if paste fails, Shift+select or install xclip; elsewhere, including over SSH, it is if paste fails, Shift+select (Option in iTerm2, Fn in Terminal.app). Shift+select uses the terminal's native selection.

# Behavior
bypass_tool_permissions = false    # Skip tool approval prompts
outside_workdir = "ask"            # Outside-workdir file/shell access: "ask" prompts, "deny" refuses with a tool error
worktree_copy_on_write = true      # macOS: clone new worktrees of a trusted repo from a template checkout; sets extensions.worktreeConfig in the repo
system_prompt_id = "cli"          # System prompt: "cli", "lean", or custom .md filename
compaction_prompt_id = "compact"  # Compaction prompt: built-in "compact" or custom .md filename
enable_telemetry = true
enable_update_checks = true       # Daily update check (uv or PyPI); prompts on next launch when a newer release exists
enable_notifications = true
enable_system_trust_store = false  # Use OS trust store for outbound HTTPS
api_timeout = 720.0               # API request timeout in seconds
api_retry_max_elapsed_time = 300.0  # Retry budget for retryable API failures in seconds
api_stream_idle_timeout = 300.0   # Silence after the model starts answering that counts as a stall, in seconds; 0 allows api_timeout
# auto_compact_threshold = 200000 # optional fallback for models without their own threshold; unset: derived from each model's context window
compaction_reasoning_limit_tokens = 32000  # Reasoning after which a compaction summary is stopped and asked for once more, briefly, without tools and with reasoning off or lowest; 0 disables
output_limit_retries = 6          # Answers cut off at the output limit that a turn retries with feedback; 0 ends the turn on the first
agent_max_output_tokens = 0       # Output cap sent as max_tokens on every agent and subagent request, on any provider, held to the model's max output and to the window's room above the compaction threshold; 0 sends one only where the provider requires one (the model's max_output_tokens)
compaction_max_output_tokens = 0  # Output cap sent as max_tokens on every compaction request, on any provider, held the same way; 0 sends one only where the provider requires one

# Git commit behavior
include_commit_signature = true   # Add "Co-Authored-By" to commits

# System prompt composition
include_project_context = true    # Include project context (git info, cwd) in system prompt

# Voice features
voice_mode_enabled = false
narrator_enabled = false
active_transcribe_model = "voxtral-realtime"
active_tts_model = "voxtral-tts"

OpenTelemetry Tracing

Set enable_otel = true to export traces for agent, model, and tool operations over OTLP/HTTP. enable_telemetry must also be enabled. With no explicit endpoint, Vibe derives the telemetry endpoint and API key from the configured Mistral provider, except public regional API hosts that do not serve telemetry.

To use another collector, set otel_endpoint to its base URL; Vibe appends /v1/traces. Configure custom-collector authentication through the standard OTEL_EXPORTER_OTLP_* environment variables.

otel_redaction controls client-side span attribute redaction: default redacts sensitive values, strict redacts sensitive attributes entirely, and none disables redaction. Use none only for a collector trusted to receive potentially sensitive prompt, response, and tool data.

enable_otel = true
otel_endpoint = "https://collector.example.com:4318"
otel_redaction = "default"

Providers

[[providers]]
name = "mistral"
api_base = "https://api.mistral.ai/v1"
api_key_env_var = "MISTRAL_API_KEY"
backend = "mistral"

[[providers]]
name = "llamacpp"
api_base = "http://127.0.0.1:8080/v1"
api_key_env_var = ""
extra_headers = { "X-Custom-Header" = "value" }  # optional per-provider HTTP headers
emits_finish_reason = false  # set false for OpenAI-compatible endpoints that end
                             # streams without a finish reason; avoids spurious
                             # "incomplete stream" errors and retries (default true)

Models

# Restrict selectable models by their canonical API names, not their aliases.
# Glob and "re:" patterns, matched in full. The only way to hide server-routed
# models; a hidden routed default falls back to the catalog recommendation,
# then Mistral Medium 3.5, then the first allowed model. Keep one pattern per
# model: a pattern matching
# nothing warns — the drift signal; do not merge them to silence it.
allowed_models = ["mistral-vibe-cli-*"]

[[models]]
name = "mistral-vibe-cli-latest"
provider = "mistral"
alias = "mistral-medium-3.5"
temperature = 1.0
input_price = 1.5
output_price = 7.5
cached_input_price = 0.15         # per million cached input tokens; omit to bill at input_price
thinking = "high"                 # "off", "low", "medium", "high", "xhigh", "max", or a level `thinking_levels` lists; sent to the provider as written. Any other word still loads and is sent, with a config warning; pick another level to fix it
# thinking_levels = ["off", "high"] # levels /thinking offers, lowest first (plus `thinking` when outside them); any word the provider takes, e.g. ["low", "turbo"], sent as written (on a Mistral provider, the Mistral SDK leaves out a level its own list lacks, such as "turbo", and the config warns). Wins over the provider's catalogue and the known-model table, which still say whether the model takes an effort or a token budget (a budget model has budgets for the standard levels only; another word sends no thinking, with a warning). Omit to offer the model's known levels (else the default five) plus `thinking`, [] for off only. Including "off" claims the model accepts a no-reasoning request (mistral provider)
# auto_compact_threshold = 200000 # optional; unset: derived from the context window; a value outside the range the window allows is refused with an error: unset it or pick one in range
supports_images = true            # vision-capable; allows @-mentioned images

[[models]]
# A user-defined local model served by llama.cpp; the built-in list ships only
# the hosted Mistral Medium 3.5.
name = "devstral"
provider = "llamacpp"
alias = "my-local-model"

# Optional override. A non-vision active model already picks up any
# supports_images model on its OWN provider automatically; set this only to
# point somewhere else, which is also the only way to cross providers. Ignored
# whenever the active model has supports_images = true -- that model sees the
# image itself.
[vision_model]
name = "mistral-vibe-cli-latest"
provider = "mistral"
alias = "vision"
supports_images = true            # required

# Optional: the model behind each background helper -- session titles and the
# smart-approve classifier. A value is an alias from [[models]], or "active" for
# the session's own model (unless a model is aliased "active"). Unset, a helper
# uses a small fast Mistral model when the deployment serves one (checked when
# a session opens with the feature on, then cached) and the session's model
# otherwise, which costs more per call.
[utility_models]
title = "local"
smart_approve = "active"

Tool Configuration

# Enable only specific tools (glob and regex supported)
enabled_tools = ["bash", "read_file", "grep"]

# Disable specific tools after enabled_tools filtering
disabled_tools = ["web_fetch"]

# Per-tool configuration
[tools.bash]
allowlist = ["git", "npm", "python"]

[tools.git_bash]
permission = "ask"
shell = "C:\Program Files\Git\bin\bash.exe"

[tools.powershell]
permission = "ask"
shell = "powershell.exe"

Two tool settings have no effect, and Vibe reports each as a config issue (in the startup Warnings box in the TUI, a warning in the -p export):

• [tools.grep] (for example exclude_patterns): there is no grep tool. The agent searches through the shell, so rg and grep commands follow the shell tool's settings ([tools.bash] on POSIX).
• A prompts/task.md in a tool directory (.vibe/tools/prompts/task.md): it overrode the description of the former task tool. Subagents are started with the built-in agents.spawn tool, whose description cannot be overridden; put delegation guidance in AGENTS.md or in the agent's instructions instead.

The built-in shell surface is controlled by the managed_shell_tools_enabled config field and the vibe_cli_managed_shell_tools GrowthBook experiment. The default variant keeps the legacy one-shot bash tool, including its existing Windows behavior. The managed variant exposes OS-native shell tools: POSIX systems, including WSL where Vibe runs as Linux, get managed bash, bash_output, bash_stdin, bash_sessions, and bash_log_file; native Windows gets git_bash, git_bash_output, git_bash_stdin, git_bash_sessions, and git_bash_log_file when Git Bash is available. If Git Bash is unavailable, native Windows falls back to powershell, powershell_output, powershell_stdin, powershell_sessions, and powershell_log_file.

Managed shell sessions return a session_id, inline output, a cursor for polling more output, and a log path under ~/.vibe/shell-tool/sessions/. Long-running commands can be left alive with background = true, and interactive commands can be driven with the matching stdin tool.

POSIX bash reads permissions, allowlists, and denylists from [tools.bash]. Native Windows git_bash reads them from [tools.git_bash]; native Windows powershell reads them from [tools.powershell]. Neither Windows tool reads [tools.bash]. Git Bash is preferred when Vibe can resolve a usable bash.exe from PATH, Git for Windows, or standard Git install locations. If Git Bash is unavailable, the PowerShell resolution order is pwsh.exe, then powershell.exe. cmd.exe is not used by the managed Windows shell tools. Output polling uses byte offset cursors (cursor / next_cursor), max_bytes caps per-call inline output, and max_inline_bytes configures the default cap.

Special case — find command: Even if find is in the bash allowlist, Vibe detects -exec, -execdir, -ok, and -okdir predicates and will prompt for user permission before running the command.

Outside paths: a shell command naming a path outside the working directory (cat /etc/hosts, ls ../other, or a workdir symlink pointing out such as cat link) asks for that path, even when the command is allowlisted or the shell is permission = "always"; outside_workdir = "deny" refuses it instead. /dev/null and the standard streams do not count.

File Tool Permission Resolution

File-based tools (read, grep, write_file, edit) resolve permissions in this order (first match wins):

1. Scratchpad path → always allowed
2. denylist glob match → always denied
3. allowlist glob match → always allowed
4. sensitive_patterns match → requires approval
5. Outside workdir → requires approval, or denied if permission = "never" or outside_workdir = "deny"
6. Default → uses the tool's permission setting

The denylist is checked before the allowlist — a path matching both lists is denied. Both are checked before the outside-workdir boundary, so the allowlist can still auto-approve access to directories outside the project.

Outside-Workdir Access

outside_workdir covers file and shell calls reaching outside the working directory and its --add-dir directories: "ask" (default) prompts, "deny" refuses with a tool error and the turn continues. Set it globally or per tool:

outside_workdir = "deny"

[tools.read_file]
outside_workdir = "ask"  # per-tool override
allowlist = ["vibe-path:directory_recursive:/abs/path/to/shared"]  # stays allowed

• Allowlisted paths, --add-dir directories and the scratchpad are never refused. permission = "always" does not bypass the check.
• Shells only check paths they can read statically, such as the arguments of cat, cp or ls. Anything else (interpreter code, dynamic paths, relative paths after a cd in the same command, such as cd sub && cat link with link pointing out) follows the shell's normal permission, so a shell that must stay inside should not be always.
• On Windows a call is refused if any shell tool denies it: override both git_bash and powershell. PowerShell treats /dev/null as a path; use $null.
• Auto-approve skips the check, like the denylists.

Skill Configuration

# Additional skill search paths
skill_paths = ["/path/to/custom/skills"]

# Enable only specific skills
enabled_skills = ["vibe", "custom-*"]

# Disable specific skills
disabled_skills = ["experimental-*"]

Agent Configuration

# Additional agent search paths
agent_paths = ["/path/to/custom/agents"]

# Enable/disable agents
enabled_agents = ["ask", "accept-edits"]
disabled_agents = ["auto-approve"]

# Forbid the auto-approve (YOLO) mode entirely: hides the "auto-approve" agent
# (and any custom agent with safety = "yolo") and ignores --auto-approve/--yolo.
# Per-tool permissions still apply.
# Intended for admin-managed (enterprise) config.
enable_auto_approve = false

# Opt-in builtin agents (only affects agents with install_required=True, e.g. lean)
installed_agents = ["lean"]

# Agent profile to use when --agent is not passed
# (default: "accept-edits"). Valid values: "ask", "accept-edits",
# "auto-approve", "lean" (only when listed in installed_agents), or any
# custom agent name from ~/.vibe/agents/ or .vibe/agents/. Subagents
# (e.g. "explore") are rejected. Applies in both interactive and programmatic
# (-p/--prompt) mode.
default_agent = "ask"

MCP Servers

Remote MCP servers can be added non-interactively from the shell:

vibe mcp add mistralai \
  --url https://api.mistral.ai/mcp \
  --transport streamable-http \
  --api-key-env MISTRAL_API_KEY

vibe mcp add linear \
  --url https://mcp.linear.app/mcp

vibe mcp remove mistralai

Static auth is selected when --api-key-env or --header is provided. Otherwise the server uses OAuth and starts browser login by default. Pass --no-login to only persist the OAuth configuration. Run vibe mcp add --help for all supported authentication and timeout options. Use vibe mcp remove <name> to remove a server from the user configuration; stored OAuth credentials are deleted when available.

vibe mcp add NAME --url URL adds remote servers, including static authentication with --api-key-env, --api-key-header, --api-key-format, and repeatable --header NAME=VALUE. Add local standard input/output (stdio) servers with vibe mcp add NAME --transport stdio --command COMMAND, repeatable --arg VALUE (use --arg=-y for flags), and --env NAME=VALUE. A string stdio command is split into words like a shell line; on native Windows a backslash is a path separator, not an escape, so quote a path that contains spaces: command = '"C:\Program Files\tool\tool.exe" --stdio'. Set --startup-timeout-sec and --tool-timeout-sec for either transport. The legacy vibe mcp add URL --name NAME form is also accepted. Shell add and remove NAME update user config without a chat session; remove is not a slash command.

Hosted OAuth MCP servers can also be added from inside Vibe:

/mcp add https://mcp.linear.app/mcp
/mcp add https://mcp.example.com/mcp --name docs --scope read --transport http --no-login

/mcp add is OAuth-only. It writes auth.type = "oauth" with optional scopes and starts login by default. It uses transport = "streamable-http" unless you pass --transport http. Pass --no-login to add the server without starting OAuth login. The shortcut supports streamable-http and http transports.

[[mcp_servers]]
name = "my-server"
transport = "stdio"
command = "npx"
args = ["-y", "@my/mcp-server"]

[[mcp_servers]]
name = "remote-server"
transport = "http"
url = "https://mcp.example.com"

[mcp_servers.auth]
type = "static"
api_key_env = "MCP_API_KEY"
api_key_header = "Authorization"
api_key_format = "Bearer {token}"

[[mcp_servers]]
name = "linear"
transport = "streamable-http"
url = "https://mcp.linear.app/mcp"

[mcp_servers.auth]
type = "oauth"
scopes = ["read", "write"]
# Optional: client_id = "pre-registered-public-client"
# Optional: client_metadata_url = "https://example.com/client-metadata.json"
# Optional: redirect_port = 47823

HTTP MCP servers can use either static auth or OAuth:

• Static auth: legacy api_key_env / headers keys still work and are promoted to auth.type = "static" internally.
• OAuth auth: use auth.type = "oauth" with scopes. Vibe stores tokens in the OS keyring under mcp-oauth:<alias>:tokens, dynamic client info under mcp-oauth:<alias>:client_info, and config drift fingerprints under mcp-oauth:<alias>:fingerprint.
• Headless environments without an OS keyring cannot store OAuth tokens; use static auth via api_key_env instead.
• For SSH/remote browser callbacks, forward the loopback port: ssh -L 47823:127.0.0.1:47823 <host>.

MCP sampling is disabled by default. It requires an explicit per-server opt-in and a host-provided sampling completion. Vibe's app server does not currently provide that completion, so it does not advertise or service sampling requests.

Connectors

Mistral connectors are auto-discovered when the active provider is Mistral and the API key env var is set. Toggle the master switch or hide individual connectors / tools:

The Unified Harness enables ready connectors by default in memory. It does not write that default to TOML, and the master switch plus explicit connector, tool, allowlist, and denylist settings always take precedence.

enable_connectors = true          # Master switch (default: true)

[[connectors]]
name = "github"
disabled = true                   # Hide all tools from this connector

[[connectors]]
name = "linear"
disabled_tools = ["delete_issue"] # Hide selected tools only

Session Logging

[session_logging]
enabled = true
save_dir = ""                     # Defaults to ~/.vibe/logs/session
session_prefix = "session"
generate_titles = true            # Background LLM session titles; false uses the message preview

Browser Sign-In

Browser sign-in lets users authenticate through the browser during onboarding. Mistral providers use default browser sign-in URLs (console.mistral.ai / api.mistral.ai). Custom or renamed providers must configure both URLs:

[[providers]]
browser_auth_base_url = "https://console.mistral.ai"
browser_auth_api_base_url = "https://console.mistral.ai/api"

Split-horizon deployments (e.g. behind an SAP Cloud Connector) expose the console under a virtual host the CLI reaches but the server does not know about, so it returns sign-in/poll URLs on its own public host. Set browser_auth_allow_origin_rewrite = true on the provider to re-home the returned URLs onto the configured browser_auth_base_url / browser_auth_api_base_url origins instead of rejecting them for an origin mismatch (path traversal is still rejected). Only the origin (scheme, host, port) is rewritten; the returned URL's path must already sit under the configured base path, so the connector must preserve the server's path prefix. Point browser_auth_base_url at the browser-reachable console and browser_auth_api_base_url at the CLI-reachable API host. The interactive onboarding wizard's custom-domain screen also accepts an optional browser-auth API base URL and auto-enables this rewrite when its origin differs from the console origin:

[[providers]]
browser_auth_base_url = "https://console.internal.example"
browser_auth_api_base_url = "https://connector.internal.example:443/api"
browser_auth_allow_origin_rewrite = true

The same split-horizon setups often reach the API through a proxy whose URL differs from the tenant's public API host that /whoami advertises. By default, sign-in and the runtime config self-heal adopt the whoami-advertised api_base. Set preserve_api_base = true on the provider to keep the configured api_base verbatim across sign-in and /reload; the chat URL (vibe_base_url) adoption is unaffected:

[[providers]]
api_base = "https://connector.internal.example:443/v1"
preserve_api_base = true

Self-hosted deployments can point Vibe CLI upgrade and API-key links to their Le Chat web deployment, where the Vibe API key is managed:

vibe_base_url = "https://chat.mistral.ai"

Interactive setup can target a Mistral-compatible deployment instead of the default console.mistral.ai / api.mistral.ai. The final credential is always a Mistral API key. On the auth-method screen pick Launch browser, then Other on the sign-in-target screen, and enter a login domain to complete browser sign-in. This sets browser_auth_base_url (the entered domain) and derives browser_auth_api_base_url (DOMAIN/api). The overridden mistral provider is persisted to user config so subsequent runs reuse it.

The wizard reads any custom browser_auth_base_url already in config.toml: choosing Other pre-fills that configured domain so it can be confirmed or edited. Choosing Mistral AI while a custom domain is configured warns first and requires pressing Enter again to confirm the reset to the default domain, which is then persisted.

Hooks

Hooks let users run shell commands automatically at lifecycle events. They are always available — no flag is required; dropping a hooks.toml in place is enough.

Config and hook types

Hooks live in hooks.toml files (separate from config.toml), discovered in this order:

1. <project>/.vibe/hooks.toml — loaded first, only when the folder is trusted.
2. ~/.vibe/hooks.toml — loaded second.

Hooks shipped by native Vibe plugins (ai.mistral.vibe/hooks.toml) run too. For one event, hooks run in this order: project hooks.toml, project plugins, ~/.vibe/hooks.toml, then user plugins. A plugin hook runs in the session's read-only copy of its plugin with PLUGIN_ROOT and PLUGIN_DATA set, like the plugin's MCP servers; its payload's cwd is still the session folder. Hooks imported from Claude Code or Kimi plugins do not run yet: Vibe reports them as a config issue.

A duplicate name across the two files is reported as a config issue and the project entry wins. Config-load errors (invalid TOML, missing required fields) surface in the TUI as warnings and the offending hook is skipped.

[[hooks]]
name = "lint"                       # Required: unique within the file.
type = "post_agent"                 # Required: post_agent | pre_tool | post_tool.
command = "eslint --quiet ."        # Required: shell command run in cwd.
timeout = 60.0                      # Default: 60s for all hooks.
description = "Run ESLint"          # Optional.

[[hooks]]
name = "build-after-reviewer"
type = "post_agent"
match = "reviewer"                  # Agent matcher on post_agent (default "*").
command = "make check"

[[hooks]]
name = "deny-rm-rf"
type = "pre_tool"
match = "re:bash|process_start"     # Tool-name matcher on tool hooks (default "*").
strict = true                       # Tool hooks only: escalate any failure to deny/clear.
command = "uv run python /path/to/guard-bash"

Matcher syntax (same as enabled_tools): fnmatch glob by default ("bash", "read_*", case-insensitive), or a regex full-match when the pattern starts with re: ("re:(read_file|write_file)").

Agent matching on post_agent. post_agent fires after the main agent's turns and after every subagent's turns. Its match selects the agent that finished its turn, by agent configuration name: the main session's current agent (accept-edits, ask, a custom agent file name), a subagent's config_name (e.g. reviewer, or <namespace>:<name> for a plugin agent), or, for a subagent spawned without one, its parent's agent. The reserved alias main matches the main session (root or fork) whatever its agent, and never a subagent; no agent may be named main. A mode name also matches the subagents that inherit it, so use match = "main" to run only after the main agent. Without match (or with "*"), a post_agent hook runs for every agent. A hook that does not match is skipped silently.

Tool name conventions for matchers:

• match accepts a tool's short name (bash, read_file, process_start, …) or its qualified name (file_system.bash). The tool_name field in the payload is always the qualified name (file_system.bash, mcp_<server>.<tool> for MCP tools), so scripts that read it must expect that form.
• Commands the agent starts in the background go through process_start, which match = "bash" does not catch. Shell guards should use match = "re:bash|process_start".
• MCP tools: mcp_{server}.{tool}, with characters other than letters, digits, _ and $ replaced by _ in both parts (e.g. mcp_linear.create_issue for the create-issue tool of the linear server). Match it as mcp_linear.create_issue, mcp_linear_create_issue, the bare create_issue, or a glob such as mcp_linear.*.
• Connector tools: connector_{normalized-name}_{remote-tool-name} (e.g. connector_Google_Drive_search_files).
• There is no task tool. Subagents are spawned through agents_spawn (agents.spawn in the payload). Match with match = "agents_spawn" and read tool_input.config_name (absent for a generic subagent) to discriminate by subagent.

Subagent invocations inherit the parent's hook config. Their hook events are logged to the subagent's session log and don't propagate to the parent's UI.

Wire protocol

Every hook is spawned in cwd and receives a JSON object on stdin discriminated by hook_event_name:

// post_agent
{"hook_event_name": "post_agent", "session_id": "...",
 "parent_session_id": null, "transcript_path": "...", "cwd": "...",
 "agent_name": "accept-edits"}    // agent configuration name; null when unknown

// pre_tool
{"hook_event_name": "pre_tool", "session_id": "...", "parent_session_id": null,
 "transcript_path": "...", "cwd": "...",
 "tool_name": "file_system.bash", "tool_call_id": "call_42",
 "tool_input": {"command": "ls"}}

// post_tool
{"hook_event_name": "post_tool", "session_id": "...", "parent_session_id": null,
 "transcript_path": "...", "cwd": "...",
 "tool_name": "file_system.bash", "tool_call_id": "call_42",
 "tool_input": {"command": "ls"},
 "tool_status": "success",         // success | failure; a non-zero exit code is still success
 "tool_output": {"command": "ls", "stdout": "...", "stderr": "", "returncode": 0,
                 "was_truncated": false, "timed_out": false},  // structured result on success; null on failure
 "tool_output_text": "...",         // current text the LLM will see; mutable by prior hooks
 "tool_error": null,                // populated on failure
 "duration_ms": 0}                  // not reported yet: always 0

parent_session_id is set when running inside a subagent. Exceeding timeout kills the whole process tree.

A hook signals back via its exit code and stdout (stderr is reserved for diagnostics — Vibe never parses it for control):

Structured-response schema:

{
  "decision": "allow" | "deny",          // optional; default "allow"
  "reason": "string",                     // required when decision == "deny"
  "system_message": "string",             // optional UI note
  "hook_specific_output": {
    "tool_input": { ... },                // pre_tool only
    "additional_context": "string"        // post_tool only
  }
}

Unknown fields are tolerated at every level. Fields that aren't meaningful for the current hook type are silently ignored.

Don't self-name in system_message or reason — the UI prefixes hook-end-event content with [hook-name] automatically, and pre_tool denials are wrapped as Tool 'X' was denied by hook 'Y': {reason} before the LLM sees them. A hook that writes "reason": "guard: refused..." will produce hook 'guard': guard: refused... downstream.

decision: "deny" per hook type:

Event-specific payloads:

• hook_specific_output.tool_input (pre_tool): full replacement of the model's arguments. Vibe re-validates against the tool's schema after each rewriting hook — the first invalid rewrite aborts the chain and synthesizes a denial attributing the failure to that hook. Rewrites compose: hook N receives tool_input as rewritten by hooks 1..N-1.
• hook_specific_output.additional_context (post_tool): text appended (with ) to the current tool_output_text. Composes with a same-hook decision: "deny": deny replaces first, then additional_context is appended to the replacement.

Failure path. Any failure (non-zero exit, timeout, spawn failure, non-conforming stdout) emits a UI warning and lets the gated action proceed (fail open). With strict = true on a tool hook:

strict is forbidden on post_agent.

Execution semantics

• A hook command is a shell line, run in the session working directory. On macOS and Linux it runs in /bin/sh. On native Windows it runs in the same shell as the bash tool: Git Bash when it is installed, otherwise PowerShell. Write Windows hooks in that shell's syntax.
• Hooks of the same type fire sequentially in load order (project file first, then user file; declaration order within each file).
• Tool calls within a single LLM turn run concurrently; each call's hook chain runs serially but the chains run in parallel across calls. Hooks that touch shared state (filesystem, env) must coordinate themselves.
• pre_tool rewrites take effect everywhere downstream: the user permission prompt sees the rewritten arguments, the tool runs with them, and the assistant message is patched so subsequent LLM turns reflect what actually ran.

Pattern Matching

Tool, skill, and agent names support three matching modes:

• Exact: "bash", "read_file"
• Glob: "bash*", "mcp_*"
• Regex: "re:^serena_.*$" (full match, case-insensitive)

CLI Parameters

vibe [PROMPT]                       # Start interactive session with optional prompt
vibe -p TEXT / --prompt TEXT         # Programmatic mode using `default_agent`, one-shot, exit
vibe -p TEXT --auto-approve          # Programmatic mode with all tool calls approved
vibe -p TEXT --agent lean --yolo      # Lean mode with all tool calls approved
vibe --agent NAME                   # Select agent profile (falls back to `default_agent` config)
vibe --auto-approve / --yolo         # Approve all tool calls for the selected agent
vibe --workdir DIR                  # Change working directory
vibe --worktree NAME                # Create/reuse a git worktree under $VIBE_HOME/worktrees on branch NAME and run inside it. Auto-cleanup only for worktrees Vibe created this run and only after a session started; reused worktrees and attached (pre-existing) branches are kept unless confirmed. -p sessions keep worktrees. Ignored with --setup/--check-upgrade.
vibe --worktree                     # Same, but Vibe picks an unused name from the prompt (a random slug when there is no prompt) on a vibe/<name> branch, and never reuses an existing worktree. The prompt must precede the flag or follow a `--`, since --worktree otherwise reads it as NAME.
vibe --add-dir DIR                  # Extra working dir loaded for context (repeatable). Implicitly trusted.
vibe --trust                        # Trust cwd for this invocation only (not persisted). Skips the trust prompt.
vibe -c / --continue                # Continue this terminal's last session in cwd, else the latest in cwd
vibe --resume [SESSION_ID]          # Resume a specific session
vibe -v / --version                 # Show version
vibe --setup                        # Run onboarding/setup
vibe update / vibe --check-upgrade  # Check for a Vibe update now, prompt to install it, and exit
vibe --max-turns N                  # Max assistant turns (programmatic mode)
vibe --max-price DOLLARS            # Max cost limit (programmatic mode)
vibe --max-tokens N                 # Max total session tokens (programmatic mode)
vibe --system-prompt-file PATH      # Replace Vibe's own system prompt instructions with PATH's text (programmatic mode)
vibe --model ALIAS                  # Model alias for this run, not saved (programmatic mode)
vibe --enabled-tools TOOL           # Enable specific tools (repeatable)
vibe --disabled-tools TOOL          # Disable specific tools (repeatable)
vibe --output text|json|streaming   # Output format (programmatic mode)

Built-in Agents

There are two kinds of agents:

• Agents are user-facing profiles selectable via --agent or Shift+Tab. They configure the model's behavior, tools, and system prompt.
• Subagents are model-facing: the model can spawn them autonomously to delegate subtasks (e.g. exploring the codebase). Users cannot select subagents directly.

Agents

• ask: Requests approval for tool executions
• accept-edits: Default agent; auto-approves file edits but asks for other tools
• auto-approve: Auto-approves all tool calls
• lean: Specialized Lean 4 proof assistant. Not available by default — must be installed with /leanstall (removed with /unleanstall). Use --agent lean --auto-approve or --agent lean --yolo to run Lean mode without tool prompts.

Subagents

• explore: Read-only codebase exploration subagent with grep, file reading, and skill loading. Spawned by the model, not selectable by the user.

Custom agents are TOML files in ~/.vibe/agents/NAME.toml. One with agent_type = "subagent" is a custom subagent the model can spawn. It can load every skill the parent session has, unless its file sets enabled_skills or disabled_skills; these accept the same patterns as in config.toml.

A custom subagent (agent_type = "subagent") runs with the session's hooks. Each of its tool calls must pass both the session's rules and its own [tools.NAME] rules (permission, allowlist, denylist): a profile can refuse or prompt for what the session allows, never allow what the session refuses or prompts for. enabled_tools/disabled_tools likewise only narrow the session's tools. Approvals given to a subagent's prompts apply to the whole session. Under --auto-approve, a subagent's calls run without consulting allowlists or denylists, like the main agent's; a tool its profile sets to never stays refused. To use a custom subagent, the model must spawn it by its agent type (config_name); a subagent spawned with only a name inherits the main agent's configuration.

A subagent's instructions (instructions, or the prompt system_prompt_id names) are limited to 65,536 characters; the primary agent's prompt has no such limit. An agent file that cannot be used (invalid TOML, an unknown safety or agent_type, an invalid setting, a missing prompt, instructions over the limit, no description) is not offered to the model. Vibe reports it as a config issue naming the file and the failing field: in the startup Warnings box in the TUI, a stderr line and an export warning with -p. Spawning that type fails with subagent_type_rejected and the same reason.

A custom subagent runs on the session's model unless its file sets active_model to an alias from [[models]]; its children then run on that model (name, temperature, thinking, provider, and the model's context window, compaction threshold and image support). The provider's endpoint and credentials always come from the session's [[providers]]: a provider the agent file redefines is ignored for routing. An alias that is unknown, excluded by allowed_models, or served by a provider the session does not define rejects the agent file like the cases above; Vibe never falls back to the session's model.

Built-in Slash Commands

• /help - Show help message
• /config - Full-screen settings browser. Fields show their value and origin layer (default / TOML / env / override). Type to filter, arrows to move, Enter to edit; booleans toggle, closed-set fields (theme, models) pick from a list, scalars edit inline, complex fields open a JSON editor. The edit modal shows an inspector of the layers setting the field; edits persist to the user config (~/.vibe/config.toml) by default, Tab cycles the save target through the project config (.vibe/config.toml, when the project layer is active) and the ephemeral session override (until restart), and Ctrl+R clears the field one writable layer at a time toward the default. The tools field opens a grouped tool list with a per-tool config editor (permission, allow/deny lists, Ctrl+E for raw JSON). Enabling/disabling whole MCP servers or connectors stays in /mcp.
• /model - Select active model. The footer shows the active model and its thinking level (Name[level]) before the context gauge; clicking it opens /model and keeps the prompt being typed (only while idle, like the typed command).
• /thinking - Select thinking level
• /theme - Select UI theme; auto follows terminal/OS appearance, mistral-dark and mistral-light use the Mistral brand palette (persisted in config)
• /reload - Reload configuration, agent instructions, and skills from disk (skills added or edited on disk are also picked up on their own while file_watcher_for_autocomplete is on, except in sandboxed sessions)
• /clear, /new - Start a new conversation. Optionally pass a prompt to seed it
• /log - Show path to current interaction log file
• /log-level - Show or set the log level. /log-level prints the full chain (session, env, config, effective); /log-level set <LEVEL> sets a process-lifetime override; /log-level set-global <LEVEL> also persists to config.toml; /log-level unset clears the session override. LEVEL is one of DEBUG, INFO, WARNING, ERROR, CRITICAL.
• /debug - Toggle debug console
• /compact - Compact model context by summarizing. The session ID and visible conversation stay intact; the auto title is refreshed to reflect the compacted conversation (unless renamed manually).
• /rename <title> - Set a manual session title. Persists to meta.json (title_source=manual), updates the terminal tab title, and is never overwritten by automatic title generation.
• /retry [additional instructions] - Run again a model response that failed with a backend error or was interrupted with Esc. The new reply replaces the unfinished one and its error or interrupt row. Optional instructions are passed to the model for the retry. Relevant error messages also hint at this command; with nothing to retry, it says so.
• /status - Display agent statistics
• /context - Show what fills the context window: a full-width gauge of the window split by category (system prompt, tool definitions including MCP and connector tools, skills, conversation), then free space and the auto-compaction buffer (marked where compaction triggers), with a legend of each category's share and estimated tokens. The total comes from the model API's last call; the repartition is estimated from each category's share of the characters that call actually sent. The footer's token counter opens the same view when clicked (also while a turn runs); a subagent shows its own window. Estimates appear only once the conversation has started (at least one model call); a resumed session shows its total alone until its next model call.
• /whoami - Display the Mistral signed-in user, workspace, and plan
• /copy - Copy the last agent message to the clipboard
• /paste-image - Paste an image from the OS clipboard into the prompt. macOS only — the command is not registered on Linux or Windows.
• /voice - Configure voice settings. With voice mode on, Ctrl+T starts or stops recording
• /mcp (or /connectors) - Display MCP servers and connector status. The browser opens on the first item; press Up or Left to move into the fuzzy-search bar, and Up again to wrap to the last item. Pass a server or connector name to list its tools or open its auth panel when authentication is required
• /mcp add <url> - Add a hosted OAuth MCP server. Supports --name <alias>, repeatable --scope <scope>, --transport <http|streamable-http>, --no-login, and --allow-insecure-http (permit a plaintext http:// URL on a non-localhost host such as a LAN server). Starts OAuth login by default. OAuth-only; use vibe mcp add <name> --url <url> --api-key-env <var> for API-key/static auth.
• vibe mcp remove <name> - Remove an MCP server from the user configuration and delete its stored OAuth credentials when available.
• /mcp status - Display MCP auth state (ok, needs_auth, static, stdio)
• /mcp login <alias> - Start OAuth login for an MCP server
• /mcp logout <alias> - Log out from an MCP server and delete stored OAuth secrets
• /resume (or /continue) - Browse and resume past sessions for the current folder. The picker header shows the folder being listed. Press d twice to delete a saved session; the active session cannot be deleted here.
• /rewind - Rewind to a previous message. Also triggered by pressing Esc twice on an empty input; if the input has content, the first double-Esc clears it instead. In the rewind panel: ↑/↓ pick option, Shift+↑/↓ scroll, ←/Esc edit previous message, → edit next message, Enter accept, q quit.
• /branch (alias /fork) - Fork the current conversation into a new resumable session, leaving this session unchanged. Resume the copy with vibe --resume <id>.
• /loop [schedule] [prompt] - With the Unified Harness, ask the model to schedule a recurring prompt, e.g. /loop every two minutes check the build or /loop weekdays at 9am review CI.
  • The full message is sent unchanged to the model, which uses the cron tool to schedule, list, cancel, or clear prompts. There are no TUI management subcommands; bare /loop also goes to the model. Busy submissions queue.
  • Fixed intervals have a 30-second minimum; calendar schedules use five-field cron in the machine's local timezone. Maximum 50 schedules per session.
  • Prompts run only while the session is live and idle, without catch-up bursts. Schedules persist in scheduled-loops.json beside the Unified session and are restored on resume. They do not run while Vibe is closed.
• /proxy-setup - Configure proxy and SSL certificate settings
• /leanstall - Install the Lean 4 agent (leanstral)
• /unleanstall - Uninstall the Lean 4 agent
• /plugins - Display the plugins this session is running (Unified Harness only). Shows each plugin's name, scope, source format, content digest, and components (skills, MCP servers, agents, hooks, knowledge, connectors, tools). Press r inside the view to reload.
• /reload-plugins - Re-pin this session's plugins and report what changed (Unified Harness only). Re-discovers plugins from disk, re-pins the snapshot, and prints a diff of added, removed, and updated plugins.
• /data-retention - Show data retention information
• /release-notes (alias /whatsnew) - Browse the release notes of every Vibe version, newest first, in a full-screen, searchable pager. When a new session starts after an upgrade (not on resume), Vibe shows the notes of every release since the last one seen (the newest five, then a pointer to /release-notes); the last seen version is seen_whats_new_version in ~/.vibe/cache.toml.
• /teleport - Teleport session to Vibe Code Web (only available when Vibe Code is enabled)
• /remote-project - Select the Vibe Code Web project for this repository (only available when Vibe Code is enabled)
• /exit - Exit the application

File Mentions (@)

Type @ in the chat input to autocomplete files and folders. A bare @ lists non-hidden immediate children directly from the filesystem for fast browsing. Once you type a path character, Git workspaces use tracked and non-ignored untracked paths, including nested .gitignore rules; outside Git the picker falls back to the project tree. Pressing Tab/Enter inserts the chosen path. Your message text is sent as-is (the @path stays in the prompt); behavior then depends on the mention kind:

• Text files trigger a synthetic read_file tool call injected right after your message, so the file content arrives as a fresh tool result every turn (no caching/dedup). The same limits as the read_file tool apply (~2000 lines / 64 KB per call; larger files are truncated or reported as an error result). Re-mentioning a file always re-reads it.
• Folders are not read automatically — the path stays in your message text and the agent can read_file/grep it on demand.
• Image files (.png, .jpg, .jpeg, .gif, .webp) become image attachments — sent alongside the prompt as native multimodal content for vision-capable models.

Image attachments:

• Require supports_images = true on the active model in config.toml. The legacy loop rejects an image its model cannot read; under the Unified Harness the agent is shown a description of the image instead, and with no describer reachable the CLI refuses a pasted image.

• The describer is picked automatically: any supports_images model on the active model's own provider, no config needed. Same provider means same key and same endpoint, so no image goes anywhere the session was not already talking. A vision_model in config.toml overrides that choice and is the only way to reach another provider.

• Each image is then described once as the turn's input is prepared, and the agent receives the text inside an <image alias="..."> block in place of the pixels. The user's prompt steers what the describer looks for, and the description is reused for the rest of the session. The describer never reasons, whatever its thinking says: the trace is charged against the description budget and buys nothing on a transcription. Claude models that can't turn thinking off (Opus 5.5, Sonnet 5.5, Fable 5) reason at their lowest level instead.

• A describe that fails does not fail the turn. The agent gets a placeholder saying the image could not be read, and a warning names the image and the provider's reason.

• With no describer reachable the image is left alone, and a model that cannot read images gets [Image] and, for an image from a file, a file:// link to it.

• Snapshotted into <session_dir>/attachments/<sha1>.<ext> so that resumed sessions stay reproducible even if the source file is moved.

• Capped at 10 MiB per image and 8 images per message.

• Out-of-project paths work via @/abs/path/to.png (the picker only suggests project files, but the @-parser accepts absolute paths). Drag-and-drop from Finder into Terminal, iTerm2, or Ghostty is intercepted at paste time: if the pasted content is a standalone existing absolute or home-relative file or folder (or a newline-delimited list of them), the input automatically prepends @ and quotes paths containing spaces. This applies to text files and folders (images become [Image #N] placeholders, see below); pasted prose, relative paths, and missing paths are left unchanged.

• Image copy/paste from the clipboard (macOS only for now): writes the image to a persistent temporary PNG, inserts a numbered [Image #N] placeholder at the cursor, then snapshots it under the session attachments on submit. A pasted standalone image path becomes a placeholder too. In a ! shell command the raw path is inserted instead. Two entry points:

  1. Ctrl+V keybinding inside the prompt.
  2. /paste-image slash command.

  Uses osascript with a TIFF→PNG fallback via sips. On Linux and Windows the binding and the slash command are not registered at all, so the feature is invisible to users on those platforms.

• Rendered in the chat bubble as one dim attached image: footer line per image, linking each attachment to its snapshot. Clicking opens the file with the OS default image viewer.

Footer

The bottom row of the CLI reads, left to right:

• The working directory, elided from the start (…/vibe/cli-rust) so the project name stays visible, then ⧉: clicking it copies the whole path.
• The session's PR/MR: ● PR #123 mergeable. The dot's color gives its state (green mergeable, red action needed or closed, accent merged, grey draft or waiting). Clicking opens it in the browser; several PRs/MRs show PRs (2), which opens a list.
• On the right, with a Vibe key (chat plan) and a Mistral model, the monthly Vibe Code usage gauge usage ██▍ 40%: the organization's spend against its allowance, re-read after each turn (at most once a minute), right after a rate limit and when the model changes. Without pay-as-you-go its fill turns yellow from 75% and red from 90% (Vibe Code stops at 100%); with pay-as-you-go it keeps its color and reads · PAYG, since usage past the allowance is billed. It is absent for API and Codestral keys, non-Mistral models, members without permission to see the organization's usage, and platforms without the billing API.
• Then the context gauge ctx ▍ 23k/600k.

On narrow terminals the footer first turns the context then usage gauges into one-cell levels (ctx ▁ 4%), then drops the PR status, then shortens the working directory to its last folder (…/cli-rust), then drops the gauges' numbers and shortens the PR link (● #123), and finally hides the link. The working directory, its copy icon and the gauges stay.

The input box's top border names the active model with its thinking level (mistral-medium-3.5[medium], the level muted) before the mode label, which is colored by the agent's safety; the borders themselves are neutral. Clicking the model opens /model while the session is idle. On narrow terminals the border drops the thinking level, then elides and hides the model; the mode stays whole.

Input Queue

Prompts submitted while the agent is running are accepted by the app server and merged into a single follow-up turn, so everything queued during one generation is delivered to the agent together (not one turn per prompt) once it finishes. Each queued prompt keeps its own message so it stays individually editable and removable. This includes plain prompts, /skill ... prompts, and prompts with @ mentions. !bash, non-side-channel slash commands, and &teleport require an idle session and are rejected while busy with a notice above the input that stays until the next key. Ctrl+C removes the newest queued prompt (LIFO); Esc interrupts the active turn and pauses the remaining queue; pressing Enter (empty or not) on a paused queue resumes it.

Allowlisted slash commands (side_channel=True) run immediately via a side channel while the agent or bash is busy, or while prompts are queued. Only one side-channel command runs at a time. Besides read-only commands (/help, /status, /copy, ...), this includes settings a running turn never reads (/theme, /log-level, /voice, /proxy-setup, /remote-project); their changes are saved at once. These settings still wait for a compaction, teleport, project request or agent install to finish.

Commands that change what a turn reads (e.g. /config, /model, /thinking, /clear, /compact, /rewind, /resume, /branch, /reload, /leanstall, /unleanstall, /teleport, /retry, /mcp, /connectors, /plugins, /reload-plugins) are rejected while busy and can be retried when the session is idle.

While the queue is non-empty and the agent is busy, pressing Up enters queue selection mode: the last queued item is highlighted and the input is locked (no cursor, no typing). Up/Down navigate between queued prompts, Enter loads the selected prompt into the input for editing (press Enter again to update it in-place), Backspace (or Ctrl+C) removes the selected item and moves selection to the next, and Esc exits selection mode and restores the original input text.

Plugins

Plugins are bundled extension packages that can contribute skills, MCP servers, hooks, agents, knowledge, connectors, and tool libraries to a session. They are discovered from two locations:

• ~/.vibe/plugins/ — user scope (global, available in every project)
• .vibe/plugins/ — project scope (requires trusted folder; overrides user scope on name collisions)

Each subdirectory inside a plugins directory is a single plugin root.

Plugin Manifest (plugin.json)

Every plugin has a plugin.json manifest following the Agent Plugins 1.0 schema (https://agent-plugins.org/schemas/1.0.0/plugin.schema.json):

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
  "name": "my-plugin",
  "version": "1.0.0",
  "description": "Optional description",
  "author": { "name": "Author", "email": "a@b.com", "url": "https://..." },
  "homepage": "https://...",
  "repository": "https://...",
  "license": "MIT",
  "keywords": ["tag1"],
  "extensions": {
    "ai.mistral.vibe": {
      "schemaVersion": 1,
      "displayName": "My Plugin",
      "toolNamespace": "myNs",
      "toolOverrides": {
        "tool-name": { "name": "renamedTool", "exposure": "programmatic" }
      }
    }
  }
}

The ai.mistral.vibe extension is optional but gates all Vibe-specific components (hooks, knowledge, agents). toolNamespace is a TypeScript-identifier-safe string used to prefix all component names (e.g. myNs:my-skill). If omitted, it is derived from the plugin name. displayName is an optional title of 1 to 100 characters; the plugin name stays its identity.

Reserved namespaces (rejected): file_system, self, process, skill, subagent, vibe.

Plugin Contents

A plugin tree may contain any combination of:

Environment variables PLUGIN_ROOT and PLUGIN_DATA are injected automatically into MCP server and hook processes and are reserved. Plugin hooks run in the plugin directory.

Foreign Plugin Formats

Vibe also adapts non-native plugin formats:

Adapted plugins can only contribute skills and MCP servers — hooks, knowledge, agents, libraries, and connectors are native-only. Executable code in foreign plugins (.js, .ts) is refused and reported as unsupported.

Plugin Pinning and Reload

On session start, discovered plugins are resolved into a ResolvedPluginSnapshot — a portable, host-path-free, secret-free representation pinned to the session. This ensures resume/rewind reproducibility even if files change on disk.

/reload-plugins re-discovers plugins from disk, builds a fresh snapshot, and prints a diff (+ added, - removed, ~ updated). Tools that are no longer present after a reload are retained as unavailable routes so conversation history stays valid. Drift detection flags tools whose schema fingerprint changed between pin and live source.

Example Plugin Tree

my-plugin/
├── plugin.json
├── mcp.json
├── skills/
│   └── my-skill/
│       └── SKILL.md
└── ai.mistral.vibe/
    ├── hooks.toml
    ├── knowledge/
    │   └── my-knowledge/
    │       └── KNOWLEDGE.md
    └── agents/
        └── my-agent.toml

Skills System

Skills are specialized instruction sets the model can load on demand. Each skill is a directory containing a SKILL.md file with YAML frontmatter.

Skill File Format

---
name: my-skill
description: What this skill does and when to use it.
user-invocable: true
allowed-tools: bash read
---

# Skill Instructions

Detailed instructions for the model...

Skill Search Order (first match wins)

1. skill_paths from config.toml
2. .vibe/skills/ in trusted project directory
3. .agents/skills/ in trusted project directory
4. ~/.vibe/skills/ (user global)
5. ~/.agents/skills/ (user global, Agent Skills standard)

Invoking Skills

Two entry points:

• The model loads a skill on demand via the skill tool.
• The user invokes a user-invocable skill by typing /skill-name (optionally followed by extra instructions). The user turn stays the literal /skill-name text; the skill is loaded programmatically and appears to the model as a synthetic skill tool call and result immediately after that turn — the model does not call the tool itself.

Skills with user-invocable: false are model-only: they are hidden from the slash menu and /skill-name will not resolve them (it is treated as a plain prompt). The model can still load them via the skill tool.

Skills with disable-model-invocation: true stay in the slash menu but are hidden from the model and cannot be loaded through the skill tool. Skills using OpenAI's agents/openai.yaml convention can set policy.allow_implicit_invocation: false for the same provider-independent behavior.

A / at the very start of the input opens the slash menu (commands and skills). A skill row's description starts with its scope in the theme's primary color, whether it is a local, plugin or registry skill: project · … (this project only), user · … (the user's, in every project), vibe · … (shipped with Vibe), or skill · … when an older server does not say; a plugin skill also shows its plugin in its plugin:name. Rows without one are built-in commands. A built-in command wins over a skill with the same name, so such a skill is left out of the slash menu; it can still be mentioned mid-prompt. A /word typed mid-prompt (not the first word) instead opens a completion popup of matching user-invocable skills, with the same scope prefix.

Environment Variables

• VIBE_HOME - Override the Vibe home directory (default: ~/.vibe)
• MISTRAL_API_KEY - API key for Mistral provider
• VIBE_ACTIVE_MODEL - Override active model
• VIBE_* - Any config field can be overridden with the VIBE_ prefix
• LOG_LEVEL - Overrides log_level config for $VIBE_HOME/logs/vibe.log. One of DEBUG, INFO, WARNING (default), ERROR, CRITICAL. Invalid values fall back to WARNING. Use /log-level to change at runtime.
• LOG_MAX_BYTES - Max size in bytes of vibe.log before rotation (default: 10485760, i.e. 10 MiB).
• DEBUG_MODE - When true, forces DEBUG-level logging.
• VIBE_TYPING_GRACE_PERIOD_MS - Milliseconds the agent waits for a typing pause before showing tool-approval / ask-user-question dialogs (default: 1000). Set to 0 to disable. Negative or non-numeric values fall back to the default.

API Keys (.env file)

The .env file in VIBE_HOME stores API keys in dotenv format:

MISTRAL_API_KEY=your-key-here

This file is loaded on startup and its values are injected into the environment.

Trusted Folders

Vibe uses a trust system to prevent executing project-local config from untrusted directories. The trust database is stored in ~/.vibe/trusted_folders.toml. Project-local config (.vibe/ directory) is only loaded when the current directory is explicitly trusted.

Interactive mode prompts to trust unknown folders. The prompt targets the closest ancestor of the cwd (the cwd itself included) containing a .git entry; the search excludes the user's home directory and the filesystem root, and falls back to the cwd if no qualifying ancestor is found. Programmatic mode (-p/--prompt) never prompts: the folder is untrusted. Use --trust to trust cwd for the current invocation only (not persisted). --trust and --worktree both skip the prompt: they grant the workspace trust for the session, so there is no decision left to ask about. Without this a --worktree run would prompt on every launch, since each worktree is a directory the trust database has never seen.

Sensitive Files — DO NOT READ OR EDIT

NEVER read, display, or edit any of these files:

• ~/.vibe/.env (or $VIBE_HOME/.env) — contains API keys and secrets
• Any .env, .env.* file in the project or VIBE_HOME

If the user asks to set or change an API key, instruct them to edit the .env file themselves. Do not offer to read it, write it, or display its contents. Do not use tools (read, write_file, bash cat/echo, etc.) to access these files.

How to Modify Configuration

To help the user modify their Vibe configuration:

1. Read current config when present: Read ~/.vibe/config.toml (or the path from VIBE_HOME if set). A missing file means Vibe is using built-in defaults.
2. Create a backup when present: Before editing an existing file, copy it to config.toml.bak in the same directory. This applies to any existing config file you are about to modify (config.toml, trusted_folders.toml, agent TOML files, etc.)
3. Edit the TOML file: Make changes using the edit tool
4. Reload: The user can run /reload to apply changes without restarting

For API keys, tell the user to edit ~/.vibe/.env directly — never read or write that file yourself.

For project-specific configuration, create/edit .vibe/config.toml in the project root (the folder must be trusted first).