Current Environment

Python version:

!python3 --version 2>/dev/null || python --version 2>/dev/null || echo "Python not found in PATH"

Installed FastMCP version:

!uv run python -c "import fastmcp; print(f'FastMCP {fastmcp.__version__}')" 2>/dev/null || echo "FastMCP not installed — run: uv add 'fastmcp>=3.0' before scaffolding"

Trigger Matrix

When user intent matches, load the reference file listed — do not rely on training data for v3 API facts.

Choose Provider Type

flowchart TD
    Q1{What do you need?}
    Q1 -->|Define tools/resources in this server| LC["LocalProvider — default<br>No mount() needed<br>Source: providers/local.mdx"]
    Q1 -->|Add another FastMCP server's tools| MC["FastMCPProvider / mount()<br>mcp.mount(sub, namespace='ns')<br>Source: servers/composition.md"]
    Q1 -->|Wrap remote HTTP MCP server| PC["ProxyProvider<br>create_proxy('http://remote/mcp')<br>Source: providers/proxy.mdx"]
    Q1 -->|Serve files from disk as resources| FC["FileSystemProvider('path/')<br>reload=True for dev, False for prod<br>Source: providers/filesystem.mdx"]
    Q1 -->|Expose Claude/Cursor skill files| SC["SkillsProvider / ClaudeSkillsProvider()<br>skill:// URI scheme<br>Source: providers/skills.mdx"]
    Q1 -->|Build a custom provider| CC["Subclass Provider base class<br>Source: providers/custom.mdx"]

Choose Transport

flowchart TD
    Q1{How will clients connect?}
    Q1 -->|Local tool in Claude Code / desktop app| ST["stdio — default<br>fastmcp run server.py:mcp<br>Source: deployment/running-server.mdx"]
    Q1 -->|Web service or multi-client| HT["HTTP transport<br>mcp.run(transport='http', port=8000)<br>Source: deployment/http.mdx"]
    Q1 -->|Testing — in-process| IT["In-memory transport<br>async with Client(mcp) as client<br>Source: patterns/testing.mdx"]
    Q1 -->|Managed cloud deployment| PH["Prefect Horizon<br>fastmcp run via GitHub integration<br>Source: deployment/prefect-horizon.mdx"]

Choose Auth Approach

flowchart TD
    Q1{Auth requirement?}
    Q1 -->|No auth needed| NA["No auth — default FastMCP behavior"]
    Q1 -->|Validate bearer tokens per tool| RS["require_scopes('scope')<br>@mcp.tool(auth=require_scopes('write'))<br>Source: servers/auth/token-verification.mdx"]
    Q1 -->|Full OAuth2 server built-in| FO["Full OAuth server<br>Source: servers/auth/full-oauth-server.mdx"]
    Q1 -->|Delegate to external IdP — Auth0, Azure| OP["OIDC proxy / OAuth proxy<br>Source: servers/auth/oidc-proxy.mdx"]
    Q1 -->|Mix OAuth + JWT for hybrid clients| MA["MultiAuth — compose OAuth server<br>+ token verifiers (v3.1)<br>Source: servers/auth/multi-auth.mdx"]
    Q1 -->|Use PropelAuth| PA["PropelAuthProvider<br>OAuth + token introspection (v3.1)<br>Source: integrations/propelauth.mdx"]
    Q1 -->|Client calling protected server| CA["Client auth — BearerAuth / CIMDAuth / OAuthAuth<br>Source: clients/auth/*.mdx"]

Quick-Start Examples [1] [2] [3]

Minimal server

from fastmcp import FastMCP

mcp = FastMCP("my-server")

@mcp.tool  # RULE: no parentheses — v3 canonical syntax
def greet(name: str) -> str:
    """Return a greeting."""
    return f"Hello, {name}!"

if __name__ == "__main__":
    mcp.run()

Server composition

from fastmcp import FastMCP

weather = FastMCP("weather")
main = FastMCP("main")

main.mount(weather, namespace="weather")
# Tools from weather become weather_<tool-name> on main

Background task

from fastmcp import FastMCP

mcp = FastMCP("task-server")

@mcp.tool(task=True)  # RULE: task=True, NOT task=TaskConfig(...)
async def long_running(data: str) -> str:
    """Process data in background."""
    return "done"

Before deploying: run in-process pytest using the in-memory Client transport (references/testing.md) before switching to HTTP transport. In-process tests are the fastest signal that tools behave as expected.

v3 API Corrections

CONSTRAINT: These v2 patterns are deprecated or removed. Generate only the v3 form.

Version Gating

FastMCP 3.0 — Available

All core features (tools, resources, prompts, providers, transforms, auth, tasks, elicitation, client SDK, deployment) are available in FastMCP 3.0.

FastMCP 3.1 — Available (baseline for this skill's original content)

The following features were added in FastMCP 3.1.0 and require fastmcp>=3.1.0:

• Tool Search transforms — BM25SearchTransform, RegexSearchTransform for large tool catalogs
• CodeMode transform (experimental) — sandboxed Python execution for tool invocation (fastmcp[code-mode])
• transforms= kwarg — server-level FastMCP("name", transforms=[...]) constructor parameter
• MultiAuth — compose OAuth server + multiple token verifiers
• PropelAuth provider — PropelAuthProvider for PropelAuth OAuth + token introspection
• Prefab Apps (experimental) — @mcp.tool(app=True) with declarative UI components (fastmcp[apps])
• Google GenAI sampling handler — alternative to Anthropic/OpenAI sampling
• -m/--module flag — fastmcp run -m my_package.server for module mode
• FASTMCP_TRANSPORT env var — default transport selection without CLI flag
• http_client parameter — connection pooling for token verifiers
• include_unversioned option in VersionFilter
• Tool.from_tool() — immediate transformation at registration time [4]

FastMCP 3.2 — Available (released 2026-03-30)

The following features were added in FastMCP 3.2 and require fastmcp>=3.2.0:

• FastMCPApp — provider class for building interactive applications inside MCP; separates LLM-visible UI entry points (@app.ui()) from backend tools (@app.tool())
• Generative UI — LLM writes Prefab Python code at runtime instead of calling a pre-built tool with a fixed shape
• fastmcp dev apps — browser preview for app tools without an MCP host
• KeycloakAuthProvider — secure a FastMCP server with Keycloak OAuth; Docker-based local setup with pre-configured fastmcp realm
• run_in_thread=False on @mcp.tool() — opt sync tools out of the default threadpool dispatch for thread-affine libraries
• ssl verify parameter on Client — SSL certificate configuration for development with self-signed certs
• client_log_level parameter on Client — control client-side log verbosity
• ResponseCachingMiddleware token-partitioning security fix (v3.2.2) — cache now partitioned by access token; upgrade required for deployments with multiple users [5]

FastMCP 3.3 — Available (released 2026-05-15)

The following features were added in FastMCP 3.3 and require fastmcp>=3.3.0:

• fastmcp-slim — client-only distribution; install fastmcp-slim[client] for consumers who only need the FastMCP client without the full server framework; import namespace is identical (from fastmcp import Client)
• Storage backends — persistent cache and OAuth state storage backends [6]

Reference Files

All v3 reference files sourced from https://gofastmcp.com (published docs) and https://github.com/jlowin/fastmcp (source code):

• ./references/server-core.md — FastMCP(), tools, resources, prompts, context, lifespan, transforms= kwarg
• ./references/providers.md — LocalProvider, FastMCPProvider, ProxyProvider, FileSystemProvider, SkillsProvider
• ./references/transforms.md — Namespace, ToolTransform, Enabled, ResourcesAsTools, PromptsAsTools, BM25SearchTransform, RegexSearchTransform, CodeMode
• ./references/auth.md — require_scopes, OAuth variants, token verification, MultiAuth, PropelAuth, http_client pooling
• ./references/client-sdk.md — Client, transports, BearerAuth, CIMD, OAuth, sampling, elicitation, fastmcp discover, fuzzy matching
• ./references/apps.md — FastMCPApp provider class, Generative UI, low-level HTML API, Prefab Apps
• ./references/advanced.md — tasks, elicitation, storage backends, dependency injection, versioning, visibility, Prefab Apps, Google GenAI sampling
• ./references/middleware.md — Middleware base class, hook hierarchy, 11 built-in middleware, tag-based access control
• ./references/deployment.md — stdio, HTTP, server config, Prefect Horizon, nginx reverse proxy, module mode, FASTMCP_TRANSPORT
• ./references/testing.md — in-memory Client, FastMCPTransport, pytest patterns, inline-snapshot
• ./references/integrations.md — Anthropic, OpenAI, Gemini, Google GenAI, FastAPI, GitHub, Auth0, Azure, PropelAuth, Claude Code
• ./references/migration.md — v2 → v3 breaking changes, from MCP SDK
• ./references/observability.md — OTEL instrumentation, automatic spans, OTLP exporters, environment variable configuration
• ./references/real-world-patterns.md — ProxyProvider, mount(), SkillsProvider, showcase

Preserved references (not overwritten):

• ./references/evaluation-guide.md — evaluating server quality
• ./references/typescript-mcp-server.md — TypeScript MCP SDK (out of v3 overhaul scope)
• ./references/claude-code-mcp-integration.md — .mcp.json config, Claude Code deployment

Related Skills

• For pytest patterns and in-memory testing fixtures: Skill(skill: "fastmcp-creator:fastmcp-python-tests")
• For fastmcp list / fastmcp call / fastmcp discover CLI usage: Skill(skill: "fastmcp-creator:fastmcp-client-cli")
• For Python project setup (pyproject.toml, uv, src layout): Skill(skill: "python3-development:python3-development")
• For evaluating MCP server quality: ./references/evaluation-guide.md
• For Claude Code MCP config (.mcp.json): ./references/claude-code-mcp-integration.md

References

1. servers/server.mdx + servers/tools.mdx (accessed 2026-03-05)
2. servers/providers/mounting.mdx (accessed 2026-03-05)
3. servers/tasks.mdx — requires fastmcp[tasks] extra (accessed 2026-03-05)
4. Fastmcp releases v3.1.0, v3.1.1 (accessed 2026-05-23)
5. Fastmcp releases v3.2.x (accessed 2026-05-23)
6. Fastmcp releases v3.3.x (accessed 2026-05-23)