Alethia — Agent-Native E2E with Verifiable Safety

Alethia is the patent-pending zero-IPC E2E runtime built for AI agents. Your agent writes the tests, runs them against a real browser, and proves destructive actions are blocked by a per-step policy gate — with a cryptographic audit trail and no cloud. 2–5× faster than Playwright MCP per flow; up to 50× faster than Playwright CLI on simple flows (benchmark).

Quick start

npm install -g @vitronai/alethia

The bridge auto-installs the signed headless runtime on first use. No signup, no email gate. Works with file:// pages and http://localhost dev servers.

Starter repo — try it in 30 seconds

Fork alethia-anvil and point your agent at __alethia__/. Ships a 250-line SPA called Anvil and 14 .alethia test files covering CRUD, search, priority, bulk actions, keyboard shortcuts, tab filters, toast stack, export — plus the expect block: safety test. Drop-in GitHub Actions workflow + runner included.

MCP setup

Add to your agent's MCP config (.mcp.json, Claude Code settings, Cursor MCP, etc.):

{
  "mcpServers": {
    "alethia": { "command": "alethia-mcp" }
  }
}

MCP tools

How alethia_tell works

Send plain English instructions against any local page:

navigate to file:///path/to/demo/incident-response.html
assert CRITICAL INCIDENT ACTIVE is visible
click Acknowledge
assert Acknowledged is visible
expect block: click Delete Incident

Alethia compiles to Action IR, runs each step through the VITRON-EA1 fail-closed policy gate, executes with synchronous DOM access, and returns structured per-step results with DOM diffs, semantic page snapshot, policy audit records, and a SHA-256 integrity hash.

Verifiable safety primitive — expect block:

expect block: is an NLP assertion unique to Alethia. The step passes if the EA1 gate blocks it, fails if the gate lets it through. This turns Alethia from "a test runner" into "a verifiable-safety framework" — you can regression-test your app's safety boundary itself.

Other frameworks can assert "nothing destructive happened" by inspecting the app's state after a click; only Alethia's assertion is that the runtime itself refused to let the click through in the first place. Meaningfully different guarantee.

Structured response fields

Every alethia_tell response now returns top-level:

• nearMatches — elements close to the failed selector
• suggestedFix — corrected selector the agent can retry with
• pageContext — { title, headings, buttons, inputs }

This feeds agent self-repair loops: a mis-selected click comes back with concrete candidates the agent can choose from instead of prose it has to regex-parse.

Safety: VITRON-EA1

Default profile is controlled-web:

• read (navigation, assertions) — always allowed
• write-low (form input, drafts) — allowed by default
• write-high (delete, purchase, submit) — blocked by default
• Sensitive input (passwords, credit cards, PII) — blocked in all profiles unless explicitly opted in

The policy gate is in-process — agents cannot bypass it.

Privacy

Everything runs on your machine. The runtime listens on 127.0.0.1:47432 (loopback only). Zero telemetry by default. The only network call is the one-time signed runtime download on first install.

Security posture — local-only by architecture

Alethia refuses to navigate to any origin outside file://, localhost, 127.0.0.1, ::1, .local, and RFC1918 private ranges. The allowlist is a compile-time constant enforced at four choke points in every signed binary we ship — NAVIGATE, alethia_propose_tests, alethia_assert_safety, and the renderer-level will-navigate / will-redirect handlers. It is not exposed as a CLI flag, env var, MCP argument, policy profile, or UI toggle. Design-partner production-origin access is handled by issuing a custom-signed build with those origins baked in, never by shipping a general-purpose override. See vitron.ai/safety.

Links

• npm: npmjs.com/package/@vitronai/alethia
• Bridge source (MIT): github.com/vitron-ai/alethia-mcp
• Runtime releases: github.com/vitron-ai/alethia
• Starter template: github.com/vitron-ai/alethia-anvil
• Website: vitron.ai
• Licensing: gatekeeper@vitron.ai