Phaedo MCP Server

Phaedo — your cognitive fingerprint, on every AI surface

Phaedo learns how you think — communication style, working style, decision pattern — into a portable, user-owned cognitive fingerprint, and carries it to any AI tool. This repo is the open half of that system:

• spec/ — Anamnesis (formerly "Phaedo Protocol"), the open cognitive-fingerprint standard: machine-readable JSON Schemas (fingerprint layers, consultation request/response, deference policies, receipts) + the conformance suite.
• docs/protocol/ — the Anamnesis spec (v0.1), the MCP binding, and accepted proposals.
• mcp/ — the reference MCP server (phaedo-mcp on npm): injects your fingerprint at session start and answers agent consultations — an agent at a decision point can ask "would my human proceed here?" and get a calibrated proceed / clarify / escalate / decline signal (or an honest abstain), gated by user-authored deference policies, with an encrypted on-device audit trail. On a blocking signal it can escalate to your paired phone in real time for a live approve / deny / modify, and your overrides teach the act-as-me channel — how you want delegated work handled, not just how you work yourself.

Quick start

npx phaedo-mcp        # run the server over stdio

Or wire it into Claude Desktop / Cursor / Claude Code automatically:

npm install -g phaedo-mcp
cd "$(npm root -g)/phaedo-mcp" && npm run setup

Out of the box it serves a sample fingerprint, so every tool works immediately. Your real fingerprint comes from the Phaedo app — currently in private beta (phaedo.so) — learned on-device from your actual usage; pair with npm run pair. See mcp/README.md for the full guide (pairing, policies, receipts, the agreement metric).

Design commitments

• User-owned & local. The fingerprint lives on your devices, encrypted; the server holds no plaintext at rest. Nothing is sent to any Phaedo service.
• Signals out, never structure. Consultations return a signal + confidence — never fingerprint internals (spec §10.4).
• Honest abstention. No coverage → insufficient_signal, not a confident guess.
• User-authored guardrails win. Deference policies (§10.5) override any inferred signal; standing authorizations are explicit and auditable (§10.6).

What's not in this repo

The learning pipeline that builds fingerprints (conversation capture, on-device extraction models) is Phaedo's proprietary side. This repo is everything a client, agent, or independent producer needs to consume and interoperate: the spec, the schemas, and a complete reference server.

License

Apache-2.0. The Phaedo name and logo are trademarks; see phaedo.so.