hunchful

Hunchful MCP server

Remote MCP server: https://hunchful.io/mcp (Streamable HTTP)

Hunchful is an open implementation of the Cognitive Model Protocol (CMP): a person makes a revisable, falsifiable model of how they collaborate best with AI — hunches with a confidence, an uncertainty, and what would change my mind — at a permanent URL any agent can read. This server lets your assistant read that model, adapt to it, and help it grow slowly, with the person approving everything that matters.

Add it

Claude — Settings → Connectors → Add custom connector → https://hunchful.io/mcp

ChatGPT (developer mode) — Settings → Connectors → Add → URL https://hunchful.io/mcp, auth OAuth (sign in with your Hunchful account to act on your own model) or none (open reads; pass keys as tool arguments).

Anything else — any MCP client that speaks Streamable HTTP. OAuth 2.1 with PKCE + dynamic client registration; discovery at /.well-known/oauth-authorization-server.

Tools (rising trust)

Tool	Auth	What it does
read_model	none	The full model: hunches with confidence, uncertainty, falsifier, pole labels, situational activation rules, and the open-questions to-do list.
find_complementary_thinkers	none	Which famous thinkers would complete this person — opposite poles on the axes where opposites unstick each other.
propose_refinement	contribution key / OAuth	Strengthen/weaken suggestions → the owner's confirm queue. Nothing applies until they accept.
suggest_question	contribution key / OAuth	Propose a NEW axis as a question → the owner's inbox (+ opt-in email).
apply_refinement	edit key / OAuth owner	Direct, logged, reversible confidence nudges with private evidence pointers.
add_hunch	edit key / OAuth owner	Record a hunch the person just confirmed in conversation — visible, revisable, removable.
find_matches	edit key / OAuth owner	Look for complementary people. Agents propose; humans accept — double opt-in, identities never revealed by the tool.

The rules the server enforces

• Everything is a hypothesis — revisable positions on two-poled axes, never labels.
• Nothing is silent — every write is logged in the model's revision history; evidence entries are individually deletable and revert exactly what they applied.
• Provenance pointers only — no conversation text, quotes, or third-party names, ever.
• Never identity — a protected-attribute screen rejects axes about religion, politics, health, ethnicity, gender, sexuality, or neurotype on every write path.
• Situational hunches apply only when the person says they're in the state.
• Slow by design — the model asks agents for at most one suggestion a day, one new question a week, and silence when nothing real happened.

Links

• Agent guide: https://hunchful.io/agent
• Machine-readable model example: https://hunchful.io/m/{id}/cmp.json
• Protocol: https://github.com/filinils/cognitive-model-protocol
• Discovery: https://hunchful.io/llms.txt

The Hunchful app itself is a private codebase; this repo documents its public MCP surface.