Critical Thinking

A Model Context Protocol server for critical, narrated, sequential thinking. Think one step at a time, out loud — with required confidence calibration and adversarial self-critique on every thought.

It fuses three disciplines:

1. Sequential thinking — break problems into ordered, numbered steps; revise; branch.
2. Thinking out loud — explain each thought in first-person, exploratory voice. Putting half-formed reasoning into words is itself the double-check on it.
3. Critical self-examination — every thought is paired with confidence, assumptions, critique, and a counter-argument.

The single tool is criticalthinking. Every call must include the four critical-thinking fields — there is no opt-out, by design.

Install & usage → docs/usage.md. One-line install:

go install github.com/jacaudi/critical-thinking/cmd/critical-thinking@latest

Or install the Claude Code plugin under plugins/critical-thinking/: it auto-installs the server, adds a two-gate verification skill, and a hook that injects the two-gate protocol into every prompt. See its README.

One-call example

Request:

{
  "thought": "I think we should normalize first because reads dominate writes.",
  "thoughtNumber": 1, "totalThoughts": 3, "nextThoughtNeeded": true,
  "confidence": 0.6,
  "assumptions": ["read:write ratio is ~10:1"],
  "critique": "Drifted into solution mode without confirming the ratio.",
  "counterArgument": "If writes dominate, monolith-first is simpler.",
  "nextStepRationale": "Verify the read:write ratio before committing to normalization."
}

Response (structuredContent):

{ "thoughtNumber": 1, "totalThoughts": 3, "nextThoughtNeeded": true, "confidence": 0.6 }

The text content is a rendered transcript in first-person, exploratory voice. Every call must send thoughtNumber and totalThoughts (both required, ≥ 1); if thoughtNumber exceeds totalThoughts the server raises totalThoughts to match. Keep each field to one tight sentence; the server does not enforce a length limit. The full contract lives in the tool description itself.

Stateless by design

The server keeps nothing between calls: no history, no sessions, no running averages. Each call validates and narrates the one thought it receives. Your own context is the record — the model that wrote thought 3 already has thoughts 1 and 2 in front of it, and it judges its calibration across them itself. Over HTTP the server speaks MCP's sessionless model (protocol 2026-07-28): no Mcp-Session-Id, every POST complete in itself, so any number of clients, gateways, or replicas can share one deployment without contaminating each other.

Documentation

• docs/usage.md — install, MCP-server & CLI-pipe usage, a worked session
  • Install · As an MCP server · As a CLI pipe · Worked session
• docs/clients.md — Claude Desktop, Codex CLI, VS Code, Cursor recipes
• docs/configuration.md — env vars, HTTP endpoints, logging, CORS, stateless HTTP
• docs/migration.md — breaking changes since http-sequential-thinking
• docs/development.md — building, testing, debugging with MCP Inspector
• plugins/critical-thinking/ — the Claude Code plugin (bundled server install + two-gate skill + always-on protocol injection)

License

MIT.