self-inspect-mcp

MCP server exposing one tool, self_inspect. The agent passes a one or two sentence description of what it is doing; the tool returns a single self-inspection metathought (a concrete self-check) or a no-match note. Keyless and free.

What it does, measured: in a 30-turn design session (Claude Sonnet 4.6), agents calling self_inspect each turn surfaced ~3.5x more decision-forks (assumptions, edge cases, preconditions) than the identical agent with no tool, the forks it would otherwise pass over in silence. Data, observations, and a one-command reproduction: evaluation.

By default it calls the public endpoint https://api.ejentum.com/self-inspect. It can also run the exact same published heuristic fully offline.

Install (stdio)

npx -y self-inspect-mcp

Claude Desktop / Claude Code / Cursor / Windsurf config

{
  "mcpServers": {
    "self-inspect": {
      "command": "npx",
      "args": ["-y", "self-inspect-mcp"]
    }
  }
}

No API key. No env block required.

The tool

self_inspect({ thought: string }) -> one metathought, or a no-match note.

Pass what you are doing, not the topic. Good: "I'm about to assert the default timeout is 30s from memory". Returns the matched metathought text to absorb and act on.

Configuration

| env var | effect | |---|---| | SELF_INSPECT_LOCAL=1 | Run the vendored selector + CSV locally. No network. The selection is identical to the hosted endpoint because the core is copied, unmodified, from the canonical repo (scripts/vendor.mjs). | | SELF_INSPECT_API_URL | Override the hosted endpoint URL (self-hosting / testing). |

Auditability

The selector and the CSV that drive every answer are published in the parent repo and vendored here under vendor/. There is no hidden model and no scoring you cannot read. Run SELF_INSPECT_LOCAL=1 to confirm the offline answer matches what the endpoint returns.

License

MIT.