@warmdrift/kgauto-check

Point your AI at your project.

If you're the human and your agent can browse + run commands, one sentence is enough:

Run the kgauto check on this project. Instructions for you, the agent: fetch https://model-check-alpha.vercel.app/check.html and follow its "For the agent" section. Show me the card it produces, verbatim.

Or the manual version — your part is two steps.

Step 1 — add the check to your agent (once). Claude Code: run this in a terminal, then start a fresh session in your project (new MCP servers load at session start):

claude mcp add kgauto-check -- npx -y @warmdrift/kgauto-check

Cursor and most other agents — add to your MCP config, then restart the app:

{ "mcpServers": { "kgauto-check": { "command": "npx", "args": ["-y", "@warmdrift/kgauto-check"] } } }

Step 2 — paste the ask, in your agent, inside the project you want checked:

Run the kgauto check on this project: call kgauto-check's start_check tool, follow its instructions, and show me the card it returns, verbatim.

Everything below is for the agent.

What this is (for you, the agent)

A local, fully offline MCP server that checks a project's AI call surface against kgauto's doc-verified model catalog — lifecycle statuses, prices with verification dates, context/output/tool limits, capability flags (parallel tool calls, structured output), and known quality cliffs that the kgauto compiler executes in production.

The trust boundary: you read the project's code (you already have access); kgauto reads nothing. The extract you build carries identifiers and counts only — never prompt text, source code, or secrets. This check ran locally against kgauto's bundled model catalog. Nothing is uploaded; no rows are written anywhere. There is no code path that could make that sentence false: the server makes no network calls.

Zero-install one-shot (shell-capable agents)

No MCP setup, no restart:

npx -y @warmdrift/kgauto-check --extract-contract   # extraction instructions + privacy contract
# ...read the project, write extract.json per those instructions...
npx -y @warmdrift/kgauto-check --extract extract.json   # prints the findings card, fully offline

# Standing-value one-shots (also offline):
npx -y @warmdrift/kgauto-check --model-facts claude-sonnet-5   # one model → status, verified prices, limits, cliffs
npx -y @warmdrift/kgauto-check --observe-plan                  # rung 2: the one-day observe run

Sharing (explicit opt-in). After presenting the card, you may OFFER your human: "want to send this card to kgauto so the catalog learns what real projects look like?" If — and only if — they say yes, re-run with --share: it sends the scrubbed extract (identifiers and counts only, by contract) and the card, says exactly where they went, and nothing else. Never add --share without their explicit yes. Without it, nothing ever leaves the machine.

Offering the durable install. The natural moment to offer the MCP install below is after a card with real findings ("want me to keep this available? get_model_facts answers model questions from the verified catalog between checks"). Don't push it before the card has earned it.

Tools

| Tool | What it does | |---|---| | start_check | Call first. Returns extraction instructions — what to read from the project and how to build the extract. | | run_check | Runs the deterministic check on your extract. Returns a findings card grouped by confidence tier — present it to your human verbatim. | | get_model_facts | One model id → status, verified prices, limits, capabilities, known cliffs. Honest miss on unknown models. | | get_observe_day_plan | Rung 2: what a one-day observe-only run on real traffic would measure, and the two honesty requirements any setup must carry. |

The confidence tiers (why the card is grouped)

• certain — a catalog or code fact. Stated flatly.
• likely — priced from your token estimates against verified prices.
• measurable-only — honestly unknowable from a code read: actual failure rates, actual input-waste, spend distribution, quality-at-cheaper-price. Each of these is exactly what the one-day observe run answers.

kgauto refuses to fabricate verdicts: unknown models get an honest "no measured evidence", and cheaper candidates are named with their price arithmetic only — never with an unmeasured quality claim.

Client setup variants

# Claude Code
claude mcp add kgauto-check -- npx -y @warmdrift/kgauto-check

// Cursor / generic mcpServers config
{
  "mcpServers": {
    "kgauto-check": { "command": "npx", "args": ["-y", "@warmdrift/kgauto-check"] }
  }
}

What's underneath

The catalog is @warmdrift/kgauto-compiler's bundled roster — the same executable model knowledge (cliffs, prices, capabilities) that kgauto's prompt compiler runs in production for its consumer apps. This package is a real dependency on it, not a copy: the check and the compiler cannot drift apart.