ideaudit

ideaudit-tools

The deterministic scoring behind ideaudit, as a local MCP server. No account, no key, no network, no model calls — just the arithmetic.

claude mcp add ideaudit-tools -- npx -y @inite/ideaudit-tools

Listed in the MCP registry as studio.inite/ideaudit-tools, which carries both ways in: this package over stdio, and the hosted server over HTTP. Installing straight from this repository — npx -y github:inite-ai/ideaudit-mcp — also works and runs the same code.

Why this is published

The product's claim is that it is an audit allowed to say no. A verdict nobody can inspect is a verdict nobody has to believe, so the part that produces the number is open — including the number that says no. You can read how three lens scores become KILL, disagree with the thresholds, and argue about the arithmetic rather than about our word.

It is also the honest shape of the system. The server makes zero model calls; judgement lives in the skills, determinism lives in the tools. These are the tools.

What runs locally

Twenty tools, every one a pure function of its input:

compute_dealbreakers_v2 is the one that decides. Stage-aware weights, confidence-weighted lens scores, and a deliberately risk-asymmetric verdict: GO needs a score ≥ 80 and zero red flags and average confidence ≥ 0.6, while the KILL gate is observer-invariant — a founder profile shifts every threshold above the gate, and none of them below it. You cannot buy a better answer by describing yourself more favourably. There is a test that says so.

Same input, same output as the hosted server — because it is the same specification object, imported rather than reimplemented. A test asserts that identity rather than comparing two copies, so the two cannot drift.

Where to look

Be told where the numbers are rather than hunt for them:

• src/specs/*.ts — one file per tool: what it accepts, what it means, what it returns. compute-dealbreakers-v2.ts is the verdict.
• src/server.ts — the whole server, including what it refuses.
• src/__tests__/parity.test.ts — the claims above, as assertions. The observer-invariant KILL gate is checked against all five archetypes.
• dist/bundle.js — the arithmetic, un-minified. computeBarrier is a dozen lines and every constant is visible: 0.6 off the score per direct competitor, 0.2 per adjacent, six points of range spent on SERP noise, and labels at 18 / 12 / 6. Disagree with those numbers, not with our word.

That last one is a build artefact and reads like one — the commented source of the maths lives in the monorepo this repository is mirrored from, and is not here. If you want the reasoning behind a threshold rather than the threshold itself, open an issue and ask; that is a cheaper conversation than either of us pretending.

What needs the hosted server

Not withheld to make a point — these genuinely need somewhere to keep things and someone to pay for them:

• Persistence. Saving an audit, finalising it, reading it back, comparing two of them, the dashboard. A verdict that exists only in a chat transcript is one nobody can re-open or defend later.
• Data. SERP, keyword volume, funding rounds, hiring signals, social mentions. Those are metered third-party APIs.
• The knowledge graph. Entities, evidence, contradictions across audits.

That boundary is enforced rather than documented. The hosted compute_dealbreakers_v2 takes an auditId; this one refuses it, by name, instead of letting zod strip it silently and handing you a correct verdict plus the false belief that it had been written down.

$ compute_dealbreakers_v2 {"auditId": "…", …}
{ "error": "unknown_arguments", "unknown": ["auditId"],
  "message": "This server computes; it stores nothing. …" }

The hosted server

claude mcp add ideaudit --transport http https://api.inite.studio/mcp \
  --scope user --header "Authorization: Bearer $IDEAUDIT_API_KEY"

It serves these twenty tools and everything above them. Sign up at inite.studio; --scope user matters, or the registration is bound to one directory.

Using both

They compose. Register the local server for the arithmetic and the hosted one for everything else; tool names do not collide because the local set is a strict subset served under the same names, so whichever you register wins for those twenty and the rest resolve to the platform.

If you only want to check our maths against your own numbers, the local server alone is enough and always will be.

Contributing, and a warning about it

This repository is a projection. It is rebuilt and force-pushed from the monorepo that develops it every time the package changes, which means a pull request opened here would be overwritten by the next sync — through no fault of yours, and with no notification. Saying so is cheaper than letting someone find out by losing an afternoon.

Issues are the right channel and they are read: a threshold you think is wrong, a number you cannot reproduce, a tool that accepts something it should refuse. Bring the input you used — every tool here is deterministic, so a disagreement about an answer is always settleable.

Contribute with an AI agent

You don't need to be an engineer. Open the repo in your coding agent and say /contribute (Claude Code) — or "read AGENTS.md and help me report …" in any other agent. It runs the tool on your input, finds the constant involved, and files an issue a maintainer can settle — an issue, not a pull request, for the reason above. AGENTS.md also explains why dist/ is committed and how node scripts/verify-dist.mjs checks it against src/.

Licence

Apache-2.0. See LICENSE.