clause-mcp

You are about to trade a spread between two contracts that resolve differently, and nothing you own will tell you.

Polymarket lists "U.S. enacts AI safety bill before 2027?". Kalshi lists what reads as the same contract. Polymarket's resolves YES only on restrictive provisions. Kalshi's resolves YES on a federal framework "regardless of whether those provisions mandate private-sector compliance or promote voluntary guidelines."

A voluntary standards act resolves one YES and the other NO. They are not the same bet at any price. Every arbitrage screen in existence shows them as one market at a spread, and a language model reading both rules texts will agree with the screen — because the divergence is not a difference between the two documents. It is an explicit clause on one side matched by silence on the other, and silence is the one thing a comparator has to be told not to read as agreement.

This is an MCP server. It gives an agent the checks it cannot perform on its own.

The second failure: the book that isn't there

A market on a major venue reported $65,883 of order-book depth. Within a sensible band of its own mid it held $33. It had traded $122 in thirty days. Putting $200,000 into it is 49,276 days of volume — 135 years.

A book is an offer. The tape is a fact. When they disagree, believe the tape.

Install

pip install clause-mcp

Claude Code

claude mcp add --scope user clause -- clause-mcp

--scope user matters: the default registers the server for one project directory only. The tools appear in a new session, not the one you ran the add from.

Claude Desktop, or any stdio client

{
  "mcpServers": {
    "clause": { "command": "clause-mcp" }
  }
}

Streamable-HTTP instead of stdio: clause-mcp --http --port 8848.

The eight tools

Six of the eight work the moment you install this. They read the venues live and need no local data.

The two search tools need a Clause archive — a hash-chained daily record of every open market's terms. This package does not ship one, and it will not pretend otherwise: with no archive, both tools refuse, name the state, and say in words that nothing was searched.

{
  "ok": false,
  "error": "archive_unavailable",
  "this_is_not_a_negative_result":
    "Do NOT report this as 'no such market' or 'no results'. Nothing was
     searched. The named market may exist and be trading; this install
     simply has no corpus to look in."
}

That refusal is the point of the package rather than a limitation of it. The alternative — falling back to a live search — is what the tool exists to avoid: Polymarket's list API caps pages at 100 rows and rejects offsets past ~2,100, so a live search silently misses the quiet markets, which are exactly the ones whose liquidity most needs checking. Point CLAUSE_DB at an archive if you have one.

Why the tool descriptions read the way they do

They are written for the agent, not for a human browsing a catalogue. An agent does not know it has a problem, so each description leads with the failure it prevents rather than with what the tool returns. A tool called after the mistake has been made is a tool that did nothing.

The verdicts are built the same way. check_equivalence returns two of them, because two different things can make two contracts different bets. verdict compares thresholds and settlement sources; it can reach EQUIVALENT when both are genuinely comparable. criteria_verdict compares the provisions — subject, jurisdiction, actor, action breadth, voluntary conduct, logical structure, carve-outs — and it never returns a bare EQUIVALENT: LIKELY_EQUIVALENT is its ceiling, with the clauses to read attached. That is the one that catches the AI bills above, and roughly half the board resolves on provisions rather than on numbers.

An unknown on a material field returns REVIEW_REQUIRED, never a pass. INSUFFICIENT_DATA is a distinct verdict from LIQUID and must never render as one.

venue_conventions

Thirteen documented ways a venue API returns HTTP 200 while being wrong, each one found by losing time to it and each one pinned by a test. A sample:

• Polymarket's /book returns asks descending — asks[0] is the worst price.
• A YES bid at 0.10 is a NO ask at 0.90. One order, two books. Never add the ladders.
• data-api /trades ignores the asset parameter and returns the full tape. A five-row sample "confirms" the filter works.
• Trades are newest-first; a capped pull is recent history, so dividing by 30 days understates a busy market ~15×.
• Kalshi's orderbook is under orderbook_fp, and both sides are resting bids — the NO ask is 1 − best YES bid.
• Limitless caps limit at 25 and 400s above it, which reads as "0 markets found".

The meta-lesson, and the reason the tools refuse the way they do: a check that cannot run reads as a check that passed.

Honest status

Version 0.1.0. The verdicts are conservative by construction and the refusals are deliberate, but this is early software with no user base behind it yet. It measures and records; it does not recommend positions and it is not legal advice. Order transmission is off.

Requires Python 3.11+. Tested on 3.13.

Name collision: an unrelated PyPI package called clause (a Chinese NLP SDK) also installs a top-level clause module. Installing both into one environment will produce a broken hybrid. Run this server with uvx, pipx, or its own virtualenv — which is how MCP servers are usually run anyway.

MIT licensed.