authority-risk-mcp

A read-only MCP server that tells an AI agent who can change a smart contract, how fast, and how strong the setup is, before the agent depends on it. It reads the files that the Authority Risk Oracle Watch publishes every six hours, and checks them against the Watch's signed manifest before it answers.

No account, no key, no wallet, no transaction. The scores are published on testnet oracles for contracts of the networks named in each answer. A score is a reading of public data taken at the time shown: not an audit and not investment advice.

Install

With uvx (Python 3.9 or newer, nothing else):

{ "mcpServers": { "authority-risk": { "command": "uvx", "args": ["authority-risk-mcp"] } } }

With npx (needs Node 18 or newer and python3 3.9 or newer on the PATH; the npm package is a thin launcher around the same Python server):

{ "mcpServers": { "authority-risk": { "command": "npx", "args": ["-y", "authority-risk-mcp"] } } }

Put the block in claude_desktop_config.json (Claude Desktop) or .cursor/mcp.json (Cursor), then restart the client. With pip: pip install authority-risk-mcp, then authority-risk-mcp.

The four tools

Every answer is JSON with a status:

• ok: the answer is read.
• not_tracked: the Watch does not track this contract. That says nothing about its safety.
• not_read: a file the answer needs is missing, could not be fetched, or did not match its signature (integrity_failed: true).

A value the Watch did not read is null and listed in not_read. It is never 0 and never "ok". unread_part means some contracts of a protocol were read and others were not: that is not a pass.

Example call (what a client sends):

{"jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": {"name": "getScore", "arguments": {"chain": "robinhood", "address": "0x1f7d7550B1b028f7571E69A784071F0205FD2EfA"}}}

Chains: robinhood, ethereum-l1, arbitrum, base, tempo, plasma, monad, hyperliquid, solana.

What is checked

The Watch signs a manifest (the sha256 of every published file and the head of its hash-chained event log) with an Ed25519 key. This server:

1. fetches signature.json and verifies the signature with a key it has pinned (11baca3e...f409); an unknown key, an unsigned manifest or a bad signature is refused;
2. fetches each file it needs and compares its sha256 with the manifest; if the first source (raw.githubusercontent.com) serves a different copy, it tries the second (jsDelivr); a mismatch everywhere is refused;
3. compares the head of the event log with the manifest when it can.

Each answer carries an integrity block: whether it verified, which key signed, how old the snapshot is (snapshot_stale after 72 hours), and the event-log check. A file that does not match is never a value.

What this does not prove: the Watch's workflow holds the signing key, so a malicious change to the workflow itself could sign anything (that change would be visible in the repository history). The pinned key protects against a different server, a compromised mirror or a tampered download, not against the maintainer.

Options

authority-risk-mcp [--data-dir DIR] [--base-url URL] [--trust-key HEX] [--allow-unsigned] [--allow-unpinned] [--max-age-hours N]

• --data-dir DIR: read a local clone of the watch-data branch, offline (git clone --depth 1 -b watch-data https://github.com/RealSpap/authority-risk-oracle-showcase.git). The folder is checked against its own signature.json when it has one; without it the answers say verified: false. Nothing touches the network in this mode.
• --trust-key HEX (or AUTHORITY_RISK_TRUSTED_KEYS, comma separated): accept a manifest signed by another key, for a rotation announced in the repository.
• Network mode contacts only raw.githubusercontent.com and cdn.jsdelivr.net (or the https --base-url you give) and writes verified files to a private temporary folder that is removed when the server stops.

Licence

MIT. Source: packages/authority-risk-mcp in the repository. The tool modules are copied byte for byte from the repository's scripts/ folder at build time.