costwright — agent budget analysis

A local MCP server that statically analyzes Python LLM-agent repos for worst-case token budgets and can issue signed budget certificates.

What you can do

Analyze a local Python repo (LangGraph, CrewAI, OpenAI Agents SDK) by AST without running it, returning pass/fail and counts of certifiable, default-dependent, non-certifiable and runaway units, plus LLM calls lacking token caps. Issue a signed, publicly logged budget certificate for a repo and get its id, signature and verification link. Verify an existing certificate by id, checking validity, expiry, revocation and signature. Fetch the active Ed25519 public keys for offline verification.

When to use it

Use it before deploying an agent workflow, to catch missing token caps and runaway loops the way a type check catches type errors. Useful when you want a verifiable, signed record of a workflow's worst-case budget ceiling.

Requirements

Runs locally via npx (Node.js). The check and certify tools need COSTWRIGHT_API_KEY, sent as an X-RapidAPI-Key header by default or as a Bearer token when COSTWRIGHT_DIRECT=1. Verify and pubkey need no key. Network access to the hosted costwright API is required.