FactGrid
Reference data agents need before they write code. Dependency upgrade advisories for npm and PyPI, developer API schemas across 1,097 services, and electronic component specifications — over MCP, Ed25519-signed, with the evidence attached.
MCP endpoint: https://factgrid.co.uk/mcp · Agent-readable docs · factgrid.co.uk
Stop an agent shipping a broken dependency upgrade
Your agent wants to bump chalk from 4.1.2 to 5.0.0. Before it edits package.json, it asks what breaks:
Via MCP, the same question is one tool call:
Every finding carries evidenceField, evidenceBefore and evidenceAfter — the exact registry metadata field and its values either side of the change. An agent can verify the claim against npm or PyPI directly rather than taking our word for it.
Connect
Streamable HTTP. No install, no package to pull.
Manifest: /.well-known/mcp.json
Get a key — no signup, no card:
Tools
Catalogue
Live counts, free and unauthenticated: /api/v1/stats · /api/v1/packages/stats
Pricing
50 free queries per key per day. No card, no signup.
Pay from prepaid credit, or per call in USDC over x402 — no account and no human in the loop. Live pricing: /api/v1/pricing
Misses are never billed. A query we can't answer returns found: false, billed: false and the nearest matches, at no charge. Cached revalidations (HTTP 304) and rate-limited requests are never charged either.
What we don't claim
This section exists because the alternative is worse.
The API breaking-change tier is offline. An earlier version of this project published 8,150 breaking-change diffs for third-party APIs. A spot-check found a 100% false-positive rate — they were computed from mis-ordered specification versions. We withdrew all of them. That tier still returns no data and bills nothing, and it will stay that way until we can show the output holds up against publishers' own changelogs.
That failure is why the package engine is built the way it is:
breaking is reserved for a removal or tightening provable from declared metadata — an entry point gone from an exports map, a runtime floor raised, a release the publisher yanked.
- Additions can never be breaking. There is a test asserting exactly that.
- Where a version range can't be parsed, the engine stays silent rather than guessing.
- Across
urllib3's 103 consecutive stable releases it emits 7 advisories, not thousands. The conservatism is the feature.
Package advisories cover publisher-declared registry metadata, not behavioural changes. If a maintainer silently changes what a function returns without touching package.json, we will not see it — read the changelog. We say so rather than implying coverage we don't have.
Mainnet x402 is configured and advertised but has never been exercised with real money. It settled against two independent facilitators on testnet. We won't call it proven until a real settlement lands.
Accuracy
Every response is Ed25519-signed; the public key is at /.well-known/jwks.json. Records are source-cited. API schema data derives from APIs.guru under CC-BY 4.0.
Accuracy is best-effort, not guaranteed — see /terms. Found an error? POST /api/v1/report-error — free, no auth.