Modelmeter MCP server

Modelmeter

The changelog of record for the AI stack. Live at modelmeter.xyz.

What this is

A machine-readable, source-verified record of everything that changes in the model layer: deprecations with sunset dates and migration targets, silent model swaps behind stable IDs, price changes, releases — plus three years of pricing history and cost tooling. Every event carries its sources with verbatim quotes (most with archive snapshots), a verification status (unverified → human-reviewed verified), and corrections are never silent — a correction event points at what it supersedes.

Built primarily for AI agents to call programmatically, secondarily for humans through a dashboard. Reach it three ways: plain GET URLs returning JSON, a remote MCP server any tool-calling agent can connect to, and a visual dashboard built on the same endpoints. The wedge: clean URLs, stable JSON contracts, an OpenAPI spec, and an llms.txt discoverability file. Most sites in this space are JS-heavy SPAs that agents can't use; Modelmeter is built the other way around.

Endpoints

• GET /estimate?model=X&input=N&output=M — token cost for a hypothetical call, including upstream-markup comparison for reseller models
• GET /models — current model catalog, with filters
• GET /model?id=X — unified card for one model: normalized pricing, capabilities (context window, vision, reasoning, tags), availability, reseller markup, and a price-history summary (launch vs. current price, % change, all-time low/high)
• GET /check?models=gpt-4o,claude-sonnet-4-6 — is my stack okay? Verdict per model: scheduled retirements with days-remaining and migration targets, past retirements, breaking changes, or a clean bill
• GET /deprecations — per-model retirement rows: which model dies when, with runway and migration target
• GET /feed.xml / GET /feed.json — RSS 2.0 / JSON Feed 1.1 of breaking and action-required events
• GET /events — the changelog of record: deprecations, price changes, launches and surrounding market events, each with severity (breaking | actionrequired | informational), announcedat/effective_at dates, sources with quotes, and verification status. Filters: provider, type, model, severity, since, until, status
• GET /history — historical pricing time-series per model, with filters
• GET /pricing.json — raw pricing snapshot
• GET /events.json — raw events snapshot
• GET /history.json — raw historical time-series

See public/openapi.yaml for the full spec.

GitHub Action

Fail CI when a model your repo depends on is deprecated, sunsetting, or already retired:

- uses: modelmeters/modelmeter/actions/check@main
  # with:
  #   models: "gpt-4o, claude-sonnet-4-6"   # optional — otherwise the repo is scanned for known model ids
  #   fail-on: breaking                      # breaking (default) | retired | none
  #   warn-days: "120"                       # sunsets within N days fail; beyond N warn

The action scans your repo for known model ids (catalog + every model named in a deprecation event, dashed and dotted spellings), calls /check, annotates the exact file:line of each affected model, writes a job-summary table with migration targets, and exits nonzero per your fail-on policy.

MCP server

Modelmeter is also a remote MCP server, so any tool-calling agent (Claude, Hermes, …) can call it natively over the Streamable HTTP transport — no auth, no install:

https://modelmeter.xyz/mcp

Tools: checkmodeldependencies, listdeprecations, estimatecost, getmodel, listmodels, getpricehistory, listevents. Start with listmodels to discover ids, then getmodel or estimatecost. The MCP tools return the same data as the REST endpoints above — use whichever your runtime prefers.

Shortened. The full README is on GitHub.