apiwatch MCP server

apiwatch

Know when an API, model, or MCP server you depend on changes, and exactly which lines of your code it breaks.

Vendors retire models, rename MCP tools, and change API behavior on a schedule you don't control. Most of these changes are announced in a changelog nobody on your team reads. apiwatch reads them for you, and its GitHub Action points at the exact lines they land on, before the shutoff date.

• The feed: reviewed records of breaking, silent, and deprecation changes from OpenAI, Anthropic, Google, Stripe, Twilio, Notion, Atlassian, GitHub, xAI, and more. Every record cites the vendor's own page and is rechecked daily. Browse it at apiwatch.reignventures.co (Atom and JSON Feed included).
• The scan: finds the call sites each change lands on in your repo. It is token aware (gpt-4 does not match gpt-4o), follows constants, reads notebooks, and separates real uses from entries in model lists.
• The brief: for each hit, what changed, where, a suggested fix, and acceptance criteria you can hand to a coding agent.
• The probe: baseline a live JSON endpoint or an MCP server's tools/list and get told when it drifts, no spec required.

What it caught

The feed holds 140 reviewed changes, each checked against the vendor's own page.

• By vendor: Google 36, Stripe 30, OpenAI 26, Anthropic 17, Notion MCP 8, Notion 6, Atlassian Rovo MCP 4, GitHub MCP Server 4, xAI 4, Twilio 3, Greenhouse 1, MCP specification 1
• By severity: breaking 102, silent 26, deprecation 12

Next dates (everything in the next 180 days):

• Nov 30, 2026, Anthropic, breaking: Anthropic will retire Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) on the Claude API on November 30, 2026; requests to it will fail after that date.
• Nov 30, 2026, Anthropic, breaking: Fast mode for Claude Opus 4.8 is deprecated and will be removed on November 30, 2026; after that, claude-opus-4-8 requests with speed "fast" return an error
• Nov 30, 2026, OpenAI, breaking: OpenAI reusable prompt objects and the v1/prompts API shut down November 30, 2026
• Dec 1, 2026, Google, breaking: From December 1, 2026, Gemini API requests for the deprecated gemini-2.5-flash are routed to gemini-3.5-flash-lite, where prefilled model turns return HTTP 400.
• Dec 1, 2026, Google, breaking: From December 1, 2026, Gemini API requests for the deprecated gemini-2.5-pro are routed to gemini-3.8-flash, where prefilled model turns return HTTP 400.

GitHub Action

name: apiwatch
on: [push, pull_request]
jobs:
  scan:
    runs-on: ubuntu-latest
    permissions: { contents: read, pull-requests: write }
    steps:
      - uses: actions/checkout@v4
      - uses: ReignVentures/apiwatch@v1
        with:
          fail-on-breaking: "true"   # fail CI when a breaking or silent change touches your code
          comment: "true"            # keep one up to date report comment on pull requests
          only-changed: "true"       # on PRs, gate only on the files the PR changes

Every run writes a short report to the job summary: which changes touch your code, how severe, whether the date is past or coming, and links to each location. Add SARIF upload to see hits as code scanning annotations:

      - uses: github/codeql-action/upload-sarif@v3
        if: always()
        with: { sarif_file: apiwatch.sarif }

Tip: add a schedule: trigger (for example daily) so you hear about new changes even when nobody is pushing.

CLI

pip install git+https://github.com/ReignVentures/apiwatch
apiwatch scan . --fail-on-breaking --markdown report.md
apiwatch probe docs "mcp:stdio:npx -y @modelcontextprotocol/server-everything"
apiwatch probe docs mcp:https://example.com/mcp --header 'Authorization: Bearer $TOKEN' --fail-on-drift

Scans run locally with no network calls. Your code never leaves your machine or your CI runner.

To leave generated or copied files out of a scan, add a .apiwatchignore file at the repository root: one glob per line (agents/*.json), a trailing / for a directory (generated/), # for comments. Translation folders (locales/, i18n/) are skipped by default.

MCP server

AI agents can ask the feed directly while they work: does this model id or API have a published change, when, and what replaces it? One public URL, no install, no account: https://apiwatch.reignventures.co/mcp (Streamable HTTP, both the 2026-07-28 and the earlier initialize protocol eras).

claude mcp add --transport http apiwatch https://apiwatch.reignventures.co/mcp

Tools: check_model, check_identifiers, list_upcoming, search_changes, get_record. It answers only from the reviewed records the site publishes, stores nothing it is sent, and says "no record" (never "safe") for an id it doesn't know. Setup for other clients: https://apiwatch.reignventures.co/mcp-server

Two more tools, list_watched_mcp_servers and mcp_server_changes, answer from apiwatch's history of observed tool changes on some public MCP servers (https://apiwatch.reignventures.co/mcp-servers/): tools removed or renamed, new required parameters, rewritten descriptions, changed hints. They never return a server's own tool descriptions.

How records are verified

A record is published only after a reviewer that did not write it checks every field against the vendor's primary source, and an automated check confirms the quoted evidence is still on that page. The full rules are in docs/REVIEW.md. If a vendor edits or removes the page, the record comes off the site until it is reviewed again.

Found a change we missed, or a record that's wrong? Open an issue with a link to the vendor's page.

apiwatch Pro

Pro watches every repository in your GitHub organization each night, with nothing to add to your workflows, and opens an issue when a new change lands on your code. $19 per month. See how Pro works.

License

MIT. Built by Reign Ventures LLC.