Dependabot for the APIs your code calls.
AI models retire every few weeks. API versions sunset. Vitasta finds where your code uses one that's going away, tells you the date, and opens the pull request that fixes it.
Or try it on your repo: npx @vitasta/scanner scan .
$ npx @vitasta/scanner scan . Scanned 412 files · 37 API calls · 9 models ● broken src/summarise.py:18 claude-3-5-sonnet-20241022 retired 28 Oct 2025 → move to claude-sonnet-5-5 ● 28 days src/agent.ts:42 gpt-5 retires 11 Dec 2026 → move to gpt-6.1-sol ● broken infra/models.yml:9 anthropic.claude-3-haiku-20240307-v1:0 retired on Bedrock 10 Sep 2026 (Anthropic API: 20 Apr) Fix PR opened: vitasta/swap-gpt-5
From "model not found" in production to a merged fix.
Vitasta reads your code with real parsers, not text search, so it knows the difference between a model you call and a model you mention in a comment.
Finds every API call
SDK calls, raw HTTP, model IDs in config and env files, pinned API versions, Bedrock and Vertex model IDs. It follows clients across files.
Checks a live change feed
OpenAI, Azure, Bedrock and Vertex retire the same model on different days. Vitasta knows which one you're on.
Opens the pull request
Swaps the retired model for its replacement and removes dead parameters, then re-parses and re-scans the result. It never touches a branch a human has edited.
Wherever you work.
Same scanner, same feed. Use one, or all four.
CLI
Scan any repo or folder on your machine.
GitHub Action
Fails a PR that adds a call to something retired or due within 30 days. Older problems don't block.
GitHub App
A check on every PR, one tracking issue per repo, and fix pull requests.
Coding agents
Claude Code, Cursor and Codex ask Vitasta over MCP which models are current before they write the call.
What it tracks today.
AI models first, because they change fastest. More platforms every week. Add your own internal APIs in a config file.
Also mapped, with change tracking on the way: xAI, DeepSeek, Perplexity, Twilio, Slack, HubSpot, Notion, Intercom, Google Maps, SendGrid, PayPal and Zoom. OpenAI-compatible endpoints are credited to the real provider.
Languages: TypeScript, JavaScript, Python and Java in depth. Go, Ruby, PHP and C# in basic mode. Plus YAML, JSON, TOML, .env files and notebooks.
Your code stays yours.
The first question every team asks, answered plainly.
✓Local tools send nothing
The CLI, the Action and the MCP server run on your machine or your CI. They never send your code anywhere. No LLM reads it.
✓The App never runs or keeps it
It fetches one commit into a private temp folder, parses it, posts findings back to the same repo, and deletes the folder. No database of your code, no copy, no cache.
✓Secrets are stripped
Snippets are one line. Anything that looks like an API key, token, password or private key is removed before it's printed, posted or logged.
What it gets right, and what it misses.
We'd rather you trust the numbers than be impressed by them.
Strong
- SDK and HTTP calls, followed from the client to the call.
- Model settings in code, config and env files. Models that are only offered in a list or picker are marked as such and never block a PR.
- Separate dates per platform (Azure, Bedrock, Vertex).
- Clients pointed at a proxy or relay are flagged, not guessed.
Still missing
- Model names in other places than settings, lists and calls, such as some TypeScript types.
- Go, Ruby, PHP and C# are string-level for now.
- Test mocks are reported, marked as tests. They never block a PR.
- Providers that don't publish dates can't be tracked until they do.
Free for public repos. Join the beta for private ones.
We're onboarding a few teams at a time. Tell us where to reach you.