hivemindunit/llmintel-mcp

An MCP server that tells your coding agent whether a model id is
safe to use.
LLMs are trained on a snapshot of the world and will confidently write gpt-4-32k into your code
long after it stops answering. This server gives the agent a live lookup for whether a model is
deprecated and when it stops working. It returns the replacement too. Answers are normalized across
OpenAI, Anthropic, Azure AI Foundry, AWS Bedrock, Google, and Cohere, and parsed from each
provider's own deprecation pages.
No API key, no signup. The catalog is public.
Install
Add it to any MCP host. The package runs straight from npm via npx.
Cursor
In .cursor/mcp.json:
Claude Code
Claude Desktop
Same shape as the Cursor block above, in claude_desktop_config.json.
Hosted endpoint (no install)
The same five tools are served over Streamable HTTP at https://llmintel.ai/v1/mcp. Hosts that take
a URL need no Node and no package:
The endpoint is stateless and read-only. It answers from the same catalog the npm package queries.
Tools
| Tool | Returns |
| --- | --- |
| check_model | Whether one model id is safe to use: lifecycle state, the retirement deadline in days, the replacement, and the source link. |
| list_retiring_models | What breaks in the next 90 days. Past-due models are listed first, then upcoming ones soonest-first. |
| suggest_replacement | The provider's own recommendation for what to move to. Falls back to same-provider active models when none was published. |
| search_models | Catalog search filtered by provider and lifecycle state. |
| recent_lifecycle_changes | The change feed across all providers, for questions like "what was deprecated this month". |
Example
> **You:** Before we ship this, check the model ids in src/agents/.
The agent calls check_model for each one and gets back:
Deadlines are always given in days, because a model cannot reliably judge whether 2026-07-30 is
soon.
Design notes
A failed lookup is never a safety verdict. If the catalog is unreachable, the tool returns an MCP
error and says so. An agent that read a network failure as "no deprecation found" would happily ship
a retired model id. A model that simply isn't tracked gets the same treatment: it returns "not in the
catalog, verify with the provider", never "OK".
Pass whatever string is literally in the code (gpt-4o, anthropic/claude-opus-4-1, azure/gpt-4o)
and it resolves to the canonical tracked model.
When the provider's own deprecation notice names a successor, that is what you get. Otherwise the
fallback list of same-provider active models is labelled as candidates to evaluate, so an agent can
tell the two apart.
Anything past its retirement date is broken now, so it gets its own heading instead of sitting in