Coding agents write model calls quickly, but model selection runs on training-time memory. New models keep shipping, prices move without notice, and benchmark scores keep refreshing — the “best model” from six months ago may be overpriced and outperformed today. The MCP server OpenRouter launched on June 25 turns this from a memory problem into a query problem: the agent reads live model data, benchmark rankings, pricing, and docs directly, and can even send test prompts to candidate models.
Why model selection needs live data
The hidden cost in model selection is the freshness of your evidence. Training data has a cutoff; the model market does not. An expanded context window, a price cut, or a provider reshuffling its routing can invalidate last month’s conclusion. The traditional fix is a human opening a browser and checking docs — but that step lives outside the agent’s workflow.
The design intent of this MCP server is to hand the query to the agent. It exposes a single remote endpoint (https://mcp.openrouter.ai/mcp), installs nothing locally, and every MCP-capable client connects through the same address. What the agent gets is the catalog as it stands right now, not a snapshot frozen into training memory.
The workflow impact runs deeper than “a few more tools.” Selection used to happen outside the agent: a human researched, concluded, and wrote the conclusion into a prompt. Now the middle steps — filtering, price comparison, spec checks — move into the agent’s tool-call chain, and the human reads only the shortlist and the test results. That is also why OpenRouter calls it a development assistant: it changes the speed of exploration, not the ownership of the decision.
Eleven tools, only one costs money
The toolbox holds eleven tools across four kinds of work:
| Category | Tools | Purpose |
|---|---|---|
| Catalog lookup | list-models, get-model, list-model-endpoints |
Filter models by condition, read a single spec, compare per-provider price and latency |
| Rankings and benchmarks | list-benchmarks, list-daily-model-rankings, list-app-rankings |
Third-party evaluations, daily usage rankings, app-category rankings |
| Test inference | send-message, get-generation |
Send a test prompt to any model, inspect per-call cost and provider |
| Docs and account | search-docs, get-credits, list-providers |
Full-text search of OpenRouter docs, check credit, list providers |
The billing boundary is crisp: only send-message spends money; the other ten are read-only queries. Benchmarks come from Artificial Analysis and Design Arena, alongside OpenRouter’s own usage rankings; models accept the :online, :nitro, :floor, and :free suffixes for web access, speed, lowest price, and free variants respectively.
The sharpest tool is send-message: the agent sends the same prompt to candidate models and gets back the response, the cost, and the provider that actually served it — a side-by-side comparison before committing. search-docs lets the agent look up API documentation itself instead of waiting for a human to open a browser.
Setup in three environments
Three mainstream environments are one-liner setups. Claude Code takes two commands: claude mcp add --transport http openrouter with the endpoint, then claude mcp login openrouter to complete OAuth. Codex CLI takes two commands as well: codex mcp add openrouter --url https://mcp.openrouter.ai/mcp, then codex mcp login openrouter. Cursor adds the same remote URL under the mcpServers block in ~/.cursor/mcp.json.
Claude Desktop and Web are the awkward case: OpenRouter is not in the connector directory yet, so you add a custom connector manually, paste the remote URL, and finish OAuth. Some organizations disable custom connectors entirely, so whether this path exists depends on policy — ask IT before trying in a corporate environment.
The documented example: list, filter, then test
The announcement includes an example worth copying verbatim as a process. The need: a model suited to structured JSON extraction with low input pricing and a long context. The agent first filters with list-models by condition, cross-references scores with list-benchmarks, and lands on google/gemini-3-flash-preview — the numbers in the documented example are $0.10 per million input tokens and a 138k context. It then sends a real test prompt with send-message to confirm structured output behavior, rather than ordering off the spec sheet alone.
What deserves remembering is the sequence, not the conclusion: filter, evaluate, then actually test. Specs and rankings narrow the field; test inference casts the deciding vote. Side-by-side comparison works the same way — send one prompt to three candidates, pull each cost with get-generation, and the decision has evidence in three dimensions: output quality, per-call cost, and the provider that actually served it. The third dimension gets ignored most often, yet routing stability and throughput frequently matter more to the live experience than the sticker price.
Security boundaries and billing
OAuth produces a dedicated key: a seven-day expiry and a default ten-dollar spend cap. The key is displayed separately from your other keys on the dashboard, can be revoked at any time, and the cap is adjustable. That default deserves credit — it shrinks “the key an agent holds” into a disposable temporary card.
The principle still stands: do not hand your primary production key to a coding agent. Test calls go through the dedicated key; production traffic goes through the real API. Keep those two paths separate and the blast radius stays contained.
What it is and is not
OpenRouter states the positioning plainly: the MCP server is a development aid for coding agents, and your application’s real traffic should still call the OpenRouter API directly. Use MCP to research and experiment during exploration; once you reach production, encode the chosen model and routing in code. That is the correct dividing line.
A practical switch point: the moment the question changes from “which model fits” to “is this model dependable,” leave the MCP and move to the API. The first is exploration; the second is a reliability question — what you need is no longer a catalog query but your own monitoring, retries, and cost ledger.
The other thing to police yourself is the interpretation of benchmarks. Artificial Analysis and Design Arena scores answer public workloads, not your task distribution. Keep your own test set and compare cost, format adherence, and error rates together — the MCP hands you a shortlist; the final vote stays with you.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.