Summary: OpenRouter launched an MCP server on June 25 that gives coding agents live catalog data, benchmark standings, pricing, documentation, and paid test inference from one remote endpoint. Eleven tools are available, with only send-message billable, and the recommended workflow is filter, benchmark, then run a real prompt before choosing. The server is a development aid rather than a production API replacement, and benchmark scores still need testing against a team’s own workload.
Calling a model is easy for a coding agent; choosing which one is what trips it up, since that call leans on whatever impressions training left behind. The market refuses to sit idle: releases keep landing, prices shift without warning, benchmark boards keep being redrawn, and the leader of half a year back may now cost more while trailing newer rivals. The MCP server OpenRouter launched on June 25 turns model selection from a memory exercise into an on-the-spot lookup: live model data, benchmark standings, pricing, and docs are one query away, and real test prompts can go straight to any candidate.
Why model choice depends on live data
Choosing a model has a seldom-counted cost: the evidence goes stale. Training data carries a cutoff date; the model market does not. One wider context window, one price cut, a provider reworking the routing it applies — any can invalidate a judgment that held up only last month. Standard fixes meant a person with a browser and the docs, a step that lives in the human and outside anything an agent can run.
The design answer: lookups become the agent’s job. One remote endpoint (https://mcp.openrouter.ai/mcp) is all it exposes; nothing needs installing on your side, and any MCP-speaking client connects at that single address. The catalog view coming back shows this very moment, not a picture preserved inside training memory.
Its effect on the workflow runs deeper than “a couple more tools.” Selection once played out entirely offstage — the human gathered data, reached a verdict, and typed it into a prompt. Now the intermediate steps (filtering, price shopping, spec checks) run as a string of calls the agent itself makes; the person only reads over the shortlist that made the cut plus the test results. The development assistant name fits: exploration speeds up, and the final say stays with the human.
Eleven tools, one of which costs money
The full set holds eleven of them, sorted into four kinds of work:
| Category | Tools | Purpose |
|---|---|---|
| Model catalog | list-models, get-model, list-model-endpoints |
Filter the list against conditions, inspect one model’s spec, weigh provider prices and latency against each other |
| Benchmarks and rankings | list-benchmarks, list-daily-model-rankings, list-app-rankings |
Scores from two outside evaluators, per-day usage boards, app-category boards |
| Live test inference | send-message, get-generation |
Deliver a trial prompt to any model, then confirm that call’s cost and the provider behind it |
| Documentation and credit | search-docs, get-credits, list-providers |
Keyword search over OpenRouter’s docs, a balance inquiry, the roster of providers |
The money line is sharp: across the whole set, send-message alone touches a bill; the remaining ten are read-only lookups. Benchmark figures come from Artificial Analysis and Design Arena, plus usage boards of OpenRouter’s own; a model name can also carry four possible suffixes — :online, :nitro, :floor, :free — for web-enabled lookup, faster performance, a lowest price, and the no-cost variant.
The heaviest hitter among the eleven is send-message: the agent sends one prompt to every shortlisted model and gathers back each response and its cost, plus the provider that ended up taking the request — all three laid out together before any decision. search-docs lets the agent dig through the API documentation unaided; nobody has to open a browser.
Hooking up the big three
Connecting any of the big three takes a line or two. Claude Code needs two commands: claude mcp add --transport http openrouter sets the target, then claude mcp login openrouter carries OAuth through. Codex CLI takes the same pair: codex mcp add openrouter --url https://mcp.openrouter.ai/mcp, then codex mcp login openrouter. Cursor means editing ~/.cursor/mcp.json and putting the same remote URL into the mcpServers block.
Claude Desktop along with the web version are the awkward cases: OpenRouter is missing from the connector directory, so you add a custom connector by hand, paste the remote URL in, and run OAuth to the finish. Some workplaces disable custom connectors outright, which closes this route — in a company setting, check with your IT team before trying.
The official walkthrough: from filtering to a live test
The official documentation lays out a procedure you can copy as-is. The goal: a model that extracts structured JSON well, keeps input prices low, and offers plenty of context. The agent filters with list-models, cross-checks list-benchmarks, and lands on google/gemini-3-flash-preview — quoted there at $0.10 per million input tokens and 138k of context. A real prompt then goes out over send-message to see structured output in action, instead of trusting the spec sheet alone.
The takeaway is the sequence, not the pick: filter, then benchmark, then test live. Specs and rankings narrow the pool; a test inference confirms. Run candidates against each other with the same script — one prompt to the three finalists, get-generation pulling each call’s price — and the verdict sits on three supports: quality of the returned output, spend per call, and the provider that wound up handling it. The third gets left out most, yet steady routing plus throughput that holds often matter more after launch than the sticker suggests.
Security design and spending ceilings
Finishing OAuth grants a dedicated-purpose key: seven days of life and a ten-dollar default spend cap. Your dashboard keeps it in a section of its own, separate from your other keys; revoke it at any moment, and the cap figure is editable. That default deserves credit — the agent’s key becomes a prepaid, use-and-toss card instead of a standing credential.
The bottom line stays where it was: keys that run production never go near a coding agent. Test calls use the purpose-made key; production-bound traffic rides the real API. Two separate paths keep an incident small and contained.
A development helper, not an API replacement
On its own role, OpenRouter is blunt: this MCP server works as a development aid for coding agents; an app’s production-grade traffic still goes straight to the OpenRouter API. Use MCP for experiments and research during the exploring stage; when production starts, nail down the final model and the routing it uses in code. That’s where the divider belongs.
A practical switching rule: when the concern moves from what model would work best to whether this one can hold steady, leave MCP for the API. The first is about exploring options; the second is one of dependability, which needs monitoring you run, retries you control, and a ledger of costs you keep — none of it comes from a directory search.
Benchmarks, finally, deserve guarded interpretation: Artificial Analysis and Design Arena measure workloads run in public, and that report card will not transfer cleanly to your mix of tasks. Keep the private evaluation set you built, and judge cost, error rates, and format-following together — MCP delivers the shortlist; the final vote is still yours to cast.