OpenRouter has released an official OpenRouter MCP server — announced on 25 August 2026 — that puts its live model catalogue, per-provider pricing and latency data, third-party benchmarks and even test inference directly inside Claude Code, Cursor, Codex CLI, OpenCode and Claude Desktop, so you can pick the right model for a job without ever leaving your editor. For anyone who builds with AI daily, that turns model selection from a browser-tab ritual into a thirty-second question you ask your own coding agent.
📺 Watch: Hermes Agent: NEW SEO Link Building MCP!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
Model choice is one of those decisions I make dozens of times a week — which brain for the research agent, which for the bulk rewrite job, which for the client build — and the tab-hopping tax on that decision is real. So an MCP server maintained by the routing layer itself, feeding live data straight into the tools I already work in, got my attention immediately. Here is what it exposes, how to wire it in, and where it genuinely changes the workflow.
What the OpenRouter MCP Server Actually Exposes
MCP — the Model Context Protocol — is the plug standard that lets AI tools call external services as if they were built in. According to OpenRouter's announcement, its server exposes a genuinely useful spread of data:
- The live model catalogue — every model on the platform with its metadata, straight from the source rather than a stale list.
- Per-provider pricing and latency — the numbers that decide whether a "cheap" model is actually cheap once speed matters.
- Third-party benchmarks — the announcement names Artificial Analysis and Design Arena, so you get outside scores, not just vendor claims.
- Usage rankings and task classification market share — what people actually route to models for, which is often more honest than any leaderboard.
- Documentation search — OpenRouter's docs, queryable from inside the editor.
- Test inference — a send-message tool that lets you fire a real prompt at a candidate model and read the answer without switching context.
That last one is the sleeper feature. Reading a benchmark is research; sending your actual prompt to three candidate models and comparing the answers in-place is a decision.
Setting It Up in Claude Code, Cursor and Codex
Setup is one command per tool. Per the announcement, in Claude Code you run: claude mcp add --transport http openrouter https://mcp.openrouter.ai/mcp — then authenticate with: claude mcp login openrouter. In Codex CLI it is the same shape: codex mcp add openrouter --url https://mcp.openrouter.ai/mcp, followed by codex mcp login openrouter. In Cursor you add an entry for the server pointing at the same https://mcp.openrouter.ai/mcp address in your MCP configuration file, and OpenCode and Claude Desktop are supported the same way.
Once connected, the workflow is conversational. Ask your coding agent which model currently has the best price-to-latency ratio for long-context summarisation, let it pull live pricing and benchmarks through the server, then have it test your actual prompt before you commit. If you are new to wiring MCP into a coding setup, my free Claude Code guide covers getting the base tool running without paying for it.
Want my exact model-picking workflow — the decision tree, the prompts and the routing setup that keeps quality up and costs down? I teach the whole system inside AI Profit Boardroom, with five live calls a week → Steal my model stack
Why an MCP Server for Model Data Matters
The obvious objection: you could just open the OpenRouter website. True — and that is exactly why this matters. Data you have to go and fetch gets checked occasionally; data your agent can fetch gets checked every time. Model pricing moves constantly, new checkpoints land weekly, and the "best" model for a task changes more often than anyone updates their defaults. Putting live rankings and pricing one tool-call away means your agent can re-justify its own model choice mid-project.
It also fits a pattern I keep hammering: the moat is not the model, it is the system around the models. That is the entire thesis of the Agent OS we built and run — treat models as swappable parts, keep the workflows, memory and tooling as the permanent layer. A live model-data feed inside the editor is precisely the kind of part-swapping infrastructure that thesis needs. And on the evaluation side, this is why we run Goldie Bench, our own benchmark, on every serious release: third-party scores like the ones this server surfaces are the starting point, and testing on your own work is the finish line.
📺 Watch: This All-in-One AI Agent Manager Is SCARY Good
OpenRouter MCP Server vs the Hermes MCP Server
A naming clash worth untangling, because I have covered MCP servers before and they do opposite jobs. The Hermes MCP server exposes your Hermes agent as a tool that other AI clients can call — it makes your agent callable. The OpenRouter MCP server exposes OpenRouter's model data as a tool your agents can call — it makes the model market queryable. One is a worker you plug into your tools; the other is a reference desk your tools consult. In a serious setup you run both: Hermes doing the work, and live model data informing which brain the work runs on.
It also pairs naturally with the rest of the OpenRouter platform. The Fusion API answers "what if I asked several models at once?", and my Hermes agent OpenRouter setup covers the day-to-day wiring of an agent to the routing layer. The MCP server is the missing piece between them: the thing that tells you which models deserve a seat on the panel in the first place.
What I Would Ask It First
To make this concrete, here are the questions I would put through the server in the first week, because they map to decisions that actually move money. Which models under a given output price currently score well on the third-party coding benchmarks? Which provider serves my default model with the lowest latency right now, and has that changed since I set it? What are people actually routing translation, extraction or long-context summarisation to, according to the usage rankings? And before any bulk job: send this exact prompt to the two cheapest candidates and show me both answers side by side.
None of those questions is new — what is new is that each one used to cost a browser session and now costs a sentence. When checking becomes that cheap, you check more often, and defaults stop silently rotting. That is the entire value proposition in one habit: interrogate the market before you spend, every time, because the tool made it effortless.
OpenRouter MCP Server: Quick Reference
| Detail | What the announcement says |
|---|---|
| Announced | 25 August 2026 |
| Server address | https://mcp.openrouter.ai/mcp over HTTP transport |
| Data exposed | Live model catalogue, per-provider pricing and latency, third-party benchmarks, usage rankings, docs search |
| Test inference | send-message tool for firing real prompts at candidate models |
| Supported clients | Claude Code, Cursor, Codex CLI, OpenCode, Claude Desktop |
| Claude Code install | claude mcp add --transport http openrouter, then claude mcp login openrouter |
OpenRouter MCP Server FAQs
What is the OpenRouter MCP server?
It is an official Model Context Protocol server from OpenRouter, announced on 25 August 2026, that gives MCP-compatible tools live access to the model catalogue, per-provider pricing and latency, third-party benchmark scores, usage rankings, documentation search and test inference.
Which tools can use it?
The announcement covers setup for Claude Code, Cursor, Codex CLI, OpenCode and Claude Desktop. Any MCP client that speaks HTTP transport points at the same server address.
Can I actually run prompts through it?
Yes — the server includes a send-message capability for test inference, so you can trial a candidate model on your real prompt from inside your editor before committing to it.
Is this the same thing as the Hermes MCP server?
No. The Hermes MCP server makes your own agent callable by other tools; the OpenRouter MCP server makes OpenRouter's live model data callable by your tools. They complement rather than compete.
Verdict: Small Ship, Big Workflow Change
The OpenRouter MCP server is not a new model or a flashy capability — it is plumbing, and that is exactly why I rate it. Model selection is a recurring, margin-affecting decision, and this moves the data for that decision into the place where the decision actually happens. One command to install, and your coding agent goes from working with a hard-coded model string to reasoning about the live market. That is the kind of quiet upgrade that compounds every single day you build.
If you want every new drop like this turned into a working money system — not just news — the AI Profit Boardroom is where I break down what to install, what to skip and how to profit from it, live every week → Come build with us
