hyperroute-mcp

The official Model Context Protocol server for HyperRoute.

HyperRoute is a router for AI agents. Give it a task and it picks the best external tool for that task — measured, not advertised — then runs the tool for you with your own key held server-side, and learns from how it went. This MCP server is how a coordinator agent (Claude Code, Codex, Goose, Cursor, LangGraph, …) drives it:

recommend → onboard a key → execute the tool server-side → report the outcome

It talks to the router only over its public HTTP API and holds no product logic of its own.

Why route at all

An agent with 100 tools bolted on has a context problem and a quality problem. HyperRoute replaces both with one verb: your agent learns recommend, and HyperRoute decides which of hundreds of tools actually answers this task, whether you can already do it better yourself, and what it will cost.

• Measured, not advertised. Every capability score is backed by real graded probes you can inspect (describe(tool_id, ["evidence"])).
• Your keys never leave the server. You connect a key once; HyperRoute runs the tool with it and returns only the result. The key is never sent to your agent, never logged.
• It tells you when NOT to route. If nothing beats what your coordinator already does, the verdict is use_native — do it yourself. That only works if the server knows which coordinator it runs inside; see Declaring your coordinator.

Install

pip install hyperroute-mcp

Or with pipx, so the command is always on your PATH regardless of which virtualenv is active — which is what MCP clients need, since they launch the server themselves:

pipx install hyperroute-mcp

From source, for development:

git clone https://github.com/HyperRouteAI/hyperroute-mcp
cd hyperroute-mcp
python -m venv .venv && source .venv/bin/activate
pip install -e .

Requires Python ≥ 3.10.

Add it to your coordinator

Claude Code

claude mcp add hyperroute -- hyperroute-mcp

If you installed into a virtualenv rather than with pipx, hyperroute-mcp is only on your PATH while that venv is active — and MCP clients launch the server themselves, outside your shell. Give them the absolute path in that case: /path/to/.venv/bin/hyperroute-mcp.

OpenCode — copy opencode.json into your project. OpenCode is bring-your-own-model, so the server can't infer what you're running from the client name alone: add an "environment" block setting HYPERROUTE_COORDINATOR (or HYPERROUTE_NATIVE_TOOLS) to match the model you actually point it at, or HyperRoute will have no baseline for you. AGENTS.md carries the operating loop and the methodology — drop it in so the agent can both act correctly and explain how the routing works.

Any MCP client (mcp.json / claude_desktop_config.json / equivalent):

{
  "mcpServers": {
    "hyperroute": {
      "command": "hyperroute-mcp"
    }
  }
}

No base URL to set. The package talks to https://hyperroute.io out of the box; you only set HYPERROUTE_BASE_URL to point it somewhere else — a local dev router, or your own self-hosted instance:

"env": { "HYPERROUTE_BASE_URL": "http://localhost:8077" }

Then just ask: "Use HyperRoute to find the best tool for searching recent papers, connect my key, and run it." The agent calls recommend → connect_info → onboard → execute on its own.

Authenticate once

recommend and browsing are public — no account. Connecting keys and running tools need one.

Preferred: mint a personal access token at hyperroute.io and hand it to the use_token tool (or set HYPERROUTE_API_KEY). Your password never enters the conversation.

The token is then cached on disk (~/.hyperroute/token.json, mode 0600, keyed by router URL), so every new MCP session restores your login silently. You are asked to authenticate again only if the router invalidates the token. A full inline register → verify email-code flow is also available for headless use.

Declaring your coordinator

HyperRoute compares external tools against what you can already do. That baseline is the set of coordinators that are free to you — and it is empty by default, because the router never assumes you have one. An MCP server that does not declare itself gets an external tool recommended for every task, including tasks the coordinator does better itself.

This server declares it for you. It reads the MCP client identity your coordinator sends on connect and maps it to the coordinator HyperRoute models (claude-code → claude_code, …). Check what it resolved with the session_info tool: if native_tools is empty, set it yourself.

"env": {
  "HYPERROUTE_COORDINATOR": "claude_code",       // or codex / cursor / goose / …; "none" disables
  "HYPERROUTE_HELD": "anthropic_max_5x"          // subscriptions you already pay for → priced at $0
}

HYPERROUTE_NATIVE_TOOLS takes exact tool ids if you want to pin one model variant instead of the whole product family.

Configuration

Tools

The wire is deliberately lean

recommend answers with a compact table, not a catalog dump:

session: s-6d6c5a95f9f84d9a
verdict: interpose
refine:  freshness, cited_references, source_quality

  tool              name                        price  use        why
→ opencitations     OpenCitations Index         free   ready      highest-ranked: capability 0.81 …
  semantic_scholar  Semantic Scholar Graph API  free   needs_key  lower capability (0.75 vs 0.81).

confidence: med (on the pick)
act: execute('opencitations', <query>)

Everything else — descriptions, per-plan pricing, per-facet breakdowns, the probe evidence behind a score — is pulled for the one tool that matters via describe. That keeps a route roughly an order of magnitude cheaper in tokens than shipping the full object on every call.

The use column is the whole auth story: ready (run it) · needs_key (connect first) · native (do it yourself) · soon (not runnable server-side yet).

Your own tools

Tell your agent "I have my own web search, always use it for research" and it calls declare_my_tool with your words. HyperRoute maps them onto named capability regions and, from then on, routes to your tool for that kind of work — in every session, with a use_own verdict — instead of ranking a catalog tool. Your agent runs it with the access it already has; HyperRoute never executes it and holds no key for it.

Two things it deliberately does not do:

• It is scoped. Outside the region you declared, your tool is not in the ranking at all. A declaration is never a blanket override.
• It is unscored. HyperRoute has never tested your tool, so it carries no capability number and never pretends to. It wins because you said so.

my_tool_report later shows your own outcome record per capability, beside whether HyperRoute holds tested alternatives there. If you want a better-scoring catalog tool to be able to displace yours in some region, switch that tool to stance="benchmarked".

Your preferred tools

Tell your agent "I like Kagi — whenever it's even remotely acceptable, use it" and it calls prefer_tool. Unlike your own tools, a preferred tool is one HyperRoute already has and scores; the preference only changes which acceptable tool wins: it is served whenever it clears the capability bar for the task and sits within your margin (default 0.10) of the best-scoring tool. Beyond the margin, or under the bar, the ranking is served as it would have been and the ranking's preferred: line says your tool was passed over and by how much — it never disappears silently. A hard requirement you set (a must_be compliance check) still excludes it: your constraints outrank your bias.

preferred: Kagi Search — served: 0.08 behind Brave Search on capability, within your 0.10 margin

Two-pass refinement

Pass 1 always returns a usable ranking. The refine: line names the unset preferences that would reorder these candidates; fill the relevant ones and call recommend again with facets for a personalized result. Durable constraints (GDPR, a budget cap) belong in set_preferences instead — stored once, applied to every future route.

Development

pip install -e ".[dev]"
pytest
ruff check .

The suite is fully offline — the router is faked, so no network and no real account are touched. Set HYPERROUTE_BASE_URL to run against a local router instead of the hosted one.

License

MIT — see LICENSE.