emissary-router 0.0.6

Pure-Python Claude Code routing gateway

Emissary Router

Emissary Router is a local Claude Code gateway that routes each request to a supported model, keeps provider-specific caching intact, and records lightweight cost/cache telemetry.

Install

pip install emissary-router

With uv: uv pip install emissary-router.

This installs the er command. If a global install is blocked (an "externally-managed environment"), install inside a virtualenv — or use an isolated installer like pipx install emissary-router.

Then set up config and API keys:

er init

er init creates ~/.emissary-router/config.json and prompts for your keys (it skips any already in your environment), writing them to ~/.emissary-router/.env. At any prompt press Enter to skip that key and set it later — or, when re-running, to keep the current value. Run it again any time to change a key. You can also just export the keys instead:

export EMISSARY_ROUTER_API_KEY=...
export ANTHROPIC_API_KEY=...
export OPENROUTER_API_KEY=...

Don't have an EMISSARY_ROUTER_API_KEY yet? Sign up at withemissary.com and create one (Dashboard > Settings > Credentials). See API keys for where each provider key comes from — you only need the ones your enabled models use.

Then run Claude Code through the router:

er code -- [claude args]

er code starts the local gateway automatically if it is not already running. The gateway keeps running after Claude Code exits; stop it with:

er stop

Installing from a clone instead? Run bash install.sh (editable install), then er init.

Dashboard

er code and er start open a local dashboard in your browser showing cost savings, recent requests, and per-session usage — plus a Settings tab to toggle models live:

http://127.0.0.1:8788/dashboard

It stays up while the gateway runs, so reopen the URL any time. See Dashboard.

Supported Models

Toggle models in ~/.emissary-router/config.json:

{
  "models": {
    "claude-sonnet-4.6": { "enabled": true, "provider": "anthropic" },
    "claude-haiku-4.5": { "enabled": true, "provider": "anthropic" },
    "gemini-3.1-flash-lite": { "enabled": true, "provider": "openrouter" },
    "glm-5.2": { "enabled": true, "provider": "openrouter" },
    "kimi-k2.7-code": { "enabled": true, "provider": "openrouter" }
  },
  "default": "claude-sonnet-4.6",
  "confidence": 0.8
}

Built-in models:

• claude-sonnet-4.6 — Anthropic or OpenRouter
• claude-haiku-4.5 — Anthropic or OpenRouter
• gemini-3.1-flash-lite — OpenRouter
• glm-5.2 — OpenRouter
• kimi-k2.7-code — OpenRouter (always reasons; thinking can't be disabled)

Set enabled: false to drop a model, and provider to choose how it's served. Users cannot add arbitrary upstream models in V1; model id and pricing are owned by the built-in catalog.

Routing is confidence-gated and cache-aware by default: candidates the classifier is confident about are compared by cache-adjusted cost, so the router only switches models when it is genuinely cheaper after accounting for the prompt cache it would give up (naive per-request switching can cost more than not routing at all). Where a provider has no reliable cache signal this simply reduces to price-ordered routing. See Configuration for details.

Docs

• Configuration
• Commands
• Dashboard
• Providers and Caching
• Pricing
• Thinking
• Telemetry
• Troubleshooting