https://router.workweave.ai. Weave runs and operates it — you do not deploy anything, run Docker, or hold provider infrastructure. (The open-source router is a separate, self-hosted distribution with its own setup instructions.)
Weave Router is not the Claude Code plugin. The plugin is telemetry: it observes your Claude Code sessions and feeds Weave reports. The router changes which model serves each request. They are independent — you can run either or both.
Request flow
A subscription turn runs on plan capacity, not credits. The router either passes through the token your client sends or uses an enrolled subscription (stored as an encrypted refresh token), trying yours first and a teammate’s when sharing is on. Turns Weave serves from its own accounts draw down prepaid credits at upstream cost, and turns on your own provider keys draw down only Weave’s platform fee.How it works
1
Your agent points at Weave
The installer rewrites your client config so requests go to
https://router.workweave.ai instead of directly to Anthropic, OpenAI, or Google. Your Weave Router key rides in an X-Weave-Router-Key header, separate from your provider credentials.2
Weave scores the request
Every LLM call is routed independently — in an agentic loop that means every tool call, not just the first prompt. The scorer weighs what the request needs against price, your quality-vs-price setting, your enabled models, session state, and cache warmth.
3
Weave forwards it upstream
The request goes to the selected model. If you are on a subscription (Claude or ChatGPT/Codex) your credentials pass through untouched so that turn bills to your plan. Otherwise the turn bills against your Weave Router credits or your own provider key.
4
You see the decision and the cost
The routed model is shown in your client (status line, in-band marker, or the pi route display) and every decision lands in the Router report with cost and savings.
What it changes
- Cost. Cheap requests (file reads, small edits, tool plumbing) stop running on your most expensive model. Hard requests still get the strong model — the router allows tier upgrades when a turn needs one.
- Plan headroom. With subscription routing on, plan-covered turns keep using your plan, and overflow goes to paid models instead of blocking your engineers at their limit.
- Visibility. Per-request model, cost, and savings, broken down by engineer and model.
What stays the same
- Your agent, your prompts, your workflow. No forked CLI, no wrapper binary.
- Your provider subscriptions. Claude Code and Codex plan credentials pass through to the provider.
- Streaming, tool calls, caching, and the native request format of each client.
Supported clients
Anything that speaks the Anthropic Messages, OpenAI Chat Completions, or Gemini
generateContent API can point at the router directly.
Where things live
Next steps
- Get started — create a key and install the router
- Authentication — keys, subscriptions, and SSO
- Billing and credits — credits, spend limits, and BYOK
- Models and routing — control what the router picks
- FAQ — the questions customers ask most