POST /v1/messages specifically for this - it converts the request to Routor’s internal pipeline, runs the full routing/fallback/billing logic, and translates the response back to Anthropic’s shape (including streaming).
Why Use Routor with Claude Code
A Claude Code session isn’t one request - it’s dozens. Every tool call, file read, and multi-step edit is its own model call, and on a direct connection every one of them bills at whatever model you’re pinned to. Most of that traffic is trivial: reading a file, checking a lint result, confirming a small edit. Routor routes each of those independently instead of paying flagship price for all of them. Claude Code also sends its own internal bookkeeping prompts between your real turns - things like[SUGGESTION MODE: ...], [CONTINUE MODE: ...], and [SUMMARIZE MODE: ...]. Routor recognizes these specifically and routes them to the cheapest tier automatically, since they’re not something you asked for - they’re the tool talking to itself.
Setup
1. Get your API key
- Sign up at routor.io
- Go to API Keys in the dashboard and create one - it starts with
sk-routor-
2. Point Claude Code at Routor
Claude Code reads its endpoint and key from environment variables. Set these before launching it:/v1 on the base URL - Claude Code appends /v1/messages (and /v1/messages/count_tokens) itself, and Routor’s Anthropic-compatible router is mounted at exactly that path.
Routor’s auth accepts the key either as x-api-key (which is what Anthropic-format clients like Claude Code send by default) or as Authorization: Bearer - you don’t need to configure which one, both work against the same key.
3. Verify it’s routing
Start a Claude Code session and check your dashboard’s Logs page - you should see requests coming in with a routed model, tier, and savings %, the same as any other request. Claude Code’s own model field is ignored for routing purposes on this endpoint - Routor always decides, regardless of which Claude model Claude Code thinks it’s talking to.Recommended: Set a Tier Floor
Coding sessions are almost always at least moderately complex. Without a floor, an occasional trivial exchange might route to a very cheap model that’s too weak for a real coding task. Create a Routing Profile with aSTANDARD tier floor and use its dedicated API key instead of your main one, the same pattern recommended for Cline - it keeps every request at least at a capable coding tier while still letting genuinely hard requests escalate to COMPLEX.
What You’ll See
Every response still carries the full routing decision, even though Claude Code’s own UI doesn’t surface it - check/v1/routing/debug with a sample prompt, or the dashboard Logs page, to see the model, tier, category, confidence, and savings for any request.
Where to Go Next
- Use with Cline / Roo Code - same Anthropic-compatible endpoint, different client
- Set up a routing profile - lock in a tier floor for coding work
- How routing decisions are made