claude-code-router Anthropic-to-OpenAI translation over a local gateway
Parent: Mac local LLMs: Agent clients, context and compaction · Published reference · snapshot 2026-10-05
↓ Facts as markdownall context files
CCR listens on one local endpoint (default http://127.0.0.1:3456); agents point at it, and CCR selects provider, model and account per request from profiles, conditions and rewrites.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Facts
- CCR listens on one local endpoint (default http://127.0.0.1:3456); agents point at it, and CCR selects provider, model and account per request from profiles, conditions and rewrites. [source]
- A provider entry carries a protocol choice (OpenAI Chat, OpenAI Responses, Anthropic Messages, Gemini); the README advertises "protocol probing" and "connectivity checks" when a custom endpoint is added. [source]
- Routing supports conditions on headers and bodies, prefixes, rewrites, retries and ordered fallback models. [source]
- The project is no longer only a Claude Code shim: version 3.1.1 (2026-09-16) presents a desktop app plus npm CLI plus Docker image, and lists ten agent clients. [source]
- The management UI runs on port 3458 (CLI `ccr ui`, Docker), separate from the gateway on 3456; the npm CLI needs Node.js 22 or newer. [source]
- Whether CCR repairs malformed tool calls from a weak local backend is not stated in the README; its listed features are routing, retry, fallback and observability, not repair. [source]
- The README documents tool-call logging ("tool calls; agent traces" in request details), which is the only built-in aid for diagnosing translation faults. [source]
- Older guides (Dec 2025) call a translation proxy mandatory for Claude Code on local models; local-llm-server-as-a-coding-agent-backend-on-mac.md already resolves this by date. CCR's own README no longer frames local models as its purpose, so the "proxy needed" argument now rests on third-party guides only. [source]
- Exact transformer behavior for `tool_use` / `tool_result` blocks against llama-server or Ollama backends (source code not read; docs site ccrdesk.top not fetched). [source]
- Whether CCR hoists non-leading system messages (see anthropic-compatible-local-endpoints-and-role-system-messages.md). [source]
- CCR's README describes it as "a local model gateway and control plane for coding agents" giving agents one stable local endpoint. [source]
- The default CCR gateway address is http://127.0.0.1:3456. [source]
- The CCR management UI is served on http://127.0.0.1:3458 by the npm CLI (`ccr ui`) and by Docker. [source]
- The npm package is `@musistudio/claude-code-router` and requires Node.js 22 or newer. [source]
- The README lists supported protocols as OpenAI Chat / Responses, Anthropic Messages, Gemini Generate Content / Interactions, plus custom compatible providers. [source]
- CCR's provider setup includes protocol probing, model discovery and connectivity checks for custom endpoints. [source]
- CCR routing supports header and body conditions, prefixes, rewrites, retries and ordered fallback models. [source]
- CCR lists ten agent clients: Claude Code, Claude Design, Codex, Grok CLI, Kimi CLI, Kilo Code, OpenCode, Pi, ZCode and WorkBuddy. [source]
- As of 2026-10-04 the repository shows 37.5k stars, 953 commits and release v3.1.1 dated 2026-09-16. [source]
- CCR request logs record resolved provider, model, credential, status, latency, tokens, estimated cost, tool calls and agent traces. [source]
- The README does not describe any tool-call repair step; documented resilience features are retries, credential pools and fallbacks. [source]
Children
- No children recorded.