Claude Code Router (CCR, repo name claude-code-router) lets you keep Claude Code’s terminal interface while sending the actual model requests somewhere else, such as DeepSeek or OpenRouter. In February we reviewed version 1.0.8, which you set up by editing a JSON file. This update covers the current musistudio/claude-code-router release, what the Claude Code router setup looks like now, what it can save you at today’s prices, and the one fact the old version never mentioned: Anthropic does not support this way of running Claude Code.
⚡ TL;DR: The Bottom Line
What It Is: A free, MIT-licensed local gateway that sits between a coding agent (Claude Code, Codex, Grok CLI, OpenCode and others) and the model providers you choose. Version 3.1.3 ships as a desktop app, an npm command-line tool or a Docker container.
Best For: Developers who want to point Claude Code at a cheaper model on purpose and are happy to debug the result.
Price: $0 for CCR itself. You pay each provider’s token prices, for example DeepSeek V4.1 Flash at $0.15 per million input tokens off-peak, against $2 for Claude Sonnet 5.5.
Our Take: The savings are real on paper, but Claude Haiku 5.5 at $0.10 per million input tokens (prompts up to 100,000 tokens) is priced in the same range, and Anthropic does not support non-Claude models behind Claude Code.
⚠️ The Catch: Anthropic does not support this setup, and CCR is an independent open-source project with 931 open issues.
📑 Quick Navigation
🔀 What Claude Code Router Is in 2026
Think of CCR as a switchboard operator for AI coding agents. Your agent still talks to one local address, and the operator decides which provider’s model answers. The project describes itself as a “local model gateway and control plane for coding agents” with “one stable local endpoint”. As a Claude Code model router, it forwards each request to the provider you picked, applies your Claude routing rules, retries on errors, and logs what happened.
Three things are different from the version we covered in February:
- It is an app now. The README recommends the desktop app for Windows, macOS and Linux. The npm command-line tool (
npm install -g @musistudio/claude-code-router, thenccr ui) and Docker are the other two options, and the npm tool needs Node.js 22 or newer. - It is not only for Claude Code. The README lists Claude Code, Codex, Grok CLI, Kimi CLI, Kilo Code and OpenCode among the supported agents.
- Configuration moved into a UI. Our February version (1.0.8) was set up by editing
~/.claude-code-router/config.jsonand starting it withccr startandccr code. The current README describes provider, agent and routing setup through the management screens and does not mention that file. If you rely on the old commands, check the project’s CLI reference before upgrading.

Claude Router on GitHub: stars, issues and release pace
The Claude router GitHub page for musistudio/claude-code-router shows 37.6k stars and 3.2k forks on October 10, 2026, under the MIT license. The claude-code-router GitHub stars count was 26,400 in our February version, so interest has grown by about 11,000 stars. It also shows 931 open issues and 203 open pull requests, which is what an active but busy project looks like. Releases are frequent: v3.1.0 on September 10, v3.1.1 on September 16, v3.1.2 on October 8 and v3.1.3 on October 9, which brought a rebuilt model catalog served from a CDN. Fast releases mean fast fixes and also fast change, so pin a version if you depend on it.
🚀 Claude Code Router Setup (Version 3)
For the Claude Code router setup, CCR’s quick-start guide boils down to the four steps below. The interface labels in the docs are in Chinese, so the names here are translations:
- 1. Install and start. Launch the desktop app, or use the npm tool:
ccr uiopens the management screen andccr serveruns the gateway. With the npm tool the management UI is at http://127.0.0.1:3458 and the gateway at http://127.0.0.1:3456 by default, and the desktop app’s guide lists 127.0.0.1:3456.ccr uiis the main CCR command to remember. - 2. Add a provider. Pick a built-in preset or enter a custom API address and your key. CCR detects which protocols and models the endpoint supports, and a connectivity check sends a real request to confirm the key and model work.
- 3. Configure Claude Code. In Agent Config, select Claude Code, set the model and the small fast model, apply, and launch Claude Code from CCR.
- 4. Check the logs. Turn on Request Logs and Agent Observability under Settings, send a request, and look at which model actually answered.

CCR’s own troubleshooting page lists the usual failures: 401 or 403 means a key or base-URL problem, “model not found” means the model name does not match across the provider list, the routing choice and the Agent config, and “unexpected routing” means your rules are in the wrong order. That is a useful reminder of how many places a model name has to line up.
🎯 Smart Routing in v3: Rules, Fallbacks and Subagent Models
Claude routing in CCR v3 is rule-based. According to the routing docs, a rule can match request headers or body fields, rewrite fields (the common case is setting the model), retry the same model on 408, 409, 429 and 5xx errors, or fall back to backup models in order on any 4xx or 5xx. Claude Code gets a built-in route that uses the default model from your Agent config when the request does not name a model CCR recognizes. Subagents can be sent to a different model too: their prompts can carry a model tag that CCR reads and routes on, which needs a description set on each model.
One thing readers of our old version will look for is missing. The classic default, background, think and longContext scenario names do not appear on the routing docs page we read, so an equivalent has to be built with custom rules. CCR also lists Fusion (which adds things like vision or web search to a model) and ToolHub, which we have not covered in depth.
🔍 REALITY CHECK
Marketing Claims: Our February version repeated the older headline that routing could “reduce API costs by 10x”.
What It Really Means: The current README makes no savings claim at all. What you save depends on which tasks you route, to which model, at which price, which is what the next section works out from list prices.
Verdict: A 10x gap exists on token price alone (Opus or Sonnet against DeepSeek Flash), but real savings depend on retries and on whether the cheaper model finishes the task.
💰 What It Costs: Claude Code Router vs Paying Anthropic
CCR costs nothing, so the bill is whatever your providers charge. Start with what Anthropic says Claude Code costs. Its cost page puts the average across enterprise deployments at around $13 per developer per active day and $150 to $250 per developer per month, with costs below $30 per active day for 90% of users. (Our February version quoted $6 a day, which the current cost page does not state.) On a subscription, Claude Code is included: Pro is $20 a month billed monthly ($17 a month on the annual plan), Max is $100 a month with 5x Pro’s usage every 5 hours and every week, and Ultra is $200 a month with 10x Pro’s usage every week and no 5-hour limit for verified accounts.
Now the token prices. These are list prices per million tokens, from Anthropic’s pricing page and DeepSeek’s pricing page:
| Model | Input (per 1M) | Output (per 1M) | Example: 1M in + 200k out |
|---|---|---|---|
| Claude Opus 5.5 | $4 | $20 | $8.00 |
| Claude Sonnet 5.5 | $2 | $10 | $4.00 |
| Claude Haiku 5.5 (prompts up to 100,000 tokens) | $0.10 | $0.50 | $0.20 |
| DeepSeek V4 Pro (off-peak / peak) | $0.66 / $1.32 | $1.98 / $3.96 | $1.06 / $2.11 |
| DeepSeek V4.1 Flash (off-peak / peak) | $0.15 / $0.30 | $0.60 / $1.20 | $0.27 / $0.54 |
The last column is our own arithmetic: a workload of one million input tokens and 200,000 output tokens at list price, cache-miss rates, no discounts. DeepSeek’s peak window is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. Above 100,000 prompt tokens, Haiku 5.5 is $0.50 input and $2.50 output, which is dearer than DeepSeek V4.1 Flash even at peak. That matters because Anthropic says Claude Code sends your whole conversation with each request, so a long session can cross the line.
📊 List Price of a 1M-Input, 200k-Output Workload (USD)
That Haiku line is the biggest shift in the price picture. Anthropic’s own small model is priced in the same range as DeepSeek’s cheapest for prompts up to 100,000 tokens, and it needs no router, because Anthropic’s cost page already points Claude Code users at Haiku for simple subagent work. Anthropic’s cost page also recommends Sonnet for most coding tasks and Haiku for simple subagent work. Whether Haiku 5.5 is good enough as your main coding model is a quality question the price table cannot answer.
💡 Key Takeaway: Before you set up a router, price the same workload on Claude Sonnet 5.5, Claude Haiku 5.5 and your target model. The router only wins if the cheaper model actually finishes your tasks.
⚖️ Is It Allowed? What Anthropic Says About Gateways
Claude Code can be pointed at a gateway, and Anthropic documents how. Its LLM gateway page is also blunt: “Anthropic doesn’t endorse, maintain, or audit third-party gateway products, and doesn’t support routing Claude Code to non-Claude models through any gateway.” The same page says a gateway has to keep forwarding new Claude Code features, or the matching features break.
The billing detail matters too. When you set a gateway credential, Anthropic says requests use that credential instead of your claude.ai login, your subscription’s usage limits do not apply to them, and the traffic is billed per token to the owner of the credential the gateway forwards. Setting only the base-URL variable does not replace your subscription. So, by Anthropic’s description, a gateway is not a way to stretch a Pro or Max plan, it is a way to swap models and pay per token elsewhere.
🔍 REALITY CHECK
Marketing Claims: “No Anthropic account required” was a headline benefit in our February version.
What It Really Means: It is true that CCR does not need one. Anthropic’s position is that it supports gateways, but not non-Claude models behind them. Feature gaps, odd errors and breakage after a Claude Code update are yours to solve.
Verdict: Fine for a personal side project, a harder sell for work that your employer or clients depend on.
⚠️ Limitations and Gotchas
- Tool calling is on you. Claude Code relies on tool calls for edits and commands, so check that the model you pick supports them. Anthropic adds that a gateway has to keep forwarding Claude Code’s newer features, or the matching features break.
- Your code passes through a local proxy that can log it. With Request Logs on, CCR records the request body and the response body, which for a coding agent means source code. Keep the logs on a machine you control, and read the Docker guide before exposing the gateway beyond localhost.
- Everything has to line up. A model name must match in three places (provider list, routing choice, Agent config) or you get a “model not found” error.
- Heavy change. Four releases since September 10 and 931 open issues mean features move. Our February instructions may no longer apply.
- Provider coverage differs. The README lists OpenAI Chat and Responses, Anthropic Messages, Gemini, OpenRouter, DeepSeek, SiliconFlow, Moonshot, Kimi Code, Mistral, Z.AI, Bailian and custom compatible providers. Local-model hosts such as Ollama are not named in the README we read, so check the provider presets for your setup.

📬 Enjoying this breakdown?
Get honest AI tool analysis delivered weekly. No hype, no spam.
👤 Who Should Use It (and Who Shouldn’t)

Use CCR if: you already pay per token for Claude Code (API or cloud billing) and want to try cheaper models for low-stakes tasks, you like having a request log that shows which model answered, or you want one local gateway for several agents (Claude Code, Codex, OpenCode) rather than one setup each.
Skip CCR if: you are on a Pro or Max plan and happy with it (a gateway credential bypasses plan limits and moves you to per-token billing), you need a setup your team can get official support for, or you do not want to maintain a local proxy.
🔄 Claude Code Router Alternatives
If you want to cut the cost of AI coding, a Claude Code router alternative may be simpler:
- Use cheaper Claude models inside Claude Code. Anthropic’s cost page recommends Sonnet for most coding tasks, Haiku for simple subagent work, and clearing context between unrelated tasks. No gateway, no unsupported setup. See our Claude AI review for plan limits.
- Point Claude Code at a gateway you already run. Anthropic documents the base-URL method, and its cost page says several large enterprises reported using LiteLLM, an open-source tool that tracks spend by key, while noting it is unaffiliated with Anthropic and not audited for security.
- Switch tools instead of models. Goose is an open-source agent that lets you bring your own model, Google Antigravity is an agent IDE, and Claude Code vs Cursor compares two popular options. Our AI agents for developers guide maps the rest.
- Consider a subscription. If you hit usage limits rather than a bill, a Max or Ultra plan ($100 or $200 a month) keeps Claude Code included at a flat price.
❓ FAQs: Your Questions Answered
Q: Is Claude Code Router free?
A: Yes. CCR is MIT-licensed and free to install. You pay only the providers you route to, at their own token prices.
Q: What is musistudio/claude-code-router on GitHub?
A: It is the open-source repository for Claude Code Router. On October 10, 2026 it showed 37.6k stars, 3.2k forks, 931 open issues and v3.1.3 as the latest release.
Q: How do I start CCR, and what does the ccr command do?
A: After npm install -g @musistudio/claude-code-router (Node.js 22 or newer), run ccr ui to open the management screen at 127.0.0.1:3458. Or use the desktop app. The ccr start and ccr code commands are how version 1 worked; the README now shows only ccr ui.
Q: Does Anthropic support running Claude Code on other models?
A: No. Anthropic documents gateways but says it doesn’t support routing Claude Code to non-Claude models through any gateway, and it doesn’t endorse or audit third-party gateway products.
Q: Do I need an Anthropic account to use it?
A: CCR itself does not ask for one. You need a key for whichever provider you route to. If you do use an Anthropic subscription alongside a gateway credential, requests that carry the gateway credential skip your plan limits and bill per token.
Q: Does CCR work with agents other than Claude Code?
A: Yes. The README lists Claude Code, Codex, Grok CLI, Kimi CLI, Kilo Code, OpenCode and other compatible API clients.
Q: Is it cheaper than paying for Claude Code?
A: Per token, DeepSeek V4.1 Flash costs far less than Claude Sonnet 5.5 ($0.15 against $2 per million input tokens off-peak), but Claude Haiku 5.5 is $0.10 for prompts up to 100,000 tokens. Whether you save money depends on how often the cheaper model fails and has to be retried.
Q: Where does my code go with a local router?
A: It is forwarded to the providers and tools you configure, which can include add-ons such as vision or web search, so their privacy policies apply. Request logs can store request and response bodies, and the README does not describe telemetry.
Final Verdict: Is Claude Code Router Worth It?
Claude Code Router has grown from a clever proxy into a real local gateway, with a desktop app, rule-based routing and 37.6k stars. It is a good tool for someone who wants to experiment with cheaper models on purpose. It is a weaker answer to “how do I cut my Claude Code bill” than its headline suggests, because Claude Haiku 5.5 costs about what DeepSeek’s cheapest model does for prompts up to 100,000 tokens, and because Anthropic does not support non-Claude models behind Claude Code. For most people, the cheapest fix is a smaller Claude model or a subscription.
Use CCR if: you pay per token, you are comfortable on your own for support, and you will check the request logs to see what each task costs.
Stay with standard Claude Code if: you want Anthropic’s support, you work on a Pro or Max plan, or you need your setup to survive the next Claude Code update. Check the project yourself at its GitHub page before installing.
Related Reading
- Claude AI Review 2026
- Claude Code vs Cursor 2026
- Claude Code Plugins Review 2026
- Claude Agent Teams Review
- Goose AI Review
- Google Antigravity Review
- Top AI Agents for Developers 2026
- DeepSeek Review
- ChatGPT Codex Review
Stay Updated on AI Developer Tools
The AI coding tool market changes weekly. New models launch, prices move and plans change. Don’t miss the next change that affects your bill.
- ✅ Honest Breakdowns: We read the pricing pages and docs, not press releases
- ✅ Price Tracking: Know when tools get cheaper or go free
- ✅ Comparison Updates: As the market shifts, we update our verdicts
- ✅ Feature Launches: Major updates covered within days
- ✅ No Hype: Just the AI news that actually matters for your work
Free, unsubscribe anytime.
Last Updated: October 31, 2026
Claude Code Router Version Covered: 3.1.3 (released October 9, 2026)
Next Review Update: November 30, 2026
Have a tool you want us to cover? Suggest it here | Questions? Contact us