Claude Code Router: Set Up CCR v3 and Route Claude Code to Any Model

Claude Code Router is a community project by the GitHub user musistudio, MIT-licensed, created in February 2025 and published on npm as @musistudio/claude-code-router (binary ccr). When we read the repository in 2026 it had 38,096 stars, 3,203 forks and a latest release of v3.1.3, and npm counted 193,927 downloads in the 30 days before collection. It is not an Anthropic product.

The README now calls it "a local model gateway and control plane for coding agents". Claude Code, Claude Design, Codex, Grok CLI, Kimi CLI, Kilo Code, OpenCode, Pi, ZCode and WorkBuddy all get one stable local endpoint, and behind it you manage providers, models, accounts, routing rules and tools in one place. Supported upstream protocols are OpenAI Chat and Responses, Anthropic Messages, and Gemini Generate Content and Interactions, and named support for OpenRouter, DeepSeek, SiliconFlow, Moonshot, Kimi Code, Mistral, Z.AI, Bailian and custom compatible endpoints.

Why people use it. Claude Code is built for Anthropic's models and its own billing. CCR lets you keep the Claude Code interface and tools while paying a different provider per token, mixing a strong model for planning with a lower-cost one for background work, or running a local model. The README lists the jobs it does: switch providers without editing agent config files, keep requests running with retries, credential pools, key rotation and ordered fallbacks, add capabilities (vision, web search, MCP tools) to models that lack them through its Fusion feature, and show request logs with the resolved route, latency, tokens and an estimated cost.

The project has had three major versions on npm: 1.0.0 in June 2025, 2.0.0 in January 2026 and 3.0.0 in July 2026. The table shows what moved.

There are three distributions, and they read the same local configuration directory (~/.claude-code-router on macOS and Linux, %APPDATA%\claude-code-router on Windows).

Desktop app. Download the installer for your OS from the GitHub releases page (.dmg for Apple Silicon or Intel, .exe for Windows, .AppImage for Linux). It adds a tray icon, notifications and automatic updates. The docs call its launcher command ccr-app; do not mix it up with the npm ccr command.

npm CLI. For servers, SSH sessions or anyone who does not want Electron. It needs Node.js 22 or newer and starts the same gateway plus a browser-based management UI. If ccr is not found after install, check that the directory from npm prefix -g is on your PATH.

Docker. From a source checkout, the compose file publishes one Nginx endpoint on http://127.0.0.1:3458 that serves both the UI and the gateway routes, and stores data under /data/.claude-code-router. The docs ask you to read the Docker guide before exposing CCR beyond localhost.

First-run order (all three). Add a provider and at least one model, create a CCR client key on the API Keys page, set the default model and any routing rules, start the gateway on the Server page, then connect an agent. The management token that protects the UI and the CCR client keys that authenticate model requests are separate credentials. The ccr ui URL contains a ccr_web_token parameter, so treat the whole URL as a password.

Verify. The docs' check is a /health request on the gateway address that should return 200, one small model request with the client key, and a look at Logs for the requested model, the resolved provider and model, status and latency. A UI that loads does not prove the gateway works.

In v3 you no longer export environment variables by hand. On the Agent Config page, choose Add profile, pick Claude Code and give it a name such as Claude Code - Work. The fields that matter:

Tier mapping is the useful part. Claude Code picks a model per tier, so you can put a strong provider model on the Opus tier and a lower-cost one on Haiku, the small and fast tier. Leave a tier empty to keep Claude Code's own choice. Inside a CCR-launched session, /model lists the models CCR exposes, including Fusion models.

Launch. Run ccr "Claude Code - Work" (or the copied ccr-app command from the desktop card). Add Claude Code's own flags after --. If requests do not show up in CCR's logs, the docs list four causes in order: the service is not running, Claude Code was opened directly instead of from CCR, the profile is not applied or not enabled, or its scope does not cover the current project.

Subscriptions and billing. Anthropic's gateway docs say that while a gateway credential is active, requests carry it instead of your claude.ai login, the subscription's usage limits do not apply, and the traffic is billed per token to whoever owns the upstream credential. With CCR that means your DeepSeek, OpenRouter or other provider account pays, not your Claude plan.

Built-in Claude Code route. CCR detects requests from Claude Code and sends main requests to the profile's model when the client has not selected a model CCR recognizes. An explicit model you choose with /model wins. CCR also strips the billing helper system message Claude Code injects so it does not skew routing.

Subagent routing. When Claude Code spawns an Agent, Task or Workflow, CCR can let it choose a model. Fill the Description field for models on the Models page (for example "good for code search, file triage, summaries, small edits"). CCR then injects the list into the Agent / Task tool description, Claude Code writes a <CCR-SUBAGENT-MODEL>provider/model</CCR-SUBAGENT-MODEL> tag at the start of the spawned prompt, and CCR removes the tag and routes that request to the tagged model. With no descriptions, nothing is injected.

Custom rules. The Routing page holds an ordered list. Each rule has a condition on request.header or request.body (field, operator, value), one or more rewrites of the request, and its own on-failure behavior. The first enabled rule that matches wins, so order is priority.

Script rules. When one condition is not enough, choose Node.js script. The file is an async function body that receives a read-only input (full body and headers, an estimated tokenCount, a summary with lastUserText, toolNames, messageCount and hasImage) and returns null for no match or an object with a model, rewrites and fallback. Timeouts run from 10 to 30,000 ms, files may be up to 5 MiB, and CCR re-reads the file on each execution. Errors and timeouts fail open: CCR logs a diagnostic and moves to the next rule. The example below sends small turns that involve an X search tool to a lower-cost model.

Fallbacks. A global fallback and rule-level fallbacks retry or fail over to other models when a request fails, which is what keeps a long agent run alive when one provider request fails.

If you have an older install, your file looks like the one below. The v2 README defined Providers (each with name, api_base_url, api_key, models and optional transformer) and a Router block whose values are provider,model pairs: default for everything, background for small background calls, think for reasoning or plan mode, longContext above longContextThreshold tokens (60,000 in the README example), and webSearch. API keys could be written as $VAR or ${VAR} and were read from the environment.

What happens on upgrade. v3's configuration-file doc is explicit: a legacy config.json is read once as a migration source when no SQLite config exists, and after migration editing config.json has no effect. Back it up, upgrade, open the UI and check that each provider passes Check Connection, then re-create your routing intent with tier models and rules: background maps naturally to the Haiku tier, think to the Opus tier, and longContext to a rule on the estimated token count.

Commands that moved. ccr code launched Claude Code through the router; the v3 CLI reference lists start, ui, stop, serve, web and ccr <profile>, not code. ccr restart after editing JSON is gone because the UI applies changes. Do not edit config.sqlite while CCR runs; use Settings, Export data for backups.

How many people still describe the old flow? Among the 118 posts we collected from after the 3.0.0 release, 2.5% still mention ccr code or config.json, and 11.0% mention the desktop app, ccr ui, profiles or v3. Most posts do not name a version at all, so check the date of any tutorial you follow.

A router is no longer the only way to put another model behind Claude Code. Several providers now expose an Anthropic-compatible endpoint, and Claude Code only needs ANTHROPIC_BASE_URL plus a credential. DeepSeek's docs give ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic with your DeepSeek key, and map model names that start with claude-opus to deepseek-v4-pro and claude-sonnet / claude-haiku to deepseek-flash. OpenRouter's Claude Code guide sets ANTHROPIC_BASE_URL to https://openrouter.ai/api, the OpenRouter key in ANTHROPIC_AUTH_TOKEN and an explicitly empty ANTHROPIC_API_KEY, and says no local proxy server is required.

What breaks either way. Anthropic's LLM gateway page says any gateway that exposes a supported API format works, but Anthropic does not endorse or audit third-party gateways and does not support routing Claude Code to non-Claude models through any gateway. OpenRouter says its Claude Code path is only guaranteed with the Anthropic first-party provider. DeepSeek's Anthropic endpoint lists unsupported or ignored fields, including mcp_tool_use blocks, cache_control and budget_tokens. Expect some Claude Code features, tool-calling quality and prompt caching to vary by model, and keep a Claude model on the tier that does your hardest planning if results matter.

Which to pick. One provider with an Anthropic endpoint: skip the router. Several providers, per-tier mixing, OpenAI-only or local models, or you want logs and fallbacks: use CCR. Model and price detail for the two providers people pair most often is on our OpenRouter API and DeepSeek V4 Flash pages.

We searched X for posts that mention Claude Code Router by name, by its repository URL or by the ccr code command, and kept 405 original posts by 322 distinct accounts (151 from 2025 and 254 from 2026). The first question was which models people route Claude Code to.

44.7% of posts name at least one non-Anthropic provider or model family. OpenAI, GPT or Codex lead at 19.0% (Codex is also one of the agents CCR v3 serves), then OpenRouter 12.8%, Gemini 11.9%, Ollama or other local setups 9.6% and Kimi 9.1%. Claude model names (Sonnet, Opus, Haiku) appear in 24.0%, usually as the thing being replaced or kept for one tier.

What they talk about. 32.6% of posts are setup or configuration talk (install, config, tutorial, guide), 29.1% mention cost, price or free use, and 10.4% report an error or problem. The conversation is international: English 58.8%, Chinese 17.8%, Japanese 9.9%. 31.9% link the GitHub repository and 54.8% carry an image or video, which fits a topic where people share screenshots of config and logs.

Reach is concentrated. The median post has 8 likes and 1,354 views, while the 10 most-liked posts hold 62.7% of all likes in the sample (the top one has 9,615). Posts in X's Top tab had a median of 10 likes against 6 for posts that appeared only in Latest. The project account linked from the README, @musistudio2026, had 65 followers at collection, and 23.7% of its last 93 non-repost posts mention CCR, so most discovery happens through other people's threads and the GitHub page rather than the maintainer's feed.

The second chart puts the topic buckets next to the language mix. Buckets are keyword lists fixed before collection and overlap, so a post about a free model setup counts in both setup and cost.

Reading it. Setup and cost dominate: people arrive at CCR to run Claude Code on a model they already pay for, or on a free or lower-cost tier, and the first obstacle is configuration. Only 10.4% read as problem reports, which is lower than setup talk by a factor of three. Since 3.0.0, the share of posts that mention the v3 flow (11.0%) is above the share still describing ccr code or config.json (2.5%), but both are small next to the 118 posts in that period: most people do not say which version they run.

A practical use of this data. Before you move a tier to a new model, search X for that model plus Claude Code and read the problem posts. The Python example at the end of this page does that for any list of model names.

CCR routes model traffic; it does not give the model new data. For an agent that has to answer "what are developers saying about this model" or watch a launch on X, add X data as a tool. Tools are called by Claude Code on the agent side, so they work the same whichever provider CCR routes the model to, as long as that model handles tool calls well.

MCP server. TwitterAPI.io runs a hosted MCP server at https://mcp.twitterapi.io/mcp with 12 tools for public X data (search_tweets, get_user_info, get_user_last_tweets, get_tweet_replies, get_trends and others). It takes your API key as a Bearer token. No X account and no X developer approval are needed for reads. New accounts that pass the automatic bot check get up to $1 of free credits (at most 100,000 credits, no card, one per person or network, expiring after 31 days).

REST. Scripts and hooks can call the same data directly: GET https://api.twitterapi.io/twitter/tweet/advanced_search with an X-API-Key header, reading tweets[], has_next_page and next_cursor. That is how the study on this page was collected.

The X data bill. TwitterAPI.io charges per result: $0.15 per 1,000 tweets returned (15 credits per tweet), so a full 20-tweet search page costs $0.003 and an empty search costs the 15-credit minimum, $0.00015. The official X API pay-per-use rate card at docs.x.com/x-api/getting-started/pricing lists $0.005 per post read, $5.00 per 1,000, about 33x more, with pay-per-use capped at 3 million post reads a month. Profiles cost $0.18 per 1,000 on TwitterAPI.io against $0.010 per user read on X.

Rate limits start at 0.2 requests per second before the first payment and rise to 20 once paid, which is plenty for an interactive coding session. Plans and credit packs are on the pricing page.

Question. How do people on X talk about Claude Code Router: which models they pair it with, what they ask about, which languages they post in, how much reach a typical post gets and whether they describe the v3 or the legacy flow.

Queries. /twitter/tweet/advanced_search with "claude code router" OR "claude-code-router" OR "ccr code" in Latest (all pages the search returned) and Top (10 pages), plus url:claude-code-router OR musistudio OR ccrdesk in Latest; and /twitter/user/info plus /twitter/user/last_tweets (5 pages, replies included) for the project account musistudio2026. 46 requests, 588 tweets returned, about $0.09 at our rates.

Sample. 488 unique search posts; 405 matched a Claude Code Router pattern in text or links and were posted after the repository was created (February 25, 2025); 5 older ccr code matches were unrelated and dropped. No reposts were returned. Collected in 2026.

Method. Provider, topic and version buckets are case-insensitive keyword lists fixed in the study protocol before collection; one post can fall in several. Engagement figures are at collection time. The v3 split uses the 3.0.0 npm publish date.

What the sample cannot show. X search does not return every post, so this is a sample of the visible conversation, not a census. Keyword buckets miss posts that describe a model without naming it, and "Codex" counts both OpenAI's model family and the Codex agent. Mentions are not usage: a post naming DeepSeek is not proof that DeepSeek works well in CCR. Nothing here measures CCR's install base.

Privacy. Aggregates only. Search results come mostly from private individuals, so raw pages were not kept in the repository; the published sample has no IDs, URLs, text or handles. The only account named is the project's own account linked from its README.