A single-process proxy that routes Claude Code requests to multiple AI providers. Use your Claude Max subscription for Anthropic models while routing subagents to cheaper or free alternatives.
Claude Code (OAuth) --> proxy (port 3456)
|-- Anthropic models --> api.anthropic.com (passthrough, OAuth preserved)
|-- Cerebras models --> api.cerebras.ai (OpenAI Chat Completions)
|-- Codex models --> chatgpt.com Codex API (OpenAI Responses API)
Claude Code is excellent but expensive. Opus 4.6 costs $15/$75 per million tokens. Most subagent work (implementation, code review, data analysis, background polling) doesn't need Opus-level intelligence.
This proxy lets you keep Opus for your main session while routing subagents to:
- Cerebras ($0.35-$2.75/M tokens) for fast implementation work
- Codex (free via ChatGPT Plus/Pro subscription) for code review, architecture, and data analysis
Token savings of 10-60x on delegated work, with zero changes to Claude Code itself.
- Node.js 20+
- Claude Code with Claude Max subscription (OAuth auth)
- Cerebras API key (free tier available at cerebras.ai)
- ChatGPT Plus/Pro subscription with Codex access (optional, for Codex routing)
git clone git@github.com:daviddawson/claude-code-multirouter.git
cd claude-code-multirouter
npm install
npm run buildCreate ~/.claude-code-router/config.json:
{
"port": 3456,
"providers": {
"anthropic": {
"passthrough": true
},
"cerebras": {
"baseUrl": "https://api.cerebras.ai/v1/chat/completions",
"apiKey": "YOUR_CEREBRAS_API_KEY",
"models": ["qwen-3-235b-a22b-instruct-2507", "zai-glm-4.7", "gpt-oss-120b"]
},
"codex": {
"baseUrl": "https://chatgpt.com/backend-api/codex/responses",
"authFile": "~/.codex/auth.json",
"models": ["gpt-5.1-codex-mini", "gpt-5.1-codex", "gpt-5.1-codex-max", "gpt-5.2-codex"]
}
},
"usage": {
"logFile": "~/.claude-code-router/usage.jsonl",
"costs": {
"qwen-3-235b-a22b-instruct-2507": { "input": 0.60, "output": 1.20 },
"zai-glm-4.7": { "input": 2.25, "output": 2.75 },
"gpt-oss-120b": { "input": 0.35, "output": 0.75 }
}
}
}For Codex, you need an auth token from the Codex CLI. Install and authenticate:
npm install -g @openai/codex
codex --loginThis creates ~/.codex/auth.json which the proxy reads for OAuth tokens.
# Start the proxy
node dist/server.js
# In another terminal, launch Claude Code through the proxy
ANTHROPIC_BASE_URL=http://localhost:3456 claudeOr add shell helpers to your .bashrc:
ccr-start() {
if ! lsof -ti :3456 >/dev/null 2>&1; then
nohup node ~/path/to/claude-multi-proxy/dist/server.js > /tmp/claude-proxy.log 2>&1 &
disown
sleep 1
fi
echo "claude-multi-proxy ready (port 3456)"
}
ccr-stop() {
kill $(lsof -ti :3456) 2>/dev/null
echo "claude-multi-proxy stopped"
}
claudem() {
ccr-start
ANTHROPIC_BASE_URL=http://localhost:3456 claude --dangerously-skip-permissions "$@"
}Then just run claudem.
The proxy inspects each request's system prompt for a routing tag:
<CCR-SUBAGENT-MODEL>provider,model</CCR-SUBAGENT-MODEL>
<CCR-SUBAGENT-MODEL>cerebras,zai-glm-4.7</CCR-SUBAGENT-MODEL>routes to Cerebras<CCR-SUBAGENT-MODEL>codex,gpt-5.1-codex</CCR-SUBAGENT-MODEL>routes to Codex- No tag = Anthropic passthrough (your main Opus/Sonnet session)
You place this tag in Claude Code custom agent definitions (~/.claude/agents/*.md). The tag is in the system prompt, invisible to the conversation.
---
name: fast-dev
description: "Implementation agent using Cerebras GLM 4.7"
model: sonnet
color: green
---
<CCR-SUBAGENT-MODEL>cerebras,zai-glm-4.7</CCR-SUBAGENT-MODEL>
You are an expert implementation engineer...The model: sonnet in the frontmatter is what Claude Code sends in the API request, but the proxy overrides routing based on the CCR tag.
| Agent | Provider | Model | Cost (per M tokens) | Use Case |
|---|---|---|---|---|
| Main session | Anthropic | Opus 4.6 | $15/$75 | Planning, coordination, complex reasoning |
| fast-dev | Cerebras | GLM 4.7 | $2.25/$2.75 | Code implementation, refactoring |
| fast-dev-gptoss | Cerebras | GPT-OSS 120B | $0.35/$0.75 | Simple tasks, background polling |
| watcher | Cerebras | GPT-OSS 120B | $0.35/$0.75 | CI/CD monitoring, deploy watching |
| code-reviewer | Codex | gpt-5.1-codex | Free (subscription) | Code review |
| software-architect | Codex | gpt-5.1-codex | Free (subscription) | Architecture analysis |
| code-reader | Codex | gpt-5.1-codex | Free (subscription) | Deep code understanding |
| data-analyst | Codex | gpt-5.1-codex | Free (subscription) | Data processing, test analysis |
Route implementation work to Cerebras models. GLM 4.7 is a purpose-built coding model at 5-30x cheaper than Anthropic. GPT-OSS 120B is the cheapest option for simple tasks.
You: "implement the login form"
Opus: spawns fast-dev agent → routed to Cerebras GLM 4.7
fast-dev writes code, Opus reviews the result
Route reviews to Codex models via your ChatGPT subscription. No per-token cost.
You: "review the changes"
Opus: spawns code-reviewer → routed to Codex gpt-5.1-codex
reviewer analyzes code, returns findings to Opus
Long-running watch tasks burn tokens on repeated polling. Route them to the cheapest model in a background agent.
You: "push to git, watch the build, monitor the k8s deploy"
Opus: spawns watcher in background → routed to Cerebras GPT-OSS
watcher polls for 30 minutes at $0.35/M tokens
main session stays free for other work
watcher returns structured summary when done
Without the watcher, this would be 20+ Opus turns at $15/$75 per M tokens just to repeatedly check job status.
JSON parsing, test result analysis, log pattern detection. Keep heavy data processing out of your main Opus context.
You: "analyze the comparison test results"
Opus: spawns data-analyst → routed to Codex (free)
analyst parses JSON, categorizes failures, produces summary
Opus receives concise findings, not raw data
The proxy converts between API formats transparently:
| Direction | Conversion |
|---|---|
| Claude Code --> Cerebras | Anthropic Messages API --> OpenAI Chat Completions |
| Cerebras --> Claude Code | OpenAI SSE --> Anthropic SSE |
| Claude Code --> Codex | Anthropic Messages API --> OpenAI Responses API |
| Codex --> Claude Code | Codex SSE --> Anthropic SSE |
| Claude Code --> Anthropic | Raw passthrough (zero transformation) |
Tool schemas are passed through with strict: false for Codex compatibility. Unsupported JSON Schema keywords (format, pattern, propertyNames, etc.) are accepted by Codex in non-strict mode.
Access the dashboard at http://localhost:3456/dashboard while the proxy is running.
Shows:
- Request counts by provider
- Token usage over time
- Cost tracking (Cerebras only, Anthropic/Codex via subscription)
- Recent request log
All requests are logged to ~/.claude-code-router/usage.jsonl:
{"timestamp":"2026-02-07T11:00:00.000Z","provider":"cerebras","model":"zai-glm-4.7","inputTokens":1200,"outputTokens":500,"cacheReadTokens":0,"cacheCreationTokens":0,"totalTokens":1700}Query via the API: curl http://localhost:3456/api/usage
src/
server.ts -- Fastify entry point
router.ts -- CCR tag detection, provider routing
config.ts -- Config loader (~/.claude-code-router/config.json)
codex-auth.ts -- OAuth token manager for Codex
usage.ts -- JSONL usage logging
providers/
anthropic.ts -- Passthrough proxy (OAuth preserved)
cerebras.ts -- Anthropic -> OpenAI Chat Completions
codex.ts -- Anthropic -> OpenAI Responses API
format/
anthropic-to-openai.ts -- Request converter (Cerebras)
openai-to-anthropic.ts -- SSE stream converter (Cerebras)
anthropic-to-codex.ts -- Request converter (Codex)
codex-to-anthropic.ts -- SSE stream converter (Codex)
dashboard.html -- Usage dashboard (Chart.js)
- Cerebras models reject
reasoningandthinkingfields. The format converter strips these. - Codex strict mode is incompatible with Claude Code's tool schemas. The proxy uses
strict: false. - Codex OAuth tokens expire and are auto-refreshed, but initial authentication requires the Codex CLI.
- Image content in messages is converted to data URLs for Cerebras but not supported by Codex.
- The proxy binds to localhost only. Not designed for network exposure.
MIT