A load-balancing proxy for the Claude API — distributes requests across multiple accounts and providers to avoid rate limits, with request compression, prompt-cache keepalive, and full request-level analytics.
# via bun (recommended)
bun install -g @adamlangepl/tc-proxy
tinyclaude
# via npm
npm install -g @adamlangepl/tc-proxy
tinyclaude
# no install — npx/bunx
npx @adamlangepl/tc-proxy@latestPre-compiled binaries for all platforms are on the Releases page; building from source is covered in Getting Started.
Add an account and point Claude Code at the proxy:
tinyclaude --add-account myaccount --mode claude-oauth --priority 0
export ANTHROPIC_BASE_URL=http://localhost:8080
claudeDashboard: http://localhost:8080/dashboard
See Getting Started for Docker, systemd, SSL/HTTPS, and Codex CLI setup.
- Load balancing — session-based, least-used, and session-affinity strategies with automatic failover across account pools
- Multi-provider support — Claude OAuth, Claude Console API, AWS Bedrock, Vertex AI, OpenAI-compatible, z.ai, Minimax, OpenRouter, Kilo, Codex, xAI/Grok, Ollama, and more
- Combos — named cross-provider fallback chains, one per model family (Opus/Sonnet/Haiku)
- Request compression — automatically compresses large tool-result payloads to cut token usage
- Prompt-cache keepalive — replays cached requests on a schedule to keep Anthropic's server-side prompt cache warm
- Auto-fallback / auto-refresh — restores preferred accounts and starts new usage windows automatically as rate limits reset
- Real-time analytics — per-request token usage, latency, cost, and error tracking in the dashboard
- CLI + REST API — full account and configuration management from the command line or API
- Optional API-key auth — protect the proxy with its own API keys
- Flexible deployment — Docker, systemd, or a pre-compiled cross-platform binary (no runtime dependencies)
Full docs in docs/ — Getting Started, CLI Commands, Architecture, Configuration, Load Balancing, Combos, Troubleshooting.
This project builds on ideas and code from:
MIT — see LICENSE.