cc-opencodego-proxy

Tiny Bun proxy. Use OpenCode Go, OpenRouter, or Baseten inside Claude Code. Listens on 127.0.0.1:8082, accepts Anthropic-shaped /v1/messages, forwards upstream, re-frames OpenAI streams as Anthropic SSE.

Routing

Provider	Format	Path
OpenCode Go	Anthropic-native	/v1/messages passthrough
OpenCode Go	OpenAI-compatible	/v1/chat/completions + translator
OpenRouter	OpenAI-compatible	/v1/chat/completions + translator
Baseten	OpenAI-compatible	/v1/chat/completions + translator

OpenCode is hybrid per-model: anything in OPENCODE_OPENAI_MODELS routes through the OpenAI translator; rest passthrough. Inbound Claude model id is matched case-insensitively against opus/sonnet/haiku and routed to MODEL_OPUS/MODEL_SONNET/MODEL_HAIKU, with MODEL as fallback.

Install

bun run build                # host binary → dist/cc-opencodego-proxy
bun run build:all            # linux/darwin/windows × x64/arm64

Or run from source:

bun install && cp .env.example .env && bun run start

Config

Env	Default	Notes
MODEL	opencode/minimax-m2.7	Fallback. Must be <provider>/<model>
MODEL_OPUS / MODEL_SONNET / MODEL_HAIKU	—	Family override
OPENCODE_API_KEY	—	Required for opencode/*
OPENCODE_OPENAI_MODELS	—	CSV of opencode models routed via OpenAI translator
OPENCODE_BASE_URL	https://opencode.ai/zen/go/v1
OPENROUTER_API_KEY	—	Required for open_router/*
OPENROUTER_BASE_URL	https://openrouter.ai/api/v1
BASETEN_API_KEY	—	Required for baseten/*
BASETEN_BASE_URL	https://inference.baseten.co/v1
OPENCODE_MAX_TOKENS / OPENROUTER_MAX_TOKENS / BASETEN_MAX_TOKENS	0	Default max_tokens when client omits
ENABLE_THINKING	true	Forward thinking/reasoning blocks
PROVIDER_RATE_LIMIT / PROVIDER_RATE_WINDOW	40 / 60s	Rolling-window cap
PROVIDER_MAX_CONCURRENCY	5	In-flight stream cap
HTTP_READ_TIMEOUT / HTTP_CONNECT_TIMEOUT	120s / 5s
HOST / PORT	127.0.0.1 / 8082	Non-loopback requires ANTHROPIC_AUTH_TOKEN
ANTHROPIC_AUTH_TOKEN	—	Inbound auth gate (x-api-key / Authorization: Bearer)
LOG_FILE	—	Path or silent to mute

Provider prefixes valid: opencode/, open_router/, baseten/. Anything else fails fast.

Client setup

~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:8082",
    "ANTHROPIC_AUTH_TOKEN": "any-key"
  }
}

ANTHROPIC_AUTH_TOKEN is a placeholder unless the proxy's .env also sets it (then both must match — constant-time compared).

Behavior

• Translator: text, thinking (reasoning_content or inline <think>...</think>), tool calls. No images, no heuristic text→tool_use parser.
• Token estimate: chars/4 bias-high feeds message_start.usage.input_tokens and /v1/messages/count_tokens so Claude Code auto-compaction fires before upstream window overflows.
• Rate limit: rolling window + reactive 429 cooldown + concurrency semaphore. Strict FIFO.
• Retry: up to 3 attempts on 429. Honors Retry-After; otherwise 2s→60s exponential + jitter. Retries don't recharge the rate window.
• Stream cancel: client disconnect → upstream AbortSignal fires → generator finally releases concurrency slot.
• Auth: constant-time compare via crypto.timingSafeEqual. Inbound auth headers never forwarded upstream.

Binary

cc-opencodego-proxy            start on $HOST:$PORT
cc-opencodego-proxy --version
cc-opencodego-proxy --help
cc-opencodego-proxy --quiet | --log-file PATH

.env picked up from binary's working directory.

Auto-start hook (optional)

{
  "hooks": {
    "SessionStart": [{
      "matcher": "*",
      "hooks": [{ "type": "command", "command": "bash ~/.claude/hooks/start-proxy.sh", "timeout": 10 }]
    }]
  }
}

Hook script: curl /health, nohup binary if down, wait ~3s for port.

Development

bun install
bun run dev            # --watch
bun run check          # biome + tsc
bun run lint:fix
bun test               # 59 tests across sse/config/translate/rate-limit

Layout

src/
├── cli.ts                   binary entry + log sink
├── server.ts                Bun.serve + routes + auth
├── config.ts                env parsing, model resolver
├── rate-limit.ts            rolling window + concurrency cap
├── sse.ts                   frame builder + line parser
├── translate.ts             Anthropic ↔ OpenAI request/stream
├── types.ts
└── providers/
    ├── dispatcher.ts        per-request route by provider
    ├── anthropic-passthrough.ts
    └── openai-compat.ts     /chat/completions + re-framing
tests/
├── sse.test.ts
├── config.test.ts
├── translate.test.ts
└── rate-limit.test.ts

Intentional gaps

• No tiktoken — count_tokens uses chars/4, good enough for Claude Code quota probes.
• No image blocks in OpenAI translator path. Anthropic passthrough forwards verbatim.
• No text→tool_use heuristic; native tool_calls deltas only.
• No Discord/Telegram/voice bots.

License

See LICENSE.