llm-proxy-openai

@computeragent/llm-proxy-openai

Anthropic Messages ↔ OpenAI Chat Completions translator proxy.

Accepts POST /v1/messages in Anthropic's Messages format and forwards to any OpenAI-Chat-Completions–compatible endpoint, translating both the request and the response (including streaming and tool calls).

Used by ComputerAgent so that engines whose underlying SDKs speak Anthropic Messages — claude-agent-sdk and deepagents (via ChatAnthropic) — can target OpenAI-compat backends like vLLM, LiteLLM, Together, Ollama, or any compatible inference gateway.

gitagent does not need this proxy — gitclaw natively speaks both protocols via GITCLAW_MODEL_BASE_URL + provider:model@baseUrl syntax.

Run as a CLI

UPSTREAM_BASE=https://your-openai-compat-host \
UPSTREAM_PATH=/v1/chat/completions \
UPSTREAM_TOKEN=sk-... \
PORT=8788 \
  npx @computeragent/llm-proxy-openai

Then point your ComputerAgent run at the proxy:

curl -N -X POST http://127.0.0.1:18800/run \
  -H 'content-type: application/json' -H 'accept: text/event-stream' \
  -d '{
    "source": "github.com/<org>/<gap-repo>",
    "harness": "claude-agent-sdk",
    "envs": {
      "ANTHROPIC_BASE_URL": "http://127.0.0.1:8788",
      "ANTHROPIC_API_KEY":  "via-proxy"
    },
    "message": "Reply: PING"
  }'

Run programmatically

import { startProxy } from "@computeragent/llm-proxy-openai";

const proxy = await startProxy({
  port: 8788,
  upstream: {
    base: "https://your-inference-gateway.example.com",
    path: "/v1/chat/completions",
    token: process.env.UPSTREAM_TOKEN!,
    modelOverride: "your-model-id", // optional; override client model
  },
});

// …
await proxy.close();

Environment variables (CLI)

Var	Required	Default	Notes
UPSTREAM_BASE	✓	—	Origin only, e.g. https://your-host
UPSTREAM_TOKEN	✓	—	Bearer token for the upstream
UPSTREAM_PATH		/v1/chat/completions	Some hosts use /v4/chat/completions
UPSTREAM_MODEL		—	Force this model on every upstream request (overrides client's Anthropic model)
UPSTREAM_AUTH_SCHEME		Bearer	Replace with e.g. Token if your backend wants something else
PORT		8788	Bind port
FORWARD_MAX_TOKENS		0	Set to 1 to forward max_tokens — some backends blank the response when this field is set, so it's off by default

What the proxy translates

Direction	Anthropic shape	OpenAI shape
Request	system field	first system message
Request	messages[].content text blocks	messages[].content strings
Request	assistant tool_use blocks	messages[].tool_calls
Request	user tool_result blocks	messages[] with role: "tool" + tool_call_id
Request	tools[] (Anthropic input_schema)	tools[] (OpenAI function.parameters)
Request	tool_choice: {type:"auto" | "any" | "tool"}	tool_choice: "auto" | "required" | {type:"function",function:{name}}
Response	id, usage, content blocks	reassembled from choices[0].message
Response	tool_use content blocks	from tool_calls
Response	stop_reason: "tool_use" | "max_tokens" | "end_turn"	mapped from finish_reason
Streaming	message_start → content_block_start/delta/stop → message_delta → message_stop	reassembled from chat.completion.chunk deltas (text + tool_calls.function.arguments partial JSON)

Endpoints

• GET /health → { ok: true, upstream }
• POST /v1/messages → translated and forwarded to <UPSTREAM_BASE><UPSTREAM_PATH>

Anything else → 404.

Limitations

• Text + tool calls only. No image inputs, no audio. Anthropic's content blocks for those types are passed through as JSON strings if encountered, which most OpenAI-compat backends will reject.
• No retry / circuit-breaker. A single upstream timeout fails the request.
• Stateless. No request-id correlation, no metrics export. Bring your own observability.
• session_id is informational. The proxy adds it to outgoing OpenAI requests for backends that track it, but it's a string derived from the timestamp — proper session continuity comes from the client replaying messages[], not from server-side state in the upstream.

Verified compatibility

Tool calls round-trip end-to-end with claude-agent-sdk and deepagents against OpenAI-compat backends serving /v1/chat/completions. Streaming + tool-use partial-JSON deltas reassemble correctly.