fuseproxy

A lightweight, observable proxy for Anthropic's POST /v1/messages API.

fuseproxy sits between your internal clients and Anthropic, validates inbound requests, forwards them to Anthropic, and emits tracing + generation telemetry to Langfuse via OpenTelemetry.

What It Does

• Exposes POST /v1/messages on http://localhost:5678
• Requires bearer auth using an internal API key (INTERNAL_API_KEY)
• Validates request bodies with Zod before forwarding
• Supports both non-streaming and streaming responses
• Streams Anthropic events as Server-Sent Events (SSE) when stream=true
• Records spans/generations and token/cost metadata to Langfuse

Tech Stack

• Runtime: Bun + TypeScript
• HTTP server: Express 5
• LLM provider: @anthropic-ai/sdk
• Observability: OpenTelemetry (@opentelemetry/sdk-node) + Langfuse (@langfuse/otel, @langfuse/tracing)
• Validation: Zod
• Logging: tslog

Project Structure

• src/index.ts: app entrypoint
• src/api.ts: Express server and /v1/messages route
• src/types.ts: request schema + model pricing table
• src/instrumentation.ts: OpenTelemetry/Langfuse setup
• src/logger.ts: structured logger with trace/span IDs
• compose.yaml: local Langfuse stack (web, worker, postgres, redis, clickhouse, minio)

Prerequisites

• Bun installed
• Anthropic API key
• Docker + Docker Compose (optional, if running local Langfuse)

Environment Variables

Minimum required to run the proxy:

• INTERNAL_API_KEY: bearer token your clients must send
• ANTHROPIC_API_KEY: key used for upstream Anthropic calls (can also be from an Anthopric-compatible provider like Kimi or DeepSeek)

For Langfuse tracing export:

• LANGFUSE_BASE_URL (for local stack: http://localhost:3000)
• LANGFUSE_PUBLIC_KEY
• LANGFUSE_SECRET_KEY

Optional:

• DEPLOYMENT_ENVIRONMENT (recorded in tracing metadata)
• LOG_LEVEL (silly|trace|debug|info|warn|error|fatal, default debug)
• ANTHROPIC_BASE_URL (Base URL for the Anthropic API, supports all Anthropic-compatible providers, like Kimi or DeepSeek)

Install

From NPM:

npm install -g @cle-does-things/fuseproxy

From source:

git clone https://github.com/AstraBert/fuseproxy
cd fuseproxy
bun install

Run

If installed with NPM:

fuseproxy run

With npx:

npx @cle-does-things/fuseproxy run

If installed from source:

bun run start

Server listens on http://localhost:5678.

Example Request

curl -X POST http://localhost:5678/v1/messages \
  -H "Authorization: Bearer $INTERNAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4-6",
    "max_tokens": 256,
    "system": "You are concise.",
    "messages": [{"role":"user","content":"Say hello in one sentence."}],
    "stream": false
  }'

Integrations

GitHub Copilot CLI

Follow these steps to configure Copilot to route its LLM requests through fuseproxy:

1. Start the proxy:

bun run start

2. In another shell, load Copilot env vars:

export COPILOT_PROVIDER_TYPE="anthropic" && \
    export COPILOT_PROVIDER_BEARER_TOKEN="<internal fuseproxy bearer token>" && \
    export COPILOT_MODEL="claude-opus-4-7" && \
    export COPILOT_PROVIDER_BASE_URL=http://localhost:5678 && \
    export COPILOT_PROVIDER_MAX_OUTPUT_TOKENS=32000 && \
    export COPILOT_PROVIDER_MAX_PROMPT_TOKENS=64000

3. Confirm values are set:

echo "$COPILOT_PROVIDER_TYPE"
echo "$COPILOT_PROVIDER_BASE_URL"
echo "$COPILOT_MODEL"

Important: COPILOT_PROVIDER_BEARER_TOKEN must match INTERNAL_API_KEY used by the proxy, or requests will return 401 Unauthorized.

Pi Agents Setup

You can set up fuseproxy for your Pi agents by adding this to your ~/.pi/agent/models.json config file:

{
  "providers": {
    "fuseproxy": {
      "baseUrl": "http://localhost:5678",
      "api": "anthropic-messages",
      "apiKey": "fuseproxy", // <- not needed, just a placeholder
      "headers": {
        "x-api-key": "YOUR_INTERNAL_API_KEY" // <- must match with the internal API key
      },
      "models": [
        {
          "id": "claude-opus-4-7",
          "name": "Claude Opus 4.7",
          "reasoning": true,
          "input": ["text", "image"],
          "contextWindow": 64000,
          "maxTokens": 32000,
          "cost": {
            "input": 5,
            "output": 25,
            "cacheRead": 0.5,
            "cacheWrite": 6.25
          }
        },
        {
          "id": "kimi-k2.6",
          "name": "Kimi K2.6",
          "reasoning": true,
          "input": ["text"],
          "contextWindow": 64000,
          "maxTokens": 32000,
          "cost": {
            "input": 0.95,
            "output": 4,
            "cacheRead": 0.16,
            "cacheWrite": 0
          }
        }
      ]
    }
  }
}

Streaming Behavior

When stream=true, the service:

• Sets Content-Type: text/event-stream
• Forwards Anthropic stream events as SSE (event: <type>, data: <json>)
• Emits an error event and closes the stream on failure

Local Langfuse Stack (Optional)

A full local Langfuse stack is provided in compose.yaml.

1. Review compose.yaml and replace every # CHANGEME secret.
2. Start services:

docker compose up -d

3. Open Langfuse at http://localhost:3000.

Default exposed ports include:

• 3000: Langfuse web
• 3030: Langfuse worker
• 9090: MinIO S3 endpoint

Development Commands

bun run lint
bun run prettier
bun run prettier:check

Notes

• The proxy currently implements Anthropic-style /v1/messages only.
• src/types.ts includes a MODEL_COSTS table used to estimate generation cost metadata for traces.
• Graceful shutdown flushes Langfuse spans on SIGINT/SIGTERM.