@fastagent/co2 is a local CLI gateway that translates between OpenAI-compatible and Claude-compatible protocols, so existing clients can talk to the other upstream without changing their request protocol. It runs as a local HTTP service and supports two modes: o2c (OpenAI request -> Claude upstream) and c2o (Claude request -> OpenAI upstream).

npm install -g @fastagent/co2

Or run it directly without installing:

npx @fastagent/co2 start --config ./co2.config.json

1. Create co2.config.json

{
  "server": {
    "host": "127.0.0.1",
    "port": 8000,
    "mode": "openai-to-claude",
    "logLevel": "info"
  },
  "providers": {
    "openai": {
      "apiKey": "OPENAI_API_KEY_PLACEHOLDER",
      "baseUrl": "https://api.openai.com/v1",
      "defaultHeaders": {
        "user-agent": "co2-cli/0.3.2"
      }
    },
    "anthropic": {
      "apiKey": "ANTHROPIC_API_KEY_PLACEHOLDER",
      "baseUrl": "https://api.anthropic.com",
      "version": "2023-06-01",
      "defaultHeaders": {
        "user-agent": "co2-cli/0.3.2"
      }
    }
  },
  "routing": {
    "defaultOpenAIModel": "gpt-5.4",
    "defaultClaudeModel": "claude-opus-4.6",
    "openAIReasoningEffort": "high",
    "openAIUpstreamApi": "responses",
    "openAIParallelToolCalls": true,
    "claudeOutputEffort": "high",
    "skipInboundFields": {
      "claudeMessages": ["context_management"],
      "openAIResponses": [],
      "openAIChatCompletions": []
    }
  },
  "modelMap": {
    "claude-opus-4.6": "gpt-5.4",
    "gpt-5.4": "claude-opus-4-6"
  }
}

Notes:

• The example shows both openai and anthropic providers so the full config shape is visible in one place.
• In openai-to-claude / o2c, only providers.anthropic is used. providers.openai can be omitted without affecting startup or request handling.
• In claude-to-openai / c2o, only providers.openai is used. providers.anthropic can be omitted without affecting startup or request handling.
• routing.defaultClaudeModel and routing.claudeOutputEffort only affect o2c. routing.defaultOpenAIModel, routing.openAIReasoningEffort, routing.openAIUpstreamApi, and routing.openAIParallelToolCalls only affect c2o.
• routing.openAIUpstreamApi explicitly selects the OpenAI upstream contract for c2o: the default is responses; if your OpenAI-compatible upstream only exposes POST /v1/chat/completions, set it to chat-completions. This is the switch you want when routing Claude Code through c2o into a chat-completions-only upstream.
• routing.openAIParallelToolCalls is a global hard override for c2o tool requests: if omitted, co2 does not send parallel_tool_calls; if set to true or false, co2 forwards that exact value to OpenAI only when tools are present.
• routing.skipInboundFields lets you explicitly drop known top-level request fields before validation, so you can keep a local gateway working with newer SDK/client fields without waiting for a new co2 release.

skipInboundFields is split by inbound protocol:

• claudeMessages: applies to POST /v1/messages
• openAIResponses: applies to POST /v1/responses
• openAIChatCompletions: applies to POST /v1/chat/completions

Behavior:

• Matching is exact and only applies to top-level fields.
• When a field is skipped, co2 removes it at the boundary, logs a warn, and continues processing the request.
• This is intended for fields that are known to be sent by real clients but are not yet modeled by the gateway.
• It does not relax typo protection for other fields; unconfigured misspellings such as thinkingg still fail validation.

Common examples:

• Claude Code currently sends context_management on some c2o /v1/messages requests. Add it to skipInboundFields.claudeMessages to keep those requests working.
• If a future OpenAI SDK starts sending a new top-level field on /v1/responses or /v1/chat/completions, add that exact field name to the corresponding skip list instead of waiting for a new release.

2. Start the server

co2 start --config ./co2.config.json

3. Call the route that matches the current mode

• openai-to-claude: POST /v1/chat/completions, POST /v1/responses
• claude-to-openai: POST /v1/messages

• o2c = OpenAI request -> Claude upstream
• c2o = Claude request -> OpenAI upstream
• The mode name is always incoming protocol -> upstream protocol; response protocol stays aligned with the incoming side by default.

What protocol your client speaks	What upstream you want	Mode to use
OpenAI chat/completions / responses	Claude	o2c
Claude messages	OpenAI	c2o

Common cases:

• OpenAI SDK and other OpenAI-compatible clients usually use o2c.
• Claude-compatible clients and Claude messages clients usually use c2o.

curl http://127.0.0.1:8000/v1/responses \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "gpt-5.4",
    "input": [
      {
        "role": "user",
        "content": [
          { "type": "input_text", "text": "Hello" }
        ]
      }
    ],
    "instructions": "You are concise."
  }'

After changing server.mode to claude-to-openai:

curl http://127.0.0.1:8000/v1/messages \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "claude-opus-4.6",
    "max_tokens": 128,
    "messages": [
      { "role": "user", "content": "Hello" }
    ]
  }'

co2 currently supports image input translation on all request paths:

• o2c /v1/responses
• o2c /v1/chat/completions
• c2o /v1/messages

Supported image sources:

• https://... / http://...
• data:image/...;base64,...

Not supported in V1:

• image output
• image generation
• audio output
• file-id based image inputs

When an image block or multimodal control field cannot be mapped safely, co2 drops just that field/block, logs a warn, and continues only if the message still contains supported content.

curl http://127.0.0.1:8000/v1/responses \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "gpt-5.4",
    "input": [
      {
        "role": "user",
        "content": [
          { "type": "input_text", "text": "Describe this image" },
          { "type": "input_image", "image_url": "https://example.com/a.png" }
        ]
      }
    ]
  }'

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "gpt-5.4",
    "messages": [
      {
        "role": "user",
        "content": [
          { "type": "text", "text": "Describe this image" },
          { "type": "image_url", "image_url": { "url": "data:image/png;base64,QUFBQQ==" } }
        ]
      }
    ]
  }'

curl http://127.0.0.1:8000/v1/messages \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "claude-opus-4.6",
    "max_tokens": 128,
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "image",
            "source": {
              "type": "base64",
              "media_type": "image/png",
              "data": "QUFBQQ=="
            }
          },
          { "type": "text", "text": "Describe this image" }
        ]
      }
    ]
  }'

Notes:

• For production, prefer environment variables for API keys; environment variables take precedence over the config file.
• ANTHROPIC_AUTH_TOKEN is accepted as a compatibility alias, but if it exists together with ANTHROPIC_API_KEY, they must be identical.
• Node.js >= 20.19.0 is required.