Model Gateway API Reference | 51cowork API Docs

Quick answer

Send model requests directly to https://api.51cowork.com. Prefer Authorization: Bearer; Anthropic clients may use x-api-key. Discover exact model IDs with GET /v1/models.

Gateway: https://api.51cowork.com
Auth: Bearer / x-api-key
Content type: application/json
Models: GET /v1/models

Base URL and authentication

Send model requests to /v1/... paths on the Gateway at https://api.51cowork.com. Manage your account in Dashboard.

Environmentbash

export COWORK_GATEWAY_URL="https://api.51cowork.com"
export COWORK_API_KEY="<key-from-dashboard>"
export COWORK_MODEL="<model-from-gateway>"

Client	Base URL	Path appended by the client
Direct HTTP	https://api.51cowork.com	Use the full /v1/... path
OpenAI SDK / OpenCode / Cursor	https://api.51cowork.com/v1	/chat/completions or /responses
Cherry Studio (OpenAI)	https://api.51cowork.com	/v1/chat/completions
WorkBuddy (Custom Protocol)	https://api.51cowork.com/v1/chat/completions	Full URL; no appended path
Anthropic SDK / Claude Code	https://api.51cowork.com	/v1/messages
Codex	https://api.51cowork.com	wire_api = responses

Method	Header
Recommended	Authorization: Bearer <API_KEY>
Anthropic compatible	x-api-key: <API_KEY>
Gemini compatible, when enabled	x-goog-api-key: <API_KEY>

GET /v1/models

Read the models currently exposed to the Key's assigned group.

cURLbash

curl "$COWORK_GATEWAY_URL/v1/models" \
  -H "Authorization: Bearer $COWORK_API_KEY"

Supported model families include GPT, Claude, Grok, DeepSeek, GLM, and Kimi.
Use an exact returned model ID; do not send a family name or infer an ID from a product name.
If the list is empty or a model is missing, check the Key, group, and operator configuration.
Cache briefly and provide a manual refresh path.

POST /v1/messages

Anthropic Messages-compatible endpoint.

Anthropic Messagesbash

curl "$COWORK_GATEWAY_URL/v1/messages" \
  -H "x-api-key: $COWORK_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "'"$COWORK_MODEL"'",
    "max_tokens": 256,
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Before generating, send the same model and messages to POST /v1/messages/count_tokens to estimate input tokens.

Count tokensbash

curl "$COWORK_GATEWAY_URL/v1/messages/count_tokens" \
  -H "x-api-key: $COWORK_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "'"$COWORK_MODEL"'",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Field	Requirement
model	An ID returned by /v1/models
max_tokens	Positive integer, bounded by client and model support
messages	At least one valid message
stream	Optional; true returns an event stream

POST /v1/chat/completions

OpenAI Chat Completions-compatible endpoint.

Chat Completionsbash

curl "$COWORK_GATEWAY_URL/v1/chat/completions" \
  -H "Authorization: Bearer $COWORK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "'"$COWORK_MODEL"'",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Common fields include model, messages, stream, and model-supported sampling or tool fields. Unsupported upstream fields may be rejected or handled by compatibility rules, so do not assume identical behavior across models.

POST /v1/responses

OpenAI Responses-compatible endpoint and the wire API used by Codex.

Responsesbash

curl "$COWORK_GATEWAY_URL/v1/responses" \
  -H "Authorization: Bearer $COWORK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "'"$COWORK_MODEL"'",
    "input": "Explain this repository in three bullets."
  }'

Streaming

Set stream to true and process server-sent events incrementally. cURL uses -N to disable output buffering.

Streaming requestbash

curl -N "$COWORK_GATEWAY_URL/v1/chat/completions" \
  -H "Authorization: Bearer $COWORK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "'"$COWORK_MODEL"'",
    "stream": true,
    "messages": [{"role": "user", "content": "Count to five"}]
  }'

Use a read timeout long enough for first-token and long-response latency.
Close the response body when the client cancels.
An error after streaming starts may not be rewritten as a normal JSON error; retain received events for diagnosis.

HTTP status and error handling

Status	Typical meaning	Action
400	Invalid JSON, field, or model parameter	Fix the request; do not retry unchanged
401	Missing, invalid, or disabled Key	Reload configuration; replace a rotated Key
402 / 403	Insufficient credit or access	Check Key quota and account status
404	Endpoint, model, or group capability not found	Refresh models and verify the path
429	Rate limited or upstream temporarily unavailable	Honor Retry-After and back off with jitter
5xx	Temporary gateway or upstream failure	Record context and retry only a few times