Gateway API (OpenAI-compatible)

The gateway exposes each of your workspaces as an OpenAI-compatible model, so any OpenAI-style client can run agent tasks in a real workspace.

Endpoint	Behavior
GET /v1/models	Lists the key owner's workspaces as models.
POST /v1/chat/completions	Runs an agent task in the selected workspace. Streams SSE by default.

curl http://localhost:8000/v1/chat/completions \
  -H "Authorization: Bearer sk-cptr-..." \
  -H "Content-Type: application/json" \
  -d '{"model": "cptr/my-project", "messages": [{"role": "user", "content": "Summarize the README"}]}'

Connecting Open WebUI? Follow the walkthrough at Use Computer from Open WebUI; this page is the endpoint reference.

API keys

Create keys in Settings → Admin → Gateway. The sk-cptr-... value is shown once and stored hashed, so copy it immediately. Key management itself (creating, revoking) requires a browser session; you can't manage keys with a bearer key.

Workspaces as models

Each workspace owned by the key's user appears as cptr/<folder-name>; duplicate folder names get a -2 suffix. Which underlying model executes the task, in priority order:

1. The model set in Settings → Admin → Gateway
2. A <workspace>/.cptr/model override file
3. The default chat model
4. The first model of the first enabled connection

Conversation continuity

Stateless requests create a new Computer chat each time. To continue a conversation, send X-Chat-Id (any client) or the Open WebUI header set:

{
  "X-OpenWebUI-Chat-Id": "{{CHAT_ID}}",
  "X-OpenWebUI-Message-Id": "{{MESSAGE_ID}}",
  "X-OpenWebUI-User-Message-Id": "{{USER_MESSAGE_ID}}",
  "X-OpenWebUI-User-Message-Parent-Id": "{{USER_MESSAGE_PARENT_ID}}",
  "X-OpenWebUI-Task": "{{TASK}}"
}

{{USER_MESSAGE_ID}}, {{USER_MESSAGE_PARENT_ID}}, and {{TASK}} require Open WebUI 0.10.0 or newer. Without them, basic chat works, but edit/regeneration branches and utility-task filtering don't.

Requests that Open WebUI marks as title, tag, or follow-up generation (via X-OpenWebUI-Task) are answered by the plain LLM without starting an agent task in the workspace.

Limits

Gateway tasks run with full tool approval

There is no per-action approval in the OpenAI-compatible API: the agent edits files, runs commands, and calls tools without asking. Treat a gateway key like an SSH credential and only hand it to clients you trust.

• temperature, top_p, and max_tokens are accepted but ignored; the workspace's configured model and settings apply.
• usage token counts in responses are always 0.
• Streams idle for 300 seconds are closed.