-
Notifications
You must be signed in to change notification settings - Fork 0
Gen3 Gt Api Keys And Access
- Confirm GT API is enabled for your account. Without
gtApiEnabled, the GT API sidebar route and key-management workspace are not available. - Open GT API at
/gt-apiand select the API Keys tab. - Choose Create API Key (wizard preset or manual form).
- Pick a purpose: inference, upload, or conversation-facing.
- Set inference scopes (
inference:chat,inference:embed,inference:transcribe,inference:speech,inference:image) to match what the client will actually call. - Allowlist agent aliases, raw-model catalog names, and/or dataset targets the key may reach—then save and copy the secret immediately (shown once).

API keys are the credential boundary for every external caller. Scoped keys keep discovery (GET /v1/models), chat, embeddings, uploads, and conversation flows least-privilege and attributable. Conversation-facing keys can inherit runtime allowlists (and prompt policy) from linked inference and upload keys so operators do not duplicate policy on every app integration.
In-app workspace: GT API → API Keys.
Full documentation (compatibility matrix, per-client runbooks, API reference) is in Instructions under GT API—see Overview, Compatibility Matrix, and API Reference.
| Purpose | Use when |
|---|---|
| Inference | Stateless or stateless-first OpenAI clients (/v1/models, /v1/chat/completions, optional embeddings/speech/image routes). |
| Upload | Automation that writes into fixed datasets (POST /v1/datasets/{id}/files). |
| Conversation | One external app credential that inherits linked inference + upload keys and supports GT conversation bootstrap and conversation-scoped files. |
The in-app wizard maps common clients to presets, including:
- OpenAI-compatible chat client — chat scope only; best for OpenAI SDK, curl, LiteLLM, Open WebUI, LibreChat.
-
Chat plus embeddings — adds
inference:embedfor LangChain, LlamaIndex, Dify, Semantic Kernel, and similar. - Agentic IDE or editor — dedicated chat key for Continue, Cursor, Cline/Roo, Aider, Manus, OpenClaw.
- Voice and audio — chat + transcribe + speech on one published multimodal alias.
-
Image generation —
inference:imagefor/v1/images/generations. - Dataset upload automation — upload purpose for n8n, Langflow, Flowise ingestion-only flows.
-
Conversation app with uploads — conversation purpose; requires reusable inference and upload keys first, then
POST /v1/conversations+X-GT-Conversation-Id.
Inference keys may attach prompt guardrail modules and optional additional prompt text. GT applies these server-side together with the agent (or raw-model policy) for allowed calls. Conversation-facing keys inherit linked inference key prompt modules and extra text.
- Rotate keys on a schedule appropriate to shared UIs or CI secrets.
- Revoke keys when an integration is decommissioned; callers should fail closed without fallback credentials.
- Use one key per integration or environment so usage and access review stay traceable.
Before creating keys, confirm:
- Published Endpoints exposes the inference names callers need.
- Datasets exist if the key will use upload routes.
- Linked runtime keys exist before creating a conversation-facing key.