Connect Gavana to ChatGPT
Gavana provides a hosted MCP endpoint with browser OAuth. It keeps the normal visual workflow visible: an agent can inspect a canvas, hand you a verified review link, and return durable Gavana handles after you approve a change.
Start with the strict read-only endpoint when you want an agent to review rather than create:
https://app.gavana.ai/mcp/readonlyUse the normal endpoint when you intend to save work or explicitly run generation:
https://app.gavana.ai/mcpConnect
-
In an MCP-capable ChatGPT workspace, open the custom connector setup. Your workspace may require an administrator to allow custom connectors.
-
Add one Gavana endpoint:
https://app.gavana.ai/mcp/readonly -
Choose Connect. Gavana opens a sign-in and consent screen.
-
Review the requested permissions, then approve the connection. The read-only endpoint asks only for canvas inspection. The normal endpoint requests broad canvas, asset, image, video, and job permissions by default, so use it only when you intend to create work.
ChatGPT returns to the plugin after the OAuth authorization-code flow completes. You never paste an Agent Access token into ChatGPT. OAuth uses PKCE and an exact callback binding.
Available tools
| Tool | What it does | Side effect |
|---|---|---|
canvas_list, canvas_get, get_canvas_image, open_canvas | Find, inspect, preview, or open an accessible canvas | Read only |
save_image_to_canvas | Save a ChatGPT or public HTTPS image as a durable image node | Writes canvas state |
create_canvas_workflow | Save a private reusable workflow card | Writes canvas state; never generates |
run_canvas_workflow, get_canvas_workflow_run | Explicitly run or check a workflow | A run may incur provider cost |
generate_image_in_canvas | Generate an image into a canvas | May incur provider cost |
find_video_models, generate_video, get_video_job | Discover a model, explicitly start video, or check it | Video may incur provider cost |
The read-only endpoint exposes only the first row. The normal endpoint exposes the complete hosted catalog above. The persistent Agent Canvas is the default destination when no specific canvas is selected.
Every save or generation call uses a caller-stable idempotency key, so retrying the exact same intended request does not create duplicate work. A workflow is deliberately two steps: create_canvas_workflow saves it; run_canvas_workflow is the separate explicit execution step.
What to ask
List my Gavana canvases, inspect the most relevant one for this brief, and open it for review. Do not change anything.Save this image to my Agent Canvas, then create a reusable workflow with the saved image as a fixed default. Do not run it yet.I approve one video generation. Find a connected model with image-to-video support, use this saved Gavana image as the first frame, and return the job status and review link.The agent should inspect first, use exact returned handles, ask before a provider-backed run, and never automatically retry a failed paid request.
Disconnect
Disconnect the connector in ChatGPT to stop new use from that client. You can also revoke the Gavana OAuth delegation for that client from Gavana’s Agent Access settings. Revocation takes effect on the next request.
Security notes
- The hosted endpoint accepts OAuth bearer tokens only; do not use Agent Access tokens here.
- The normal endpoint is
https://app.gavana.ai/mcp; the safer inspection endpoint ishttps://app.gavana.ai/mcp/readonly. - Image imports accept public HTTPS URLs and reject private or local network destinations.
- Image, workflow, and video runs can consume credits or incur charges with the provider you connected to Gavana. They require explicit approval in the current conversation.