Gavana Agent Quickstart
Use this path when you want an agent to understand a Gavana canvas before either of you changes it. It is deliberately read-only: the agent can list, read, preview, and open canvases, but cannot write, generate, cancel, or spend provider credits.
1. Connect the read-only MCP
Use this endpoint in an OAuth-capable remote MCP client:
https://app.gavana.ai/mcp/readonlyComplete the browser sign-in and consent screen. You do not paste an Agent Access token into this connection.
For a local Codex profile, the CLI can install the same safe endpoint:
gavana mcp install codex --read-only2. Give the agent a bounded task
Paste a request like this—replace the bracketed text with your goal:
Connect to Gavana in read-only mode. List my canvases, inspect the one most relevant to [goal], and summarize its nodes, connections, current revision, and the safest next step. Do not make changes or start generation.The expected operation order is:
canvas_list → canvas_get → optionally get_canvas_image or open_canvasThe agent should use the exact handles returned by Gavana, not titles or reconstructed IDs. open_canvas gives you a verified link for visual review.
3. Decide whether to continue
If you only need a review, you are done. To create durable work, reconnect with the normal hosted endpoint, a local MCP, the CLI, or the raw API. Say what may happen next—for example, “save this image to my Agent Canvas” or “run this workflow once.”
Agent rule: Read before a write. Ask for approval before any provider-backed image, workflow, or video generation. Do not retry a failed paid request automatically.
What comes back
Read operations return structured canvas data and stable handles such as canvas:, node:, and asset:. These are safe references to use in the next Gavana operation. Temporary image previews are for review; the durable asset or node handle is the record to keep.