Skip to content

Browser automation

Control a Chromium pane from an agent or script inside Graffitty. To give an agent the instructions, ask it to run gft skill chromium.

Open a page

gft chromium new --url http://localhost:3000

Copy the returned browser ID. The examples below use %3; replace it with yours.

Inspect, then interact

gft chromium snapshot --target %3

Use the returned element references to fill or click:

gft chromium fill --target %3 --ref <snapshot-ref> --text hello
gft chromium click --target %3 --ref <snapshot-ref>

Replace <snapshot-ref> with a reference from your snapshot. Commands return JSON; failures exit nonzero. Use --help for options.

Recover when something changes

Problem Next step
stale_ref Take a new snapshot and use its references
Page stops responding Reload or navigate again
Browser process stops Choose Restart in its pane
Target closes Create or select a live browser

Snapshots cover the main document, excluding cross-origin frames and closed shadow roots. Commands do not switch targets automatically. Restarting loses live page state; persistent profiles keep data already saved.

Other interfaces

  • Playwright attachment borrows an existing page.
  • gft mcp provides optional stdio MCP tools. Launch it in the agent's inherited workspace environment; do not save workspace tokens in global configuration.
  • Method reference lists RPC fields.
  • Host API describes permissions and request envelopes.