browser-automation · diff

git:20260520.a13d1dd to git:20260524.673ea7f

5 added, 4 removed. Audit A to A.

---

name: browser-automation

description: >-

Drive Chrome via CDP for login flows, SPAs, and screenshots. Use when fetch tools

are insufficient or the user mentions browser automation.

disable-model-invocation: true

---

# Browser automation (CDP)

Use **CDP browser tools** when you need a real Chrome tab: authenticated pages, dynamic JS, or visual verification.

## When to use CDP vs fetch tools

| Need | Tool |

|------|------|

| Public HTML, no login | `fetch_web_content` (browser executor) |

| Login, SPA, CORS-blocked | `browser_navigate` + `browser_snapshot` |

| Visual proof | `browser_screenshot` |

## Workflow

1. `browser_list` — confirm Chrome is on `--remote-debugging-port` (default `9222`).

- 2. `browser_navigate` — open URL (localhost allowlist by default).

- 3. `browser_snapshot` — get `[uid]` markers for elements.

- 4. `browser_click` / `browser_fill` — act on uids from the latest snapshot.

- 5. `browser_screenshot` — PNG appears inline in chat.

+ 2. For **new external origins**, call **`ask_question`** first (options `once`, `persist`, `deny`), then **`request_browser_origin_access`** with `{ url, decision: "once"|"persist" }`, then **`browser_navigate`**.

+ 3. If navigation was blocked, repeat the **`ask_question`** flow or call **`request_browser_origin_access`** without `decision` (shows the same question cards).

+ 4. `browser_snapshot` — get `[uid]` markers for elements.

+ 5. `browser_click` / `browser_fill` — act on uids from the latest snapshot.

+ 6. `browser_screenshot` — PNG appears inline in chat.

Set `MINNOW_BROWSER_URL` or pass `browser_url` on each call. Enable tools under **Settings → Browser (CDP)**.

Troubleshooting: [opencode-browser](https://github.com/different-ai/opencode-browser).