| name | agent-browser |
| description | Agent-browser is an LLM-driven browser-automation framework (DOM querying, form filling, navigation) that complements lightpanda-headless-browser for heavier semantic interactions. |
| source | https://github.com/browser-use/browser-use |
| integration | external-browser |
| status | descriptor-only |
agent-browser
Descriptor-only registration (P12 batch). Companion to the existing lightpanda-headless-browser skill.
Division of responsibility
| Concern | Tool | Scope |
|---|
| Liveness / "is the page alive" | lightpanda-headless-browser | ultra-fast Zig/Rust binary, no JS eval heavy-lifting |
| Semantic navigation | agent-browser | LLM drives click/type/select with accessibility-tree awareness |
| Hard assertions | playwright-webui-qa (existing) | deterministic end-to-end verification |
OctoCode's webui-e2e CI job (added in P9) already uses Playwright; agent-browser is a candidate for exploratory flows where the tests are not pre-scripted but discovered by the agent.
Integration plan (deferred to P13+)
- Expose
agent_browser_visit(url, goal) as a high-risk tool behind RuntimeConfig.integrations.agent_browser.enabled.
- Enforce the same HTTPS allowlist as
skills-install (see crates/octocode-cli/src/skills_install.rs), plus a page-size cap.
- Page snapshots are written through
WorkspaceToolExecutor (same path-traversal guardrails as file-writes).
Safety notes
- No auto-submission of forms that contain
password, card, ssn โ reject at tool level.
- Every invocation must emit an
event_feed_json entry for operator audit.
- Use
lightpanda-headless-browser first as a liveness gate before spinning up an LLM-driven session, per verification instructions the workspace already follows (WebUI verification step).
