| name | agent-browser |
| description | Use Vercel Labs agent-browser for web tasks, screenshots, page inspection, OAuth/device flows, and UI verification when the active runtime exposes the CLI. |
| runtime | agent-browser |
| requires_binary | agent-browser |
| capabilities | ["browser.open","browser.snapshot","browser.interact","browser.auth","browser.state"] |
agent-browser
Use the Vercel Labs agent-browser CLI when a task requires observing or operating a web page, including local Armorer UI testing, app setup pages, OAuth or device-auth flows, screenshots, form validation, and public web inspection.
Before running browser commands, load the canonical workflow guide from the installed CLI:
agent-browser skills get core
Do not invent browser commands. If agent-browser is unavailable, report that the runtime capability is unavailable.
Behavior
- Prefer Armorer-owned UI/API state when it answers the question directly.
- Use
agent-browser open, snapshot, find, click, fill, get, wait, and batch --json when visual state, page behavior, or a web login/setup flow matters.
- Element reads go through
get: use agent-browser get text @e1, agent-browser get html @e1, agent-browser get attr @e1 href, or agent-browser get value @e1. Do not call agent-browser text ... or agent-browser html ...; those are not valid top-level commands.
- Keep actions scoped to the user's explicit task and the active page/session.
- Summarize what the browser is doing in plain language.
- Verify the final state from the page or the backing API before reporting success.
Safety
- Do not submit credentials, payments, posts, messages, or destructive actions unless the user explicitly requested that exact action.
- If a page asks for a secret, prefer Armorer Vault or the product's secure auth flow over exposing the secret in chat.
- Do not scrape private data beyond what is needed for the user's task.
- Treat external sends, account changes, purchases, deletion, or public posting as approval-required.
Installed Agents
When this skill is enabled for an installed agent, the agent should use the Armorer-provided agent-browser runtime, session name, auth profile, action policy, and environment. Raw credentials stay in Armorer Vault; the agent may use only credential readiness/profile names surfaced by Armorer.
Installed agents must call Armorer browser tools such as armorer_agent_browser_run, armorer_agent_browser_links, and armorer_agent_browser_form_submit; they must not run raw terminal agent-browser commands, because that bypasses Armorer's profile lease, CloakBrowser runtime selection, policy checks, and observability.
