メニュー
Browser automation using Chrome DevTools MCP and the agent-browser CLI. Use when the user needs to navigate websites, fill forms, take screenshots, scrape data, or test web applications.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Backlog-scale orchestration on top of /workflow — drives PM items through refine → plan → implement → review → local-merge via role-specialized sub-agents dispatched in parallel waves, each item isolated in its own git worktree. Bare /swarm summarizes state; /swarm <goal> runs the orchestrator; /swarm:setup sets up the charter and .agent-tools/ umbrella; /swarm:continue resumes a paused run.
Capture durable knowledge from any engineering work — debugging solutions, refactors, features, design decisions, reusable patterns — and maintain memory quality so each unit of work compounds the next
Principal workflow entry — resolve portfolio mode (swarm resume/handoff or unit state machine), drive work without inventing path; hard-stop when path is not established
Initialize and maintain planning root (.agent-tools/planning preferred), conventions (tracks, gates, integration), runs ledger scaffold, and shared memory under .agent-tools/memory/ (AGENTS.md memory-link) for the workflow family.
Parent skill for the workflow family — principal session entry is /workflow:continue (also bare /workflow). Horizon mapping, brainstorm, refine, plan, execute, review, audit, compound. Vertical-slice and deliverable-partition modes.
Detect gaps in the canonical skill corpus, propose targeted improvements, validate changes, and present for review. Inventories src/** skill trees; optional run-ledger seeds for process IP gaps.
| name | use-browser |
| description | Browser automation using Chrome DevTools MCP and the agent-browser CLI. Use when the user needs to navigate websites, fill forms, take screenshots, scrape data, or test web applications. |
| allowed-tools | Bash(agent-browser:*) |
Two complementary approaches:
mcp__chrome-devtools__* against a running Chrome instance (when available)| Prefer Chrome DevTools MCP when | Prefer agent-browser CLI when |
|---|---|
| Live inspection of the user’s Chrome session / auth | Video recording (record start/stop) |
| Snapshots, screenshots, JS eval, console/network debug | Session/profile persistence (--session, --profile) |
| Form interaction + device emulation | Headless-only / cloud (Browserbase, Browserless, …) |
| Lighthouse and performance traces | PDF export, annotated screenshots, snapshot/screenshot diffs |
No universal best tool — pick by Chrome availability, recording needs, and environment constraints.
Preferred (autoConnect): open Chrome → chrome://inspect#remote-debugging → enable remote debugging. MCP connects with no extra config.
Manual: start Chrome with --remote-debugging-port=9222, then point the MCP server at --browser-url=http://127.0.0.1:9222.
# macOS
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --remote-debugging-port=9222 &
# Linux: google-chrome --remote-debugging-port=9222 &
# Windows: "C:\Program Files\Google\Chrome\Application\chrome.exe" --remote-debugging-port=9222
agent-browser open <url>
agent-browser open <url> --ignore-https-errors # self-signed certs
agent-browser open <url> --idle-timeout 5m
Isolated headless browser; also supports Brave auto-discovery alongside Chrome/Chromium.
navigate_page / new_page → wait_for as neededtake_snapshot (or screenshot) before actingclick / fill / fill_form / keys using selectors from the snapshotreferences/chrome-devtools-mcp.mdagent-browser open <url>agent-browser snapshot -i → use @eN refs from the treeclick, fill, …) with those refsreferences/agent-browser-cli.mdagent-browser --help)| Need | Open |
|---|---|
| Full MCP tool tables + server flags | references/chrome-devtools-mcp.md |
| Full agent-browser CLI catalog | references/agent-browser-cli.md |
| Worked examples (form, record, standalone) | references/workflows.md |
npx chrome-devtools-mcp@latest --autoConnect