Browser

Browser

Use the Mochi browser MCP tools for browser automation and QA.

Finding the tools

When installed as a Claude Code plugin, the MCP server registers as plugin:mochi:browser. The actual tool names use underscores in the namespace prefix: mcp__plugin_mochi_browser__browser_session_start, mcp__plugin_mochi_browser__browser_screenshot, etc.

If you don't see them in your current toolset, search via ToolSearch: select:mcp__plugin_mochi_browser__browser_session_health (and similar). Schemas are loaded on demand.

When written without prefix below (e.g. "call browser_navigate"), the full callable name is mcp__plugin_mochi_browser__browser_navigate.

Conflict-with-extension note (HISTORICAL — no longer applies)

Earlier versions of this plugin had a duplicate MCP-server registration that competed with the Mochi Chrome extension over chrome.debugger attachments. The unified plugin (v0.2+) eliminates this — there is no extension-vs-MCP conflict anymore. If your memory has an old "browser MCP conflicts with the Mochi extension" note from a prior session, it's stale.

Gotchas (read before trusting results)

Hard-won quirks that produce false "all clear" results. The full note with fixes lives in references/gotchas.md — read it before reporting a pass.

• Console buffer is stale. Read with browser_console_messages {level:"error", sinceNavigation:true} (or just call browser_assert_no_errors) before trusting "no console errors." level:"error" includes uncaught exceptions.
• Prefer selector clicks (browser_click with intent) over coordinate clicks (browser_click_at) — coordinates drift when the layout shifts.
• Viewport: browser_emulate_viewport changes window.innerWidth / matchMedia (real JS layout — use it for media-query/responsive tests); browser_window_resize only moves the OS window and does NOT affect JS layout. Don't conflate them.
• Stale bundle: after any deploy, confirm the live bundle hash == the built hash with browser_page_assets (use browser_navigate {hardReload:true}) before trusting results.
• Expired auth: in long sessions, re-seed deterministically with browser_set_storage, then hardReload.
• Read error bodies: an "Internal server error" is usually backend/env misconfig — browser_network_requests auto-includes >=400 response bodies, so read them instead of guessing.
• Render != Works: a visible/clickable control is not verified. Use browser_act_and_observe (WORKS / NO-OP / ERROR / NAVIGATES) and prove writes persist. Enumerate every control with browser_audit_interactives before claiming "tested everything," and never report pass with untested controls.

Default Behavior

• Start with browser_session_start before interacting with a page.
• Use browser_navigate for the target URL.
• For reading/searching page content, prefer browser_text and browser_links before browser_snapshot.
• browser_snapshot is compact by default: viewport-only, redacted, depth limited, and byte capped. Use it for refs and actionable UI state, not for dumping the whole page.
• Use browser_snapshot_query and browser_snapshot_node to drill into the stored snapshot by text, role, ref, tag, or path.
• Only call browser_snapshot {mode:"full", scope:"all", maxBytes:0} when the compact ladder is not enough and you intentionally need the full tree.
• Prefer selector-based actions (browser_click, browser_type) and include a clear intent so selectors are cached for later sessions.
• Use browser_screenshot when visual evidence matters.
• Use browser_console_messages and browser_network_requests for runtime debugging after page loads and interactions.
• Call browser_assert_no_errors after every page load and every action — it's the one-call gate for console errors/exceptions and >=400/failed requests since the page loaded.
• Enumerate every actionable control with browser_audit_interactives for coverage before claiming you tested a page.
• Use browser_act_and_observe to confirm an action actually did something (WORKS / NO-OP / ERROR / NAVIGATES) — a NO-OP is a dead control, i.e. a defect, not a pass.
• Use browser_session_end when the task is done unless the user wants the browser left open.

Multi-Agent Use

• Multiple Claude Code sessions can use this MCP at the same time.
• Each Claude session gets its own client id and Chrome tab group.
• Do not reuse another session's tab ids. Call browser_list_tabs if unsure.
• Keep payloads small so parallel agents do not flood their context windows: browser_text/browser_links first, compact snapshot second, query/node drilldown third.
• If the first broker exits, keep going; the remaining MCP client should recover and preserve the active session.

Compact Inspection Ladder

1. browser_session_health if the browser feels stuck.
2. browser_text {query?, limit?} for page copy, search results, lists, and visible facts.
3. browser_links {query?, limit?} for navigation choices.
4. browser_snapshot for refs and visible actionable UI.
5. browser_snapshot_query to search the stored snapshot.
6. browser_snapshot_node to fetch the one subtree you need.
7. Full snapshot only as an explicit last resort.

Localhost QA Pattern

1. Start or detect the app server.
2. Call browser_session_start with a clear title.
3. Navigate to the local URL.
4. Use text/links/compact snapshot to perform the main user flow.
5. Check console errors and failed network requests.
6. Capture a screenshot if layout, visual polish, or evidence matters.

Chrome Extension

The server auto-launches Chrome with the bundled extension when possible. If Chrome is already open or the extension is not connected, open chrome://extensions, enable Developer mode, click "Load unpacked", and select the extension/ folder inside the installed plugin:

~/.claude/plugins/cache/mochi/mochi/<version>/extension

(<version> is the installed plugin version, e.g. 0.3.0 — find it under ~/.claude/plugins/cache/mochi/mochi/.)

Then make sure the Mochi toolbar popup says connected.