An MCP server that lets AI assistants (Claude Code, Claude Desktop, Cursor…) read, screenshot and manage tabs in Firefox-based browsers — Floorp, LibreWolf, Waterfox, Zen, Mullvad, Firefox — via Floorp's automation API or Marionette.
gecko-mcp demonstrates solid tool design with 41 well-named, verb-prefixed tools (list_tabs, open_tab, navigate_tab, etc.). All tools have descriptions (avg ~120 chars, within 10-1024 baseline). Input schemas are present and typed for all tools examined. However, several tools lack parameter descriptions (e.g., real_type, real_key, real_clear have minimal guidance), and output schemas are not formally documented. Error handling is present but inconsistent, some tools note prerequisites (enable_os_input, enable_evaluate) while others do not. Tool composition is strong: tools chain well (list_tabs → open_tab → read_page), and most accept browserId for targeting. Security is reasonable: evaluate is locked by default, OS input requires explicit enable. Naming follows verb_noun convention consistently. Main gaps: output schema documentation, parameter description completeness, and error recovery guidance.
Click an element by CSS selector. Targets the active tab unless a browserId is given.
Close a tab by its browserId (from list_tabs).
Disable the evaluate tool for this session.
Disable real OS keyboard/mouse input for this session.
Double-click an element. Targets the active tab unless a browserId is given.
Enable the evaluate tool (runs arbitrary JavaScript in the page) for this session.
Enable real OS keyboard/mouse input for this session. Unlocks real_type, real_key, real_clear, move_cursor, and real_click.
Output schemas not formally documented. Tools return structured data (e.g., read_page returns Markdown/HTML/accessibility-tree, screenshot returns base64 PNG, find returns CSS selectors) but no explicit schema definitions are visible in source. LLMs cannot plan downstream tool calls without knowing return structure.
Real OS input tools (real_type, real_key, real_clear, real_click, move_cursor) have minimal parameter descriptions. real_type lacks guidance on delayMs semantics; real_key references 'src/os-input.ts' for key names instead of documenting them inline. Agents cannot determine valid key names without reading source.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Run arbitrary JavaScript in the page and return the result. LOCKED by default — requires enable_evaluate.
Fill a form by selector → value map. Targets the active tab unless a browserId is given.
Locate elements in a tab by visible text and/or tag. Returns compact, clickable CSS selectors — so the full page never reaches the model. Targets the active tab unless a browserId is given.
Return the active tab's title, URL and browserId. Note: with multiple browser windows open, 'active' is ambiguous — prefer the browserId returned by open_tab, or pick from list_tabs.
Extract article content from a page using Mozilla's Readability algorithm. Returns title, byline, and Markdown. Targets the active tab unless a browserId is given.
Get an attribute value from an element. Targets the active tab unless a browserId is given.
Get all cookies for the current page. Targets the active tab unless a browserId is given.
Get the value of a form field. Targets the active tab unless a browserId is given.
Hover over an element. Targets the active tab unless a browserId is given.
Launch a Gecko browser (Firefox, LibreWolf, Waterfox, Zen, Mullvad, Floorp) with Marionette enabled.
Launch Floorp if it isn't already running (Windows only). Ensures the automation API is reachable.
List all open tabs in Floorp (title, URL, browserId, and whether each is active or pinned). Use the browserId to target other tools.
Floorp-only: list all workspaces.
Move the OS mouse cursor to a screen position. Requires enable_os_input.
Navigate a tab to a new URL. Targets the active tab unless a browserId is given.
Open a URL in a new Floorp tab.
Press a keyboard key (or key combination). Targets the active tab unless a browserId is given.
Read a tab's content. Returns clean Markdown by default; can also return raw HTML or the accessibility tree. Output is capped (default 25 KB) to protect the context — to LOCATE a specific element use `find` (cheaper) instead. Targets the active tab unless a browserId is given.
Clear a field using the real OS keyboard (Ctrl+A, Delete). Requires enable_os_input.
Click using the real OS mouse. Requires enable_os_input.
Press a key using the real OS keyboard (not the browser's input simulation). Requires enable_os_input.
Type text using the real OS keyboard (not the browser's input simulation). Requires enable_os_input. Useful for fields that don't respond to synthetic input.
Right-click an element. Targets the active tab unless a browserId is given.
Screenshot a tab (viewport only, not full page). Returns a base64-encoded PNG. Targets the active tab unless a browserId is given.
Select an option in a <select> element. Targets the active tab unless a browserId is given.
Set a checkbox or radio button to checked or unchecked. Targets the active tab unless a browserId is given.
Floorp-only: take a snapshot of the page with a fingerprint map for reliable element targeting. Returns Markdown with embedded fingerprints. Targets the active tab unless a browserId is given.
Submit a form. Targets the active tab unless a browserId is given.
Floorp-only: switch to a workspace by ID.
Type text into a focused element (or into the element matching the selector). Targets the active tab unless a browserId is given.
Upload a file to a file input. Targets the active tab unless a browserId is given.
Wait for an element to reach a state (attached, visible, hidden, or detached). Targets the active tab unless a browserId is given.
Wait for network activity to settle (no pending requests). Targets the active tab unless a browserId is given.
Floorp-only: get the browser window's position and size.
Floorp-only tools (snapshot, list_workspaces, switch_workspace, window_bounds) lack clear degradation guidance. snapshot description notes 'Floorp-only' but does not explain what happens on non-Floorp browsers. Test shows graceful error, but this should be documented in tool descriptions to set LLM expectations.
Prerequisite dependencies not consistently documented. enable_os_input and enable_evaluate unlock other tools, but this is only mentioned in tool descriptions, not in a centralized capability map. Agents may attempt real_type before enable_os_input without clear guidance.
Error handling lacks actionable recovery guidance. Tools return errors (e.g., 'LOCKED' for evaluate before enable_evaluate) but do not consistently suggest next steps. Pattern: 'evaluate is locked. Call enable_evaluate first.' would guide LLM recovery.