gecko-mcp

gecko-mcp

[CI] [License: MIT]

An MCP (Model Context Protocol) server that lets AI assistants — Claude Code, Claude Desktop, Cursor, and any MCP client — read pages, take screenshots and manage tabs in Floorp and other Firefox-based browsers (LibreWolf, Waterfox, Zen, Mullvad, Firefox…), using your real, logged-in session.

Think "Claude in Chrome", but for the whole Firefox/Gecko family.

[Claude driving Floorp via gecko-mcp — opening the GitHub repo and the npm page from a natural-language prompt]

Get started in ~30 seconds:

npx gecko-mcp setup

Registers gecko-mcp with Claude Code, Claude Desktop, Cursor, VS Code (Copilot), Windsurf, Zed, Codex and more — per-project or global. See Setup for manual config and Requirements for the one-time browser step.

Cautious about installing this? Good — you should be. It's small (2 deps, all in src/), the OS keyboard/mouse is locked by default (browser-only until you opt in), releases ship with npm provenance (verifiable against this source), and the full threat model is in SECURITY.md. Read it before you run npx gecko-mcp.

How it works

gecko-mcp talks to the browser through one of two backends, picked automatically:

• Floorp ships a built-in automation HTTP API. Set floorp.mcp.enabled = true in about:config and gecko-mcp uses the fast http://127.0.0.1:58261 API — no extension, richest feature set.
• Any other Gecko browser — launch it with Marionette (the automation engine built into every Firefox fork) and gecko-mcp drives your live session over it. Same tools, same real session.

  Claude Code / Desktop / Cursor            │  MCP (stdio)            ▼      gecko-mcp ──► Floorp :58261 (built-in API)        ─┐   (this project) ──► Marionette :2828 (any Gecko fork)  ─┴─► your real tabs

Requirements

• A Firefox-based browser installed and running, with automation enabled:
  • Floorp: set floorp.mcp.enabled to true in about:config, restart Floorp.
  • Other forks (LibreWolf / Waterfox / Zen / Mullvad / Firefox): launch the browser with -marionette (see Browser support).
• Node.js ≥ 18.

Setup

Quick start — the setup wizard

npx gecko-mcp setup

An interactive wizard registers gecko-mcp with the AI coding tool(s) of your choice — Claude Code, Cursor, Windsurf, VS Code (Copilot), Gemini CLI, Codex, Zed, Cline (and a copy-paste snippet for Kimi Code, Antigravity, or any other MCP client) — and lets you install it for the current project or globally (all repos). It merges into existing config (and backs it up first).

Non-interactive / scriptable:

npx gecko-mcp setup --list                          # show supported toolsnpx gecko-mcp setup --tool claude-code,cursor --scope globalnpx gecko-mcp setup --tool codex --scope global --print   # dry run

Manual

Any MCP client works with this server block (no clone/build needed — npx fetches it):

{  "mcpServers": {    "gecko": {      "command": "npx",      "args": ["-y", "gecko-mcp"]    }  }}

Or with Claude Code's CLI: claude mcp add gecko -s user -- npx -y gecko-mcp.

One-time Floorp step: set floorp.mcp.enabled = true in about:config and restart Floorp so its automation API is available.

Browser support

gecko-mcp picks its backend automatically: if Floorp's :58261 API is reachable it uses that; otherwise it connects to Marionette, the automation engine built into every Gecko browser. To use a non-Floorp browser, launch it once with Marionette enabled:

Marionette listens on TCP 2828 by default. To use another port, set the marionette.port pref in the profile (e.g. via user.js) and start gecko-mcp with a matching MARIONETTE_PORT. Force a backend with GECKO_MCP_BACKEND=marionette.

Note: Marionette must be enabled at launch to attach to your live session. On the Marionette backend, Floorp-only extras (snapshot fingerprints, list_workspaces/switch_workspace, accessibility tree) return a clear "not supported" message — use find / read_page instead. Everything else (tabs, navigation, click, type, forms, screenshots, cookies, real OS input…) works.

Tools

Tabs & reading

Interaction

Most tools target the active tab by default; pass a browserId (from list_tabs) to target a specific tab.

OS keyboard & mouse — locked by default 🔒

The tools below can affect things outside the browser, so they are disabled until you turn them on. With nothing set, gecko-mcp does browser automation only. Unlock them per-session by just asking ("enable OS input", which calls the enable_os_input tool), or persistently with GECKO_MCP_ENABLE_OS_INPUT=1. Lock again with disable_os_input. While locked, these tools refuse with a clear message.

The evaluate tool (run arbitrary page JavaScript) is locked the same way — unlock with enable_evaluate or GECKO_MCP_ENABLE_EVALUATE=1.

Real OS keyboard (Windows) — for React/rich editors and bot-guarded submits that ignore synthetic input:

These produce input a page can't distinguish from a human's, so they drive React/Slate editors and submit composers that synthetic clicks/typing can't. Workflow: click the field to focus it → real_clear / real_type / real_key "Enter".

Safety guard: OS keystrokes go to the foreground window, so before sending anything these tools bring Floorp to the foreground and verify it — if Floorp isn't running or can't be focused, they abort without typing a single key, so input can never leak into another app.

Real OS mouse (Windows) — genuine isTrusted clicks at screen coordinates:

Double guard: the click is sent only when Floorp is verified foreground and the point lies inside Floorp's window rect — a stray coordinate is refused, so a click can never land in another app/window. Coordinates are screen pixels (note display scaling/DPI when mapping from a screenshot).

More interaction & queries

Security

Understand the threat model before enabling this. Two risks dominate:

1. Floorp's automation API has no authentication by default. While floorp.mcp.enabled is on, any local process can drive your logged-in browser via 127.0.0.1:58261 — not just this server. There is also no Origin check, so hostile web pages may attempt CSRF/DNS-rebinding tricks against it. Mitigations:
   • Turn floorp.mcp.enabled off when you're not using automation.
   • Set the GECKO_MCP_TOKEN environment variable — this server then sends it as a Bearer token on every request (effective on Floorp builds that enforce a token; harmless otherwise).
2. Prompt injection ("lethal trifecta"). The assistant reads untrusted page content and can act on your authenticated sessions (click, type, submit, navigate, real OS input). A malicious page could try to instruct the assistant to act against you. Treat everything read from a page as untrusted; don't run automation unattended on sites you don't trust.

Hardening built into this server:

• OS keyboard/mouse is locked by default (least privilege): the only tools that can act outside the browser refuse to run until you explicitly unlock them (enable_os_input tool, or GECKO_MCP_ENABLE_OS_INPUT=1). By default gecko-mcp can only automate the browser, never your wider machine.
• Real OS input is double-guarded: keys/clicks are sent only after verifying Floorp is the foreground window, and mouse clicks must land inside Floorp's window rectangle — otherwise it aborts without sending anything. PowerShell payloads are passed base64-encoded via process-private environment variables (no shell interpolation, no temp script files on disk).
• URL scheme + host allowlist: open_tab/navigate_tab accept only http(s) (and about:blank) by default, and refuse loopback/private hosts (127.0.0.1, localhost, 10/8, 172.16/12, 192.168/16, 169.254/16, IPv6 ULA/link-local). This stops a prompt-injected agent from pivoting the browser onto Floorp's own API or your LAN and reading the response back. Lift with GECKO_MCP_ALLOW_PRIVILEGED_URLS=1. Optionally pin navigation to a domain allowlist with GECKO_MCP_ALLOW_DOMAINS.
• Cookie values are redacted by default in get_cookies; raw values require an explicit includeValues: true.
• get_value can read secrets: browsers let same-origin JS read password fields, so this tool can return a typed password. It's flagged SENSITIVE — use it only on fields the user asked about, never to harvest credentials.
• Upload allowlist: set GECKO_MCP_ALLOW_UPLOAD_DIRS (;-separated directories) to confine upload_file. Paths are canonicalised with realpath (symlinks resolved) and checked so .., a symlink, a same-prefix sibling directory, or a UNC path can't escape the allowed folders.
• find skips hidden elements (inline display:none/visibility:hidden, hidden, type=hidden, aria-hidden) so a page can't lure the agent into clicking an invisible button via text search.
• Input bounds: numeric/text tool parameters are range- and length-capped (coordinates, timeouts, maxChars, find limit, typed text, form fields) to prevent resource-exhaustion / crash inputs.
• Truncated API errors & validated port: Floorp error bodies are truncated before reaching the model; GECKO_MCP_PORT is validated as 1–65535.
• Tool annotations for human-in-the-loop: every tool carries MCP hints (readOnlyHint/destructiveHint/…) so your client can auto-run read-only tools and confirm destructive ones (close_tab, navigate_tab, submit_form, upload_file). A server can't show prompts itself — approval is the client's job — so this is how gecko-mcp tells the client what's safe vs consequential.
• No evaluate tool: arbitrary page-JS execution is deliberately not exposed.

What is not defended (inherent / Floorp-side): a malicious local process can still read or impersonate the unauthenticated loopback API (plaintext, no TLS), and prompt injection from a page you choose to automate can still drive legitimate actions on that page. Disable floorp.mcp.enabled when idle and don't automate untrusted sites unattended.

The legacy FLOORP_MCP_* variable names still work as fallbacks (from before the rename), so existing configs keep working — prefer GECKO_MCP_* going forward.

Performance

• HTTP tool calls are cheap — a full attach → act → detach round-trip against Floorp's local API is ~5–6 ms. find searches the page server-side and returns ~1 KB of ready-to-use selectors instead of dumping the whole HTML, and read_page is capped (default 25 KB) so a page read can't flood the context.
• Real OS input uses a persistent PowerShell host. Spawning powershell.exe (~700 ms) and compiling the P/Invoke helper (~600 ms) used to happen on every real_*/move_cursor/window_bounds call (~1.9 s each). Now one host is started lazily, compiles once, and runs a read-eval loop — so the first call pays ~1.6 s but every call after is ~350 ms for a guarded key/click (~5× faster) and a few ms for a window-bounds query. The foreground/bounds safety guards still run on every command; the host is recycled if it hangs or dies.

Notes & limitations

Learned from driving real apps (incl. Google Flow):

• Rich editors: type_text handles plain inputs and contenteditable editors (Slate, ProseMirror, Lexical) — it falls back to dispatching a real text-input event when an element has no .value. Reliably clearing such editors isn't solved yet (no select-all/evaluate).
• Submitting React composers: many chat/prompt composers submit on a real Enter keydown, not on a synthetic click of the send button. Prefer press_key "Enter" over click for those.
• Trusted events: you cannot forge isTrusted=true from page JavaScript — it is a browser security invariant. Floorp injects input at a privileged layer, so ordinary clicks/keys behave like real ones; but flows guarded by reCAPTCHA or strict bot-detection may still refuse automated submission.
• evaluate: the page-JS eval endpoint returns HTTP 404 on some Floorp builds, so it is not exposed as a tool here.
• Multiple windows: when more than one window is open, the "active tab" is ambiguous (each window has its own active tab). Prefer the browserId returned by open_tab, or one from list_tabs, and pass it explicitly to every tool.

Roadmap

• Tab management, page reading, screenshots
• Interaction tools: click, type, fill forms, key presses, read field values
• Real OS keyboard (Windows): real_type / real_key / real_clear, with a foreground safety guard — drives React/Slate editors & bot-guarded submits
• snapshot (fingerprint refs + selector map) + click by ref + auto-scroll-into-view
• launch_floorp — start Floorp if not running (Windows)
• Extra tools: hover, double/right-click, select_option, set_checked, submit, upload_file, get_attribute, get_article, get_cookies, wait_for_network_idle, workspaces
• Real OS mouse (Windows): window_bounds / move_cursor / real_click, with a foreground + in-window-bounds double guard
• Marionette backend — all Firefox-based browsers (LibreWolf, Waterfox, Zen, Mullvad, Firefox…), auto-selected when Floorp's API isn't present
• macOS / Linux native-input backends
• JS evaluate (available in newer Floorp builds; older ones return HTTP 404)
• Optional bearer-token auth
• launch helper for non-Floorp browsers (start them with -marionette)

Acknowledgements

Built against the automation API exposed by Floorp. The official Floorp-Projects/floorp-mcp-server was a useful reference for mapping the endpoint surface. This is an independent, clean-room MIT-licensed implementation.

License

MIT © Frumane