Chromanche — self-hosted "Agent in Chrome"

Lets any MCP-capable coding agent drive your real, logged-in Chrome through a local MCP server + MV3 extension. Works with Claude Code, Codex, OpenCode, GitHub Copilot CLI, and other MCP clients. No browser data leaves your machine; the MCP server binds 127.0.0.1 only.

What you get

MCP tools exposed over stdio, relayed to the extension over a localhost WebSocket. Current set (v0.12.0, 30 tools):

• Tabs: tabs_list, tabs_create, tabs_close, tabs_activate

• Navigation & read: page_navigate, page_snapshot (uid-annotated a11y tree / text / dom, optional bounds; since:"last" returns only changed lines since the tab's previous snapshot — a token/speed saver on heavy pages), page_screenshot (returns MCP image content + viewport metadata for vision-driven flows)

• Interact: page_click, page_click_xy (vision-driven coordinate click, optional double-click), page_type, page_paste (clipboard write + Cmd/Ctrl+V — reliable bulk grid fill for Excel/Sheets), page_focus (loud focus reset for SPAs that grab focus), page_focus_state (inspect current focus), page_hover, page_press_key, page_scroll (js or real-wheel mode for virtualized grids), page_fill_form, page_select, page_upload_file, page_drag, page_handle_dialog. page_click/page_type auto-wait for the target to be actionable (visible/stable/enabled) — pass force: true to skip.

• Wait: page_wait (uid/selector/text/function/response/loadstate — the antidote to racing async SPAs; text waits for visible page text like "Payment complete" to appear/disappear; response mode is observational and doesn't claim the tab), page_wait_for_download (observe the user's real download completing — never redirects, never enumerates history)

• Batch: page_batch — run several tools sequentially in one MCP round-trip (click → type → screenshot, fill multi-step forms, write a whole grid row). Aborts on first error by default; pass stopOnError: false to collect per-step errors.

• Network / JS: page_fetch, page_eval_js, console_read, network_read, network_get_request (headers + body of one buffered request — drill into a failed Office365/Graph call; observational, never claims)

• Multi-profile: chromanche_list_profiles + optional profile field on every tool

• Session: session_release

The default page_snapshot mode returns a CDP accessibility tree with stable uids — each interactive element gets a [uid] you pass directly to click/type/hover. No CSS selector guessing. Set includeBounds=true to add bbox=x,y,w,h for accessible nodes when you need stronger visual positioning. All interaction tools support includeSnapshot=true to get an updated tree in the response, reducing round-trips.

Multi-profile (v0.6.0+): install the extension in multiple Chrome profiles — each generates a stable tag derived from chrome.runtime.id. Connected profiles appear in chromanche_list_profiles; pass profile: "<tag>" on any tool call to target a specific one. When exactly one profile is connected, the profile field is optional (auto-routes). When multiple are connected and the field is omitted, the tool returns a structured error with the list. Each Chrome profile can set a human-readable label in the extension popup.

Zero-config pairing (v0.5.0+): the MCP server and the extension both derive the same WebSocket port and auth token from sha256(timezone + platform + salt). No copy-paste, no port config. Set CHROMANCHE_TOKEN / CHROMANCHE_PORT on the server and paste matching values into the extension popup's advanced section if you need to override.

Note on concurrent MCP sessions: one session becomes the local WebSocket leader; concurrent sessions that hit the same derived port join it through the built-in proxy path. This lets multiple MCP clients or client sessions share the same connected extension without manual port changes.

Every interactive tool auto-claims its target tab: the tab is put into a distinct orange "Agent" tab group and gets an amber pulsing border + "Agent is using this tab" pill — so you always know when an MCP client is driving.

Related MCP server: Tabduct

Requirements

• Node 20+

• A Chromium-based browser (Chrome, Edge, Brave, Arc) — 116+

• Claude Code, Codex, OpenCode, GitHub Copilot CLI, or another MCP-capable client

Quickstart (users)

One command — downloads the latest release, registers the MCP server with supported MCP clients when their CLIs are present, and prints the Chrome steps:

curl -fsSL https://raw.githubusercontent.com/marcobazzani/Chromanche/main/scripts/install.sh | bash

Then load the extension as unpacked. No token paste — pairing happens automatically the first time your MCP client launches the server.

Windows: use WSL for the installer, or do it by hand — download the latest chromanche-extension-*.zip + chromanche-mcp-server-*.tgz from Releases, unpack to %USERPROFILE%\.chromanche\, then register the MCP server per the installer's printed instructions.

Uninstall: removes Chromanche and legacy BrowserUse MCP registrations and installed files.

curl -fsSL https://raw.githubusercontent.com/marcobazzani/Chromanche/main/scripts/uninstall.sh | bash

Known limitations

• The overlay cannot inject into chrome:// pages, the Chrome Web Store, or sites with particularly aggressive CSP — the tab group badge and an orange toolbar dot still appear, but the border does not.

• Developer-mode extension install only (no Chrome Web Store listing yet — manifest and privacy policy are ready).