Opencode-chrome MCPサーバー

opencode-chrome

opencode-chrome lets opencode use your real Chrome: the one already logged into Gmail, your dashboards, your staging site. Ask in the terminal, watch it happen in your browser. An unofficial, community-built bridge in the spirit of Claude in Chrome / Codex in Chrome.

Unofficial. Not affiliated with or endorsed by the opencode project. BYOK: your model keys stay in your opencode config. This bridge collects nothing and talks only to localhost.

What it looks like

you ▸ open github.com/notifications and tell me which repos have unread activity
you ▸ click the extension icon to attach that tab
opencode ▸ snapshot  (reads the list as a text tree)
           → acme/api (2), acme/web (1), infra/deploy (5)

you ▸ go to the staging checkout and screenshot it
opencode ▸ navigate    staging.example.com/checkout
           screenshot  → returns the PNG

It works because it drives your browser session, so anything you can see while logged in, the agent can see too.

How it works

A tiny local bridge (npx opencode-chrome, an MCP server plus a WebSocket on 127.0.0.1:19223) and an MV3 extension that connects to it and drives only the tabs you attach from the toolbar, using the Chrome DevTools Protocol. An attachment is limited to the tab's current origin. No backend, no accounts, no analytics.

Requirements

• Node 20 or newer
• Google Chrome (or Chromium) 116 or newer
• opencode (or any MCP client that launches a local stdio server)

Install

1. Extension: download the repo (or the latest release), open chrome://extensions, enable Developer mode, "Load unpacked", select the extension/ folder. The toolbar badge shows off until the token is configured and the bridge is running.

2. Bridge: add to your opencode config (opencode.json):

{
  "mcp": {
    "chrome": { "type": "local", "command": ["npx", "-y", "opencode-chrome"] }
  }
}

3. Pair the token: the bridge generates a token at ~/.config/opencode-chrome/token (or prints it to stderr and copies it to your clipboard at startup; set OPENCODE_CHROME_TOKEN to use your own). Paste it into the extension's options page (right-click the toolbar icon → Options); it shows a live connection status and reconnects the moment you save. Connections without this token are refused, so other local processes can't drive your browser.

4. Attach a tab: restart opencode, open the page you want to share, and click the extension icon. Its badge shows on for attached tabs. Cross-origin navigation detaches the tab by default; for legitimate flows (SSO redirects) list the extra origins in ~/.config/opencode-chrome/policy.json:

{ "origin_allowlist": { "https://app.example.com": ["https://sso.example.com"] } }

The bridge pushes the allowlist to the extension on connect (env override: OPENCODE_CHROME_POLICY). Entries apply per attached origin and are not symmetric; with no file the policy is strict exact-origin.

The same command works from another local MCP client. How many sessions run at once is set in the extension options, under Sesiones. It stays at one unless you change it.

Site adapters

Declarative per-origin recipes live in ~/.config/opencode-chrome/adapters/ (env override: OPENCODE_CHROME_ADAPTERS). list_recipes enumerates them; run_recipe(name, params?) executes one on an attached tab. The adapter's origin must be inside the attachment's reach: exact match or allowlisted, same rule as SSO redirects and OOPIF frames.

{
  "name": "example",
  "origin": "https://app.example.com",
  "params": ["q"],
  "steps": [
    { "navigate": "https://app.example.com/search?q={{q}}" },
    { "wait_for": "Results" },
    { "eval": "Array.from(document.querySelectorAll('.row')).map(r => ({title: r.innerText}))" },
    { "columns": ["title"] }
  ]
}

Steps: navigate, wait_for (+ optional timeout), init_script, eval, columns. init_script registers JS on new documents for the duration of the recipe only (auth/session setup), then it is removed. {{key}} placeholders come from params; inside eval each placeholder is substituted as a JSON-encoded literal (write search({{q}}), no quotes) so a param value can never break out of the script's context. The eval JS lives only in files you install: there is no free-form eval tool. See examples/adapters/hackernews.json.

Chat (side panel)

An optional chat client lives in the Chrome side panel: you type, a model drives your attached tabs through the same tool surface. It works with no MCP client running. The key lives in the bridge, never in the extension:

// ~/.config/opencode-chrome/chat.json  (chmod 600)
{ "provider": "anthropic", "api_key": "sk-...", "model": "claude-sonnet-4-5" }

provider is anthropic or openai (any OpenAI-compatible endpoint: set base_url for OpenRouter, Ollama, Groq, etc.). Optional: effort (off|low|medium|high → reasoning_effort / Anthropic thinking budget), max_turns, max_minutes, max_tokens. Env override: OPENCODE_CHROME_CHAT.

You can also configure it from the panel itself: the gear opens a form with provider, key, optional base URL, a model picker (loaded from the provider's /models, free models first), and an effort selector. The key travels only over the authenticated socket and is written to chat.json — it never lands in extension storage. The form also shows a data-retention note for the endpoint (no training / conditional / free-tier trains / local / unknown), worth a look before pasting passenger or payment data.

Run opencode-chrome --serve for a chat-only bridge (no MCP stdio); the normal MCP-spawned bridge serves chat too. Attaching a tab opens the panel. Reads run directly; mutations pause on an approve/deny prompt in the panel. Page content reaches the model only through tool results, wrapped as untrusted data — same rules as an MCP client, the key and the loop just move into your bridge.

Tools

Tool	What it does
browser_status	bridge connection and attached tabs
list_tabs	attached tabs with id, title, url
new_tab(url?, background?)	open a tab (active unless background:true)
close_tab(id) / activate_tab(id)	tab management
navigate(url?, action?, ignore_cache?, handle_before_unload?, tabId?)	go to a URL or action back/forward/reload; waits for load; handle_before_unload auto-answers that dialog
snapshot(tabId?)	accessibility-style text tree with [ref] per interactive element; selector, interactive_only, in_viewport_only, max_chars, frame scope it; sensitive fields are redacted
find(text?, role?, selector?, frame?)	matching [ref] lines only
read_text(selector?, max?, scroll?, frame?)	element innerText, hydrates lazy sections
list_frames	frame tree with OOPIF session ids and allowlist status
click(ref) / hover(ref)	trusted CDP pointer actions
drag(from, to)	trusted drag between refs of the same frame
type(ref, text)	focus + type; trailing \n = Enter
fill(ref, value)	set value with native setter + input/change events
select(ref, value)	select option by label or value
press_key(key)	key or combo on the focused element: Enter, Tab, Escape, arrows, Control+A...
form_schema(tabId?)	form fields with ref, kind, label, autocomplete, required, sensitive; honeypots counted in hidden_count
list_profile_keys(profile) / apply_mapping({ref: key}, profile)	fill a whole form from chrome.storage.local.profiles.<profile>; values never cross the wire, missing keys land in unmapped_keys
scroll(ref? or dx/dy)	scroll element into view or page by deltas
upload(ref, paths)	set files on a file input via CDP
list_network / get_network_body(id)	captured requests and response bodies; filter, resource_types, offset/limit scope it, include_headers adds redacted headers, format:'har' or output_path exports HAR 1.2
list_storage_keys	localStorage/sessionStorage key names only (values never leave the tab)
list_console_messages	console calls, uncaught exceptions and browser log entries; types/filter scope it
list_dialogs / handle_dialog	pending/recent JS dialogs; per-tab accept/dismiss policy
screenshot(tabId?, annotate?)	PNG (base64); annotate overlays [N] badges on snapshot refs
resize_page(width, height, clear?, tabId?)	viewport override via CDP; persists while attached
emulate({network,cpu,geolocation,color_scheme,reduced_motion,user_agent,locale}, clear?)	environment overrides per tab; clear:true resets
`wait_for(text	text[], timeout?)`
wait_download(timeout_ms?)	wait for a download; returns saved path, bytes, mime and sha256
list_recipes / run_recipe(name, params?)	site adapters from ~/.config/opencode-chrome/adapters/

Notes

• Mutating tools (navigate, click, hover, drag, type, fill, select, scroll, upload, apply_mapping, press_key) accept include_snapshot:true to return a fresh snapshot in the same response; its refs replace the previous ones.
• Payload tools (snapshot, find, read_text, screenshot, get_network_body) accept output_path to write the payload to a file on the bridge machine and return {path, bytes} instead of inline content.
• While the agent acts, Chrome shows the "being debugged" banner. That is expected with CDP; the extension auto-detaches after 30s idle.
• Env vars for the bridge: OPENCODE_CHROME_PORT (default 19223; the extension follows it via the port field on the options page), OPENCODE_CHROME_TIMEOUT_MS (default 30000), OPENCODE_CHROME_SESSION (optional name for that terminal, shown only when more than one session is connected).
• Security: the WebSocket binds to 127.0.0.1 only, rejects non-extension origins, and requires the shared bridge/extension token. The only data stored persistently is that token, the optional port override, and the session cap (in chrome.storage.local, and in ~/.config/opencode-chrome/config.json once the extension has connected). Attached tab origins live only in chrome.storage.session. Page content never leaves your machine through this bridge; it goes only to your model provider, exactly like any opencode prompt.

Troubleshooting

• Badge stays off: the extension is not connected. Check that the bridge is running (opencode launches npx opencode-chrome; its stderr prints the token at startup) and that the token in the options page matches. The options page shows the connection status live.
• A tool returns "Chrome extension not connected": same causes as above; the bridge is up but no extension has paired.
• cannot listen, cannot attach, or no free port: something that is not this bridge holds the port, or every port in the session range is taken. With Sessions left at one, a second copy attaches to the bridge already running. If stderr says the bridge is an older process, restart that process. If it says there is no free port, close one session or raise the cap under Sesiones in the extension options.
• Connection drops after sleep or a restart: the extension retries every three seconds while awake and uses a Chrome alarm to recover after suspension. A connection that receives no messages for 45 seconds is replaced. If a tool was interrupted, check the page before retrying; actions are not replayed.
• Testing a local fix: npx opencode-chrome runs the published package. Change the MCP command to ["node", "/absolute/path/to/opencode-chrome/src/index.js"], reload the extension in chrome://extensions, then restart opencode.
• debugger attach fails: a tab allows only one debugger client. Close DevTools on that tab (or detach other debuggers) and retry.

Development

npm install
npm test        # bridge tests (node:test, no Chrome needed)
npm run pack    # builds dist/opencode-chrome-<version>.zip for CWS upload

# real-Chrome smoke: launches Chrome, loads the extension, drives a fixture
# page end to end. Needs a Chrome binary; kept out of `npm test` for that reason.
node scripts/e2e-smoke.mjs   # SMOKE_HEADLESS=1 for --headless=new

# Playwright e2e suite (e2e/): same real path with broader coverage: tabs,
# interactions, downloads, dialogs, network, guards. Also needs Chrome
# (OPENCODE_CHROME_BIN overrides the lookup).
npm run test:e2e

Icons live in extension/icons/, sized from the 1024px masters in assets/: sips -z <size> <size> assets/logo-1024-transparent.png --out extension/icons/icon<size>.png.

Contributions welcome: see CONTRIBUTING.md. Working on the code with an AI agent? Start with AGENTS.md.

MIT license. See LICENSE.

---

opencode-chrome (español)

opencode-chrome deja que opencode use tu Chrome real: el que ya tiene la sesión iniciada en Gmail, tus dashboards, tu staging. Pídelo en la terminal y pasa en tu navegador. Puente comunitario no oficial, al estilo de Claude in Chrome / Codex in Chrome.

No oficial. Sin afiliación con el proyecto opencode. BYOK: tus keys viven en tu config de opencode. Este puente no recolecta nada y solo se comunica con localhost.

Cómo se ve

tú ▸ abrí github.com/notifications y decime qué repos tienen actividad sin leer
opencode ▸ navigate  github.com/notifications
           snapshot  (lee la lista como árbol de texto)
           → acme/api (2), acme/web (1), infra/deploy (5)

tú ▸ andá al checkout de staging y sacale un screenshot
opencode ▸ navigate    staging.example.com/checkout
           screenshot  → devuelve el PNG

Funciona porque maneja tu sesión del navegador: lo que ves con la sesión iniciada, el agente también lo ve.

Cómo funciona

Un puente local chico (npx opencode-chrome, servidor MCP más un WebSocket en 127.0.0.1:19223) y una extensión MV3 que se conecta a él y controla solo las pestañas que adjuntas desde el ícono, mediante Chrome DevTools Protocol. Sin backend, sin cuentas, sin analítica.

Requisitos

• Node 20 o superior
• Google Chrome (o Chromium) 116 o superior
• opencode (o cualquier cliente MCP que lance un servidor stdio local)

Instalación

1. Extensión: chrome://extensions → modo desarrollador → "Cargar descomprimida" → carpeta extension/ del repo. El badge muestra off hasta configurar el token y correr el puente.
2. Puente: en tu opencode.json:

   { "mcp": { "chrome": { "type": "local", "command": ["npx", "-y", "opencode-chrome"] } } }
   

3. Token: el puente genera uno en ~/.config/opencode-chrome/token (o lo imprime por stderr y lo copia a tu portapapeles al arrancar; OPENCODE_CHROME_TOKEN para usar el tuyo). Pegalo en la página de opciones de la extensión (click derecho en el ícono → Opciones); muestra el estado de conexión y reconecta apenas guardás. Sin ese token, las conexiones se rechazan.
4. Reinicia opencode, abre la página, haz clic en el ícono de la extensión para adjuntar esa pestaña y luego pídele cosas.

El mismo comando sirve para otro cliente MCP local. Cuántas sesiones corren a la vez se define en las opciones de la extensión, en Sesiones. Queda en una hasta que lo cambies.

Chat opcional en el panel lateral: configurá ~/.config/opencode-chrome/chat.json (provider anthropic u openai, api_key, model, effort opcional) o abrí el engranaje del panel y llená el form: provider, key, URL base opcional, selector de modelo (ordenado gratis primero) y effort. La key viaja solo por el socket autenticado hasta chat.json; nunca queda en la extensión. El form también muestra la política de retención del endpoint (no entrena / condicional / gratis entrena / local / desconocida). Corré opencode-chrome --serve o el puente MCP de siempre. Al adjuntar un tab se abre el panel; las lecturas corren directo y cada mutación pide aprobación ahí mismo.

Herramientas, notas de seguridad, solución de problemas y desarrollo: ver sección en inglés arriba (mismo contenido). Para contribuir: CONTRIBUTING.md; si trabajas con un agente de IA: AGENTS.md.