MeteorDroid

OpenCode  --MCP/stdio-->  MeteorDroid (this repo)  --Playwright-->  Chromium
                                |                              |
                                |                              + Visible safety overlay + STOP
                                + Auto-pause + human handoff on sensitive steps

Quick Install

MeteorDroid is a standard MCP stdio server, so it works with OpenCode and any other MCP host that can spawn a local command over stdio.

Install from npm:

npm i -g meteordroid

Entry point for MCP hosts (global install):

{
  "command": ["meteordroid"],
  "environment": {
    "MINI_COMET_MODE": "persistent",
    "MINI_COMET_PROFILE_DIR": "/Users/you/.minicomet/profile"
  }
}

Or run via npx (no global install):

{
  "command": ["npx", "-y", "meteordroid"],
  "environment": {
    "MINI_COMET_MODE": "persistent",
    "MINI_COMET_PROFILE_DIR": "/Users/you/.minicomet/profile"
  }
}

Why MeteorDroid?

• A real, visible browser (not a silent headless bot)
• Auto-pause on sensitive actions (sign in/up, submit/enter, purchases, destructive flows)
• Three session modes: ephemeral, persistent, or attach to your real Chrome via cdp
• Multi-tab read with a single write target to reduce "oops wrong tab" failures
• File uploads via browser_upload_file

What's new in v0.4

• browser_upload_file — upload one or more local files to a file input. Targets the <input type=file> directly OR a nearby drop-zone / button (the tool walks up to find the closest hidden file input). Handles the common SPA pattern where the visible button is decorative and the real input is display:none.
• Smart locator with fallback chain. Plain-English targets are now tried in this order: explicit selector syntax → role+name (button/link/textbox/ checkbox/tab/menuitem) → label → placeholder → alt text → title → text. The winning strategy is returned in every click/type/wait_for result so you can see (and one day record) which lookup style worked.
• browser_wait_for — wait until an element matching selector_or_text reaches a state (visible / hidden / attached / detached). Cuts the guesswork out of SPA transitions.
• browser_dom_summary — compact JSON of visible interactive elements ({tag, role, name, type?, id?, placeholder?}). Much cheaper than get_page_text when you just need to know "what can I click or fill?".
• Overlay polish. Pill moved to bottom-right (out of common notification badge zones), bigger STOP button, amber border flash on every write, and the last action label now lingers for ~1.8s after returning to idle so you can read what just happened.

What was in v0.3

• Visible safety overlay with pulsing pill + STOP button.
• Animated fake cursor (blue→orange while typing, red while clicking).
• Keyboard-first form filling: browser_focus_next, browser_focus_prev, browser_type_into_focused.
• browser_evaluate for one-off DOM probing.

What was in v0.2

• Three session modes — ephemeral, persistent (logins survive), or attach to your real Chrome via CDP.
• Multi-tab read, single-tab write.
• Auto-pause + manual resume handoff.

Tools

Navigation & I/O

Multi-tab

Handoff

All write tools are gated on the pause flag and return a structured {paused, reason, hint} observation when blocked. Read tools work while paused so the agent can verify the new state on resume.

Quick Start

npm install
npm run typecheck

Requires Node >= 18.17.

Add to OpenCode

Use the included opencode.jsonc as a starting point. Example:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "browser-agent": {
      "type": "local",
      "command": ["npx", "tsx", "/abs/path/to/src/server.ts"],
      "environment": {
        "MINI_COMET_MODE": "persistent",
        "MINI_COMET_PROFILE_DIR": "/Users/you/.minicomet/profile"
      },
      "enabled": true
    }
  }
}

Configuration — three session modes

Set MINI_COMET_MODE in your OpenCode MCP environment block.

Mode 1: ephemeral (default — like v0.1)

Fresh isolated Chromium. No cookies survive between runs. Best for safe, self-contained tasks (scraping public sites, smoke tests).

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "browser-agent": {
      "type": "local",
      "command": ["npx", "tsx", "/abs/path/opencode-browser-agent/src/server.ts"],
      "environment": { "MINI_COMET_MODE": "ephemeral" },
      "enabled": true
    }
  }
}

Mode 2: persistent — recommended default

Playwright stores cookies/localStorage in a directory. Log in once manually, the agent uses your session forever after. Different profile dir per use case recommended (one for dev tools, one for social, etc.).

{
  "mcp": {
    "browser-agent": {
      "type": "local",
      "command": ["npx", "tsx", "/abs/path/opencode-browser-agent/src/server.ts"],
      "environment": {
        "MINI_COMET_MODE": "persistent",
        "MINI_COMET_PROFILE_DIR": "/Users/you/.minicomet/profile-social",
        "MINI_COMET_HEADLESS": "0"
      },
      "enabled": true
    }
  }
}

The first time you launch, you'll get a clean Chromium. Log in to whichever sites you want; the agent will reuse those sessions on every subsequent run.

Mode 3: cdp — attach to your real Chrome

Connect MeteorDroid to the Chrome you already use, with all your extensions and logins. The agent creates a new tab marked [mc] and only writes to that tab, but can read any of your other open tabs.

Step 1. Launch your Chrome with debugging enabled. On macOS:

/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
  --remote-debugging-port=9222 \
  --user-data-dir=$HOME/.minicomet/chrome-cdp

Alternatively, MeteorDroid ships a small helper that can prompt you to pick an existing Chrome profile directory and launch Chrome in the background:

meteordroid cdp

(This uses your existing Chrome --user-data-dir and selected --profile-directory, so you must fully quit Chrome first.)

(Use a separate --user-data-dir so this doesn't conflict with your normal Chrome. Or omit it to use your default profile — at your own risk.)

Step 2. Configure OpenCode:

{
  "mcp": {
    "browser-agent": {
      "type": "local",
      "command": ["npx", "tsx", "/abs/path/opencode-browser-agent/src/server.ts"],
      "environment": {
        "MINI_COMET_MODE": "cdp",
        "MINI_COMET_CDP_URL": "http://localhost:9222"
      },
      "enabled": true
    }
  }
}

When OpenCode disconnects, the agent's [mc] tabs close but your Chrome keeps running.

All env vars

The handoff protocol

When the agent calls browser_click("Sign up"):

1. Server matches the safety pattern /\bsign\s*(in|up)\b/.
2. Server flips paused=true with reason and returns:

   {"paused": true, "reason": "Click \"Sign up\" matches sensitive pattern...",
    "hint": "Tell the user to complete this step in the visible browser, then ask them to say 'resume'."}
   

3. The LLM in OpenCode tells you what's needed, e.g.:

   🙋 I've paused before clicking "Sign up". Please complete the signup (including any 2FA / CAPTCHA) in the [mc] tab, then tell me "resume".

4. You do the manual step. The browser is yours during this time — click, type, switch tabs, whatever you need.
5. You say "resume" in OpenCode. The LLM calls browser_resume.
6. The LLM calls browser_get_page_text to see the new state and continues.

Every write tool is hard-blocked between steps 2 and 5. The LLM cannot "forget" and proceed.

Sensitive patterns that auto-pause

place order, buy now, pay now, confirm purchase, complete payment,
sign in, sign up, log in, create account, register, submit,
delete account, wipe, destroy

Plus pressing the Enter key. To explicitly authorize one of these, the agent must be told to call with confirm: true — and even then only when you've explicitly approved that specific action.

Multi-tab workflow example

Scenario: 4 reference tabs open, agent should write a summary into a 5th.

User: "I have 4 articles open. Open a new tab to docs.google.com,
       create a doc, and summarize what's in tabs 0-3 into it."

LLM:
  → browser_list_tabs()
    ← {tabs:[{index:0,title:"Article A"},
            {index:1,title:"Article B"},
            {index:2,title:"Article C"},
            {index:3,title:"Article D"}]}
  → browser_read_tab({index:0, max_chars:3000})
  → browser_read_tab({index:1, max_chars:3000})
  → browser_read_tab({index:2, max_chars:3000})
  → browser_read_tab({index:3, max_chars:3000})
  → browser_new_tab({url:"https://docs.google.com/document/u/0/create"})
    ← {paused:true, reason:"...sign in..."}    # if not logged in
  [user logs in, says "resume"]
  → browser_resume()
  → browser_type({selector_or_text:"Document content", text:"## Summary..."})

The agent never tries to write to tabs 0–3; only the marked [mc] tab.

MVP smoke tests

npm install
npx playwright install chromium
npm run smoke:v4   # v0.4: upload, smart locator, wait_for, dom_summary
npm run smoke:v3   # v0.3: overlay, STOP, focused-typing, evaluate
npm run smoke:v2   # v0.2: multi-tab + pause/resume
npm run smoke      # v0.1: single-tab MVP

Each prints a sequence of ✓ checks ending in ALL CHECKS PASSED.

Layout

opencode-browser-agent/
├── package.json
├── tsconfig.json
├── README.md
├── opencode.jsonc            # example config
├── src/
│   ├── server.ts             # MCP stdio server
│   ├── browser.ts            # Session: 3 modes, tabs, pause/resume
│   └── tools.ts              # Tool schemas, descriptors, dispatcher, safety
└── scripts/
    ├── smoke.ts              # v0.1 single-tab MVP test
    ├── smoke-v2.ts           # v0.2 multi-tab + pause/resume test
    ├── smoke-v3.ts           # v0.3 overlay/STOP/focused-typing test
    ├── smoke-v4.ts           # v0.4 upload/locator/wait_for/dom_summary
    └── fixtures/
        └── upload-page.html  # v0.4 local fixture

Limitations / honest caveats

• The auto-pause regex is a heuristic, not a security boundary. A site that uses an unusual button label can slip through. Read the page text and decide critically before pressing confirm:true.
• In cdp mode the agent could call browser_set_active_tab onto a sensitive tab. The cross-origin confirm:true guard mitigates this but doesn't eliminate it. Run cdp mode with a dedicated user-data-dir.
• This tool will not bypass CAPTCHAs, SMS verification, or any anti-abuse challenge — and that's deliberate. Those exist to ensure a human is present, and the handoff protocol respects that.

License

MIT.