LaunchRepo with OpenCode: browser MCP setup, tested

Short answer

• Put the chrome-devtools entry under the "mcp" key of opencode.json, with "type": "local" and the whole command as one array. The block is below.
• OpenCode 1.18.34 loaded it from the global file, from a project opencode.json (also when started in a subfolder), from .opencode/opencode.json and from both environment variables.
• It then offered 30 browser tools named chrome-devtools_… and passed the full 44,167-character page snapshot to the model without shortening it.
• After a CAPTCHA or 2FA step, resume the same conversation with opencode run --session <session-id>. That worked headless in every run: same tab, same fields, one submission.
• OpenCode allows most actions by default, so add one permission line if you want to approve browser actions yourself.

LaunchRepo works with any coding agent that can read and write files in your workspace and control a browser. OpenCode does both. This page documents a tested setup; it is not a certification or partnership.

Add the browser to opencode.json

OpenCode merges its config files instead of replacing them. That makes setup safe: you add one key and every setting you already have stays as it is.

1. Open opencode.json in the root of your LaunchRepo workspace, or create it. For all projects at once, use ~/.config/opencode/opencode.json instead.
2. Add the "mcp" block below next to your existing keys. If the file already has an "mcp" key, add only the chrome-devtools entry inside it.
3. In the Chrome you launch from, open chrome://inspect/#remote-debugging and enable remote debugging (Chrome 144+). That is what --autoConnect connects to.
4. Run opencode mcp list. A working entry shows the line below; then ask OpenCode to open https://example.com and read the page title.

Add to opencode.json

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "chrome-devtools": {
      "type": "local",
      "command": [
        "npx",
        "-y",
        "chrome-devtools-mcp@1.10.1",
        "--autoConnect",
        "--no-usage-statistics",
        "--no-performance-crux"
      ]
    }
  }
}

Expected line in opencode mcp list opencode mcp list

✓ chrome-devtools connected

--autoConnect uses the Chrome you already run. For a separate Chrome just for launches, use --browser-url=http://127.0.0.1:9222 instead; that is how our test connected to its temporary Chrome.

Which config files OpenCode loads

Same server entry, different places and shapes. “Loaded” means the chrome-devtools_… tools reached OpenCode’s request to the model.

The two rows that bite: a pasted "mcpServers" block is ignored without a word (opencode mcp list says “No MCP servers configured”), and a string command stops OpenCode before it starts (“Expected array, got "npx"”). Both are fixed by the array form above.

The handoff run, step by step

OpenCode 1.18.34 with Chrome DevTools MCP 1.10.1, Google Chrome 154.0.8037.92, macOS, Node 22, on October 1, 2026. Every run used a fresh home directory and a fresh Chrome profile, so no personal settings were read or changed.

No AI model was involved. A small local stand-in answered OpenCode’s model requests with fixed tool calls and read element IDs only from what OpenCode actually sent it. That makes every run identical and free and tests what OpenCode controls: config, tools, what the model sees, permissions and sessions. It does not test how well a model plans.

Handoff run on the demo form of our CAPTCHA guide: 5 runs. Config variants: 3 runs each.

OpenCode opened the demo form, filled it, stopped at the stand-in check and ended the run. A script played the person and completed the check in the same tab. Then a second opencode run with --session continued the same conversation and finished. CAPTCHA and 2FA in agent workflows: the controlled handoff →

Resume after the person finished

Every opencode run prints its session ID in --format json output, and opencode session list --format json lists it too.

Continue with the same ID. In our runs the continued session had every earlier step, found the same tab through Chrome DevTools MCP, saw the fields unchanged and submitted once.

Prefer the terminal UI? opencode --session <session-id> opens the same conversation there.

Find the session

opencode session list --format json

Continue headless

opencode run --session <session-id> "The check is done. Re-read the form and submit once."

Continue in the terminal UI

opencode --session <session-id>

Approvals: ask, allow, deny

Without a permission block, OpenCode runs browser tools without asking. One line changes that. We ran each rule headless, the way scripts call OpenCode:

Approve every browser action yourself

{
  "permission": {
    "chrome-devtools_*": "ask"
  }
}

Use "ask" in the terminal UI, where OpenCode offers once, always or reject. For headless runs, either accept that "ask" blocks the browser or use --auto knowingly, on a Chrome profile meant for launches.

Coming from Claude Code

The server is the same, the config is not. Four differences decide whether your entry works:

Three things that cost time

A pasted "mcpServers" block is silent

OpenCode ignores the key entirely: no error, exit code 0, no browser tools. Move the entry under "mcp" and turn the command into one array.

Headless runs wait on open stdin

When a script starts opencode run with stdin left open as a pipe, OpenCode waits for input that never comes. In all 3 runs it hung until stopped; with stdin from /dev/null it answered at once.

"ask" means no in scripts

There is nobody to answer in a headless run, so every "ask" becomes a rejection. Decide per run whether the browser may act alone.

Start LaunchRepo in OpenCode

Open OpenCode in your LaunchRepo private workspace and paste the start prompt. The kit then leads onboarding, tests your Chrome connection on Example Domain and builds the run.

• The start prompt recommends Claude Code or Codex CLI for the best results. OpenCode works with the same kit because it has the two things the kit needs: file access in the workspace and a browser through MCP.
• OpenCode reads AGENTS.md from the workspace, so it picks up the kit’s rules on its own.
• CAPTCHA and 2FA steps come back to you as handoffs; you resume the session with --session as shown above.

LaunchRepo start prompt

Recommend Claude Code CLI or Codex CLI in a terminal instead of the desktop apps for best results. If session metadata confirms desktop hosting, explicitly advise me to switch to the CLI and resume this same private workspace. Do not infer the host from a CLI engine version alone.

Read AGENTS.md, then follow "Start the dashboard for the owner". Create my private
workspace if I do not have one yet, start the dashboard, and give me the link.
Then lead ONBOARDING.md one phase at a time. Reuse recorded permissions instead of
asking again unless their scope is missing or conflicting. Actually test my visible
existing Chrome profile and intended identity as BROWSER-SETUP.md requires; use the
OS fallback if needed. Use one platform/browser worker by default. Build the normal
run only from suitable free-ready routes, keep research separate, and preserve any
draft that needs one exact human CAPTCHA/2FA action. Never claim browser control
bypasses or reduces those challenges.

• Connect your coding agent to your Chrome →
• Launch your SaaS on directories with Claude Code →
• Playwright MCP vs Chrome DevTools MCP →
• LaunchRepo with Kiro: where the browser config goes →
• MCP server →

Not covered by this test

• The terminal UI’s approval prompts and question tool. The headless behavior above is measured; the UI behavior follows the docs.
• A real AI model. The stand-in removes model quality from the result; a model can still pick a wrong element.
• --autoConnect to an everyday Chrome profile. We used a temporary Chrome through --browser-url.
• Windows and Linux, remote MCP servers, OpenCode accounts and paid model plans.

Sources

Every quote above is copied word for word from these pages, read on October 1, 2026. Tested behavior comes from our own runs described on this page.

FAQ

Where do I put MCP servers in OpenCode?

Under the "mcp" key of opencode.json, either in your project root or in ~/.config/opencode/opencode.json. In our test OpenCode 1.18.34 also loaded .opencode/opencode.json, OPENCODE_CONFIG and OPENCODE_CONFIG_CONTENT.

What is the OpenCode config for Chrome DevTools MCP?

An entry named chrome-devtools with "type": "local" and "command": ["npx", "-y", "chrome-devtools-mcp@1.10.1", "--autoConnect", "--no-usage-statistics", "--no-performance-crux"]. Run opencode mcp list to see it connected.

Why does OpenCode not show my MCP server?

Most often the entry was copied from another agent: an "mcpServers" block is ignored without an error, and "command": "npx" with separate "args" makes OpenCode reject the whole config. Use "mcp" and one command array.

Can I use Playwright MCP instead?

Yes, the same "mcp" entry shape works for any local MCP server. This page tests Chrome DevTools MCP because it attaches to the Chrome you already use; our comparison page explains when Playwright MCP is the better fit.

How do I resume an OpenCode session after a CAPTCHA?

Complete the check in the same Chrome tab, then run opencode run --session <session-id> with a short instruction, or open it in the terminal UI with opencode --session <session-id>. In our runs the session, tab and fields were all still there.

Is OpenCode officially supported by LaunchRepo?

LaunchRepo supports coding agents that have file and browser access; OpenCode has both, and this page documents a tested setup. It is not a certification or partnership, and the start prompt still recommends Claude Code or Codex CLI for the best results.