# OpenCode Browse Setup: Existing Logins and Playwright Alternatives

How to connect Browse without giving every agent access, and when a separate Playwright profile is the better choice.

By Paras Madan | 2026-11-17


A coding agent checking an authenticated dashboard needs a browser session. Before choosing the browser tool, decide whose session it should use: your existing account, a dedicated test account, or a fresh anonymous visitor.

[OpenCode Browse](https://opencode.ai/product/browse) connects an MCP server to a Chromium extension so agents can work in your real browser, including its existing logins. It supports clients beyond OpenCode, including Claude Code, Cursor and Codex. That makes it worth evaluating for supervised work inside an application you already use.

This is a documentation-based setup guide. I have not installed Browse or tested its browser connection, so the checks below are an evaluation plan, not a benchmark or a report of successful integration.

## Start with the extension

The current product page recommends installing the [Browse extension](https://chromewebstore.google.com/detail/browse/mfnicocicmmlkpjnaffgihfjhdgjkdjg) first and following its setup flow. It lists Chromium-family browsers, including Chrome, Edge and Brave; Firefox and Safari are not on that list. A direct CLI installation is optional:

```bash
npm install -g @opencode/browse && browse install
```

Do not run both paths just to be thorough. Start with the extension flow and use the command only if you deliberately choose the CLI route. Before connecting anything, choose the browser profile and account you are comfortable letting the agent operate.

The [repository README](https://github.com/anomalyco/browse) specifies Node 22+ for the npm launcher. Its terminal installer starts with detected agents selected. Deselect clients you do not intend to connect. Restart an already-running agent, or reload its MCP servers, afterward.

For missing tools or a connection problem, inspect status before reinstalling:

```bash
browse status
```

The README documents `browse pair` for manual pairing if needed. Avoid changing network settings to troubleshoot a local setup.

## Limit the first task

Browse documents `allowedOrigins` as an opt-in tab navigation restriction and `readOnly` for inspection-only tabs. Its toolbar button lends an existing tab and takes it back on another click. The README explicitly says the `execute` sandbox is not a security boundary.

My first task would be a read-only check of a staging dashboard with synthetic records. Ask the agent to use the restriction options, then inspect the actual tool arguments. A prompt saying “only use this site” is not evidence that the tool-level restriction was configured.

Give the task a small, observable finish line: report one known heading and one synthetic row count, then stop. Exclude exports, form submissions and account settings. Keep unrelated customer or personal information out of the test profile.

## Keep human steps human

Browse's product page describes handoffs for login, two-factor authentication and payments. The person completes the step and clicks Continue before the agent resumes.

For an initial evaluation, choose a task that needs no payment at all. If authentication is required, complete it yourself and confirm the account shown before continuing. Write the approval boundary into the task: reading a checkout page should never silently become permission to place an order.

## Browse versus Playwright MCP

Existing logins are not unique to Browse. The official [Playwright MCP documentation](https://playwright.dev/docs/getting-started-mcp) lists an extension mode that connects to existing browser tabs using `--extension`.

Playwright also offers a persistent profile, which keeps cookies and login state between sessions, and `--isolated`, which starts fresh. Its browser options include Chrome, Firefox, WebKit and Edge. These modes make the account-state decision explicit.

I would evaluate Browse when the work is supervised and depends on the browser session already in front of me. I would start with a separate Playwright profile when I need to repeat a flow without personal account state affecting the result. If both tools fit, compare them on the same staging task rather than deciding from a feature list.

## Record what actually happened

For each trial, keep four notes: which account was active, whether the agent read the expected data, whether it stopped at the agreed boundary, and whether you could reclaim control. Record connection failures too. A correct final answer does not prove the agent stayed within scope.

Repeat with a fresh session before relying on the workflow. If success depends on a remembered login or a previously opened tab, document that dependency. This gives the next developer something reproducible instead of a demo that only works on your machine.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help developer-tool teams turn setup decisions and technical tradeoffs into practical guides.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
