RealBrowserAttach

Real Browser Attach

Use this side-skill when the goal is to control the user's real browser context rather than an isolated automation browser.

Routing

Prefer this skill over generic browser automation when the request includes any of:

• "real browser"
• "existing browser"
• "logged-in browser"
• "Windows browser"
• "Brave"
• "attach to Chrome"
• "native messaging"
• "remote debugging port"
• "CDP"
• "RunBrowser"
• "Interceptor"

Core Doctrine

The substrate comes first. Do not start with Browserbase autobrowse, Playwright, or a headless Linux browser when the user explicitly wants the browser they already use.

Choose in this order:

1. Native-messaging host via wsl.exe stdio for real Brave Current primary build target. This is the PAI-owned path for driving the already-running Brave profile on Windows with zero network exposure.
2. dbalabka/chrome-wsl bridge Validation and fallback path for real Chrome/CDP workflows. Useful baseline, but no longer the main roadmap driver.
3. Extension relay fallback Use mcp-chrome first, then RunBrowser if needed. These remain contingency substrates if the native-messaging path stalls.
4. Mirrored networking or manual portproxy Experimental and last-resort. Only use after the higher-confidence paths fail and only with explicit validation of networking side effects.
5. Browserbase remote/browser automation Use for isolated sessions, bot-protected sites, or cloud automation. Do not treat it as the default answer to "use my real browser."

Current Guidance For This Machine

Read these references before deciding the implementation:

• tooling-comparison.md
• windows-attach.md

The current local findings matter:

• Browserbase browse is not the canonical real-browser path here.
• Interceptor is useful conceptually, but is currently macOS-only.
• Windows Chrome CDP can run, but WSL attachment is not plug-and-play on this machine.
• The active build decision is now: native-messaging relay for real Brave first; all CDP bridge paths are fallback-only.

Workflow

1. Confirm the target browser surface

Decide whether the user wants:

• the actual daily browser/profile
• a disposable Windows browser
• an isolated cloud browser

If they want the real daily browser on this machine, bias to the native-messaging Brave relay first.

2. Pick the substrate

Use this decision rule:

• Real existing Brave browser on this machine: build or extend the native-messaging relay path first
• Need a comparison baseline or fallback: try chrome-wsl
• Need a third-party real-browser fallback: prefer mcp-chrome, then RunBrowser
• Need cross-machine or bot-resistant automation: Browserbase remote
• Need standards-based debugging and the boundary is controllable: direct CDP attach only after proven connectivity

3. Validate the boundary before promising success

For WSL plus Windows browser:

• do not assume 127.0.0.1:9222 in WSL reaches Windows Chrome
• do not assume --remote-debugging-address=0.0.0.0 will expose Chrome beyond Windows localhost
• empirically test connectivity first

If CDP is chosen, validate:

curl -fsS http://<target>:9222/json/version
curl -fsS http://<target>:9222/json/list

If those fail, stop calling it a CDP solution and switch to:

• the native-messaging relay if it is the target path
• chrome-wsl if a fallback baseline is needed
• an extension relay
• a Windows-side proxy/forwarding layer

4. Keep agent compatibility unified

This workflow should work for both Claude and Codex:

• avoid Claude-only slash-command assumptions
• avoid Codex-only home-path assumptions in user-facing doctrine
• keep the browser strategy in PAI-owned skills, not in ad hoc local notes

5. Position Browserbase correctly

Browserbase autobrowse is an optional training loop once the browser substrate is healthy.

It is not the first thing to deploy when:

• the user wants the real browser
• the browser is on Windows and the agent is in WSL
• the goal is to preserve logins, cookies, tabs, and extensions

Browserbase ui-test is an optional adversarial testing layer once the same substrate is healthy.

For the current roadmap, both Browserbase layers sit downstream of the native-messaging Brave relay rather than competing with it.

Output Expectations

When you use this skill, produce:

• the chosen substrate and why
• the exact blocker if attachment is not yet possible
• the next concrete step to make it real

Do not say "browser automation is installed" if the agent still cannot reach the intended browser surface.