Agents Use Repo

Why?

As we step into AI using MCPs and Agents, additional setup can be used to leverage full potential. This repo is dedicated to make prompting and AI use easier.

Setup a browser session for agent use

One-time prerequisite: cp "config.json example" config.json and adjust paths.

Each named session launches its own visible Brave instance with an isolated profile (own cookies, tabs, logins) and binds it to the matching agent-browser session. Your personal browser is not touched — unless you explicitly ask for it with the reserved session name main — and even then you can keep it unexposed with --no-expose (see below).

node setup-brave.js            # isolated session "default"
node setup-brave.js work       # isolated session "work"
node setup-brave.js research   # another one, in parallel
node setup-brave.js main       # your DAILY Brave — agent sees ALL its tabs
node setup-chromium.js         # Chromium variant, session "chromium"

node setup-brave.js main --no-expose   # daily Brave, NOT exposed to any agent

Opening a browser without exposing it

Add --no-expose (aliases: --no-bind, --private) to launch the session's browser with no remote debugging port and no agent-browser binding. There is nothing for an agent to attach to, and the existing session binding is left untouched — agent-browser --session <name> just falls back to its own invisible headless browser.

node setup-brave.js main --no-expose      # daily browser, agent-proof
node setup-brave.js work --no-expose      # open the "work" profile to log in by hand
node setup-chromium.js web --private      # same for Chromium

• Works for any session, not just main — handy for logging into a session profile yourself before an agent uses it.
• No port is reserved in sessions.json for an unexposed run.
• If that browser is already running exposed, the script refuses: a running process keeps its debug port for life, so relaunching cannot take it away. Quit it (Cmd+Q for your daily Brave) and re-run.
• To expose it later, quit the browser and re-run the same command without the flag.

Then tell the chat/agent which session to use:

Use agent-browser with --session work for all browser commands.

or set AGENT_BROWSER_SESSION=work in that agent's environment.

Notes:

• Session names are lowercase (work, not Work — macOS paths are case-insensitive).
• Session profiles persist under ~/Library/Application Support/AgentBrowserProfiles/, so site logins survive restarts (log in once per session, by hand if needed).
• Ports are assigned per session in sessions.json (gitignored, auto-managed; don't delete it while session browsers are running).
• main requires Brave to be started by the script: if your daily Brave is already running without remote debugging, fully quit it (Cmd+Q) first.
• If you close a session's browser window, re-run node setup-brave.js <name> before agents use it again — otherwise agent-browser silently falls back to launching its own invisible headless browser for that session.
• agent-browser session list shows active sessions.

Migrating from the old setup (pre-sessions): the old script exposed your main Brave on port 9222 and bound agent-browser's default session to it. To sever that, fully quit and relaunch Brave normally, then run node setup-brave.js once to rebind default to an isolated browser.

Use of prompts

Prompts are available to use in chat under /

Use of instructions file

Enable at settings: github.copilot.chat.codeGeneration.useInstructionFiles