First, three things you need once
- A terminal. That is a window where you type commands. On Windows: press the Windows key, type
PowerShell, press Enter. A blue or black window opens; you type into it and press Enter. Keep it open for the steps below. - Python. Go to python.org/downloads, click the big Download button, run the installer, and on its first screen tick "Add python.exe to PATH" before clicking Install. That tick is what lets the commands below find Python. Then close and reopen the terminal and type
python --version: you should see a version number, 3.10 or higher. - The pack. Click "Download the bridges pack" above. Right-click the zip, Extract All, and put the folder somewhere permanent, for example
Documents\Agent Go Bridges. Inside you will see two folders:desktop-serverandwhisper-server.
A. Let the agent use your desktop and run commands
- Install the helper's parts. In the terminal, paste this and press Enter:
pip install flask pyautogui mss pillow. Lines scroll by for a minute and end with "Successfully installed". If you also want the agent to read PDF files on your disk, pastepip install PyMuPDF pytesseractas well. - Start the helper. In the extracted folder, open
desktop-serverand double-clickstart-desktop.bat. A black window opens and stays open; that window is the helper running. Leave it open while you use the agent (close it when you are done). - Copy your token. The first time, that window prints a line with a long code called the desktop token. Select it and copy it (or open the file
desktop-server\.desktop-tokenin Notepad; the token is in there and stays the same). - Give the token to the extension. In the side panel click ⚙ Settings, scroll to Advanced settings, local machine bridges, paste the token into Desktop token. Leave "Desktop control server URL" as
http://localhost:8777. Click the Test button under it: it should turn green and show your screen size, for example "screen 1920x1080". - Choose what to allow. Tick Enable desktop control if you want the agent to move the mouse, type and take screenshots outside the browser. Tick Enable Run commands if you want it to run commands like tests or builds, and type the folder they should run in under Project directory (for example
C:\Users\you\my-project). Click Save. - Try it. In the side panel type "Take a screenshot of my desktop and describe what you see". Before it acts, the panel asks you to approve; click Approve. If anything ever goes wrong, push the mouse hard into any corner of the screen: the helper stops immediately.
B. Talk to the agent, transcribed on your own computer
- In the terminal paste
pip install flask faster-whisperand press Enter. - In the extracted folder open
whisper-serverand double-clickstart-whisper.bat. The first time it downloads the speech model (a few hundred MB); wait for the window to say it is ready. Leave the window open. - In Settings, the "Local Whisper server URL" stays
http://localhost:8765; click its Test button, it turns green. The 🎤 button in the side panel now works through your computer. If the helper is not running, the button quietly uses Chrome's own speech service instead.
C. Use your Claude Code or ChatGPT subscription as the model
If you already pay for Claude Code or ChatGPT, the desktop helper from section A can pass the agent's requests to that tool, signed in with your own account. The model is then billed to that plan, and no API key is needed. This is off until you turn it on.
- Install the tool and sign in, once. This needs Node.js: if
node --versionshows nothing, install it from nodejs.org (the "LTS" button), then reopen the terminal. For Claude, pastenpm install -g @anthropic-ai/claude-code, then typeclaudeand follow the sign-in it shows you. Check withclaude -p "say ok": it should answer "ok". For ChatGPT, pastenpm install -g @openai/codex, thencodex login. - Allow the helper to use it. In the
desktop-serverfolder, right-clickstart-desktop.bat→ Edit (it opens in Notepad). Find the lineREM set DESKTOP_ENABLE_CLI_BRIDGES=1and delete theREMat its start so it readsset DESKTOP_ENABLE_CLI_BRIDGES=1. Save, close the helper window if it is open, and double-clickstart-desktop.batagain. - Tell the extension. Make sure section A steps 3 and 4 are done (token pasted, Test green). Then at the top of Settings, under Inference provider, choose Claude (SUBSCRIPTION, via local Claude Code CLI) and click Save. The badge in the panel header now names that provider, and your next task runs through your plan.
- Good to know. The ChatGPT choice is for reviews and drafts only: its tool returns text and cannot click or fill pages. Long answers can take a minute; the panel keeps showing progress. Using your plan this way is between you and Anthropic or OpenAI; their terms apply.
Stopping, and staying safe
- To stop everything, close the helper windows and untick the options in Settings. The agent goes back to browser-only.
- Keep "Enable Run commands" off when you do not need it: commands run with your own user rights, exactly like typing them yourself.
- The desktop token is the only thing that stops other programs on your computer from using the helper, so do not paste it anywhere but the extension's Settings.
- The helpers talk only to the extension on this computer. Nothing on your network can reach them, and they send nothing to us.