Empir3 Bridge

Your AI subscriptions, finally working as a team.

Connect an AI agent to a dedicated Chrome browser, local tools and, on Windows, desktop apps. Use the local console to choose which abilities to enable. You can use Bridge with a standalone MCP client or pair it with your Empir3 team.

Bridge can also call the AI CLIs you have installed and signed into, including Codex, Grok, Gemini and Claude. Each provider's subscription limits, supported authentication methods and charges still apply. Optional API providers and media tools can incur usage charges; Bridge does not make them free.

Browser actions use page observations and verified targets. Windows desktop tools add native window controls, screenshots and mouse/keyboard input. Read back the result after an action: sending input alone does not prove that the app accepted it.

Three ways to drive it — MCP, CLI, or WebSocket. MCP for the agents you talk to, CLI for your terminal and scripts, or WebSocket via empir3.com for paired remote control. Same engine behind all three — pick the door that fits.

_{One prompt, the whole team — Codex on your OpenAI seat, Gemini on your Google seat, the browser on this machine, all in one thread. (how it's wired ↓)}

Empir3 Bridge turns any MCP client (Claude Code, Codex, Cursor, …) into an agent that can:

1. Run several AI tools from one chat. Ask an agent to delegate to installed, authenticated CLIs and collect their results. Check cli_status first for each CLI's actual availability and authentication state.
2. Drive Chrome with a separate profile. Browser tools use Bridge's dedicated Chrome profile. Its cookies, logins and history are separate from your everyday Chrome profile. Observe the page, select an exact tab and verify each action's result.
3. Operate your desktop — DPI-aware mouse and screenshots across multiple monitors, with per-display click calibration.
4. Generate images and video — 40+ models through the Higgsfield CLI, discovered at runtime.
5. Build one agent from several computers. A desktop can lend the Brain, a laptop can lend Whisper as Ears, and a GPU box can lend Kokoro as Mouth plus ComfyUI as Imagination. Empir3 routes each ask to the selected machine automatically; private endpoint addresses and keys stay on that Bridge.

The Bridge runs on your machine and binds its control servers to 127.0.0.1. Cloud models, websites and paired Empir3 requests still use the network. Write-capable controls start disabled. Enable only the abilities you need; Release control stops active input, and Permissions lets you disable capabilities.

<img align="right" width="92" src="assets/zara-accent.png" alt="Zara, one of the Empir3 agents">

It's the same bridge pattern used inside Empir3, open-sourced so developers can use it directly. Empir3 pairing is optional: local MCP use works without an account, and the remote relay only turns on after you explicitly pair this PC.

It all runs from one local console — what's connected, which capabilities are enabled, and a live log of every action the agent takes:

One prompt, the whole team moving

"Have Codex scaffold the API, Gemini review the diff, then open the dashboard in a real browser and screenshot it for me."

With those CLIs configured, the agent can ask Codex to write, Gemini to review and the Bridge browser to capture the result. Availability and quotas come from your providers; the Bridge returns a failure when a requested tool cannot run.

Why It's Different

Plenty of tools let an agent click around a browser. Almost none let your agents work together. That's the whole point of the bridge:

• CLI delegation. cli_status reports installed tools and their current authentication state. cli_run invokes the selected CLI and returns its result. Provider limits or service changes can still prevent a run.
• A dedicated browser. Bridge's Chrome profile keeps task logins separate from your personal browser. Windows can show a visible browser; headless hosts can use browser tools without a display. Website compatibility still needs to be checked for the task.
• Chat from a trusted local surface. Browser chat and agent replies live on the Bridge's localhost dashboard, outside the JavaScript world of websites Chrome displays. A hostile page cannot submit a message as you or read the agent's replies. Use snapshots, element refs, screenshots, and the native ghost cursor to point at the page without lending that page the control channel.
• A visible desktop focus area. Drag a box around the area you want the agent to work on. Supported screenshots and observations default to that area. The focus box is a convenience, not an operating-system sandbox: explicit captures and other enabled tools can have broader access. The default focus expires after 30 minutes without use; close or release it when finished.
• Native Windows input. Desktop tools map observed controls to screen coordinates and send mouse or keyboard input. Foreground-window and target guards refuse input when the selected target changes. Verify the resulting page or field before continuing.
• Visible and governed by default. Read-only out of the box; a click-through overlay + ghost cursor show what's being targeted; action receipts, per-capability lend toggles, and one-click revoke keep the human in control.

Plus the table stakes: accessibility-ref interaction, multi-monitor desktop control, a safe /desktop-test harness, and a tray app with signed payload updates and version status.

Orchestrate the CLIs you already pay for

Open API & CLIs to inspect supported CLI installations and sign-in status. Enable sharing for each tool you want your Empir3 agents to use. CLI availability, provider plans and usage limits vary.

_{Installation paths are masked in this screenshot for privacy. Status and sharing controls are shown as captured.}

One agent, several of your computers

Install and pair the Bridge on each machine, give every device a recognizable name, then add the ability that computer is willing to lend. A provider can be a chat model, speech-to-text service, text-to-speech service, or image engine. The Bridge includes native wires for Whisper, Kokoro, AUTOMATIC1111, ComfyUI, and OpenAI-compatible speech/image APIs.

In Agent Builder, choose a machine for Brain, Ears, Mouth, and Imagination—or choose Any of my machines when the same provider/model is offered by more than one device. An exact pin never drifts: if that computer goes offline, the request names the unavailable device and stops without substituting another machine or a cloud service. Large images and videos move through a bounded, single-use upload grant instead of being forced into a WebSocket frame.

A browser with its own profile

Browser tools drive Bridge's dedicated Chrome profile and do not attach to your everyday Chrome profile. This separation does not restrict other capabilities you enable, such as desktop input, file access or shell commands. Review those permissions separately.

Chat with the agent from the trusted dashboard

Open http://localhost:3006/ or use Chat in the desktop toolbar. The chat surface is served by the local Bridge and cannot be inspected or driven by the website in the controlled Chrome tab. Standalone mode talks to the connected local MCP client; paired mode can talk to your Empir3 team.

Version 0.3.72 retired the former page-injected chat panel. Injecting a credentialed control panel into arbitrary website JavaScript created the wrong trust boundary: a hostile page could imitate the user. Page context now comes from explicit snapshots, element refs, screenshots, and scoped desktop tools.

You choose what it sees on your desktop

Drag a box around the app you want help with. Supported screenshots and snapshots default to that area. Close the box or use Release control when finished. Use Permissions to control broader desktop, file and shell access; the box does not replace those settings.

Install

Point your agent at this repo and tell it to install. Claude Code, Codex, Cursor — any agent with a terminal — will clone it, install dependencies, and start the bridge from source in about a minute. Or do it yourself with the 60-second Quickstart below. It runs entirely on your machine — no account required.

Signed packages are distributed from empir3.com/download for Windows, macOS, and desktop Linux, plus a lightweight Linux server/VPS package. Windows packages use Authenticode signing, macOS packages use Developer ID signing, Apple notarization, and stapling, and Linux packages are authenticated by the signed release manifest. Installed packages update through the same signed release channel.

Provider APIs, local LLM endpoints, CLI subscriptions, pairing, and MCP are cross-platform. The full native desktop-automation surface (UIA snapshot, pointer overlay, calibration, and native screenshot grids) remains Windows-only; unsupported capabilities fail explicitly on macOS and Linux.

60-Second Developer Quickstart

Clone it, install, and run — about sixty seconds:

git clone https://github.com/empir3hq/empir3-bridge
cd empir3-bridge
npm install
npm start

Or install the current public package from npm:

npm install -g @empir3hq/empir3-bridge
empir3-bridge            # starts the bridge + Chrome  (empir3-bridge --status / --kill to manage)

Open the dashboard:

http://localhost:3006

Then add the bridge to a Claude Code project with .mcp.json. Using the published package requires no local clone:

{
  "mcpServers": {
    "empir3-bridge": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "-p", "@empir3hq/empir3-bridge", "empir3-bridge-mcp"]
    }
  }
}

Or point it at a local clone:

{
  "mcpServers": {
    "empir3-bridge": {
      "type": "stdio",
      "command": "npx",
      "args": ["tsx", "<path-to-bridge>/src/mcp-server.ts"]
    }
  }
}

Try a browser task:

Use the browser bridge to open example.com and take a screenshot.

Or put the headline to work — one agent driving another, using a CLI you already pay for:

Use cli_status to see which CLIs are ready, then use cli_run to have Gemini summarize this README.

What You Get

Orchestrate Other AI CLIs (4 tools)

Use the coding CLIs you're already signed into from your connected agent. Toggle sharing in API & CLIs; the CLI runs locally with its configured provider authentication. Your provider's plan and limits apply.

• cli_status — installed/shared/authentication state, with blockers where known. Call this first, then check the actual run result.
• cli_run — run a lent CLI (codex / grok / gemini / claude) with a prompt and get its text back. mode:"text" returns the answer read-only; mode:"agentic" lets it write files in a