AgentMux

<!-- The alpha warning is the first content block by design — see
docs/specs/SPEC_EARLY_ALPHA_WARNING_2026_06_05.md. Decorative logo/title
follow below so the warning is never pushed below the fold. -->

> ## Alpha
>
> **AgentMux is alpha software.** Some features are incomplete or still
> settling, and behavior can change between releases. A few things to keep
> in mind:
>
> - **Rough edges** — parts of the UI may not work as expected, and
>   something that works today can regress in a later release.
> - **Don't rely on persistence** — settings, pane layouts, and agent state
>   may not carry over cleanly across versions, so avoid storing anything
>   you can't easily recreate.
> - **Interfaces are still moving** — config files, accounts,
>   bundles, and MuxAPI can change shape without a migration
>   path during alpha.
> - **Platform coverage varies** — Windows gets the most testing right now;
>   macOS and Linux builds are a bit further behind.
>
> Running into something? **Please open a GitHub issue** at
> https://github.com/agentmuxai/agentmux/issues — that's how alpha gets to beta.

---

<p align="center">
  <img src="./frontend/logos/agentmux-logo-brain-alternate.svg" alt="AgentMux Logo" width="120">
</p>

# AgentMux

**The agent operating environment**

Run any agent as a first-class pane — Claude Code, Codex, Gemini, Copilot, and more — each with its own identity, native memory, and a structured view of tool calls and diffs. Agents can drive the workspace itself through a local API. Multi-provider, local-first, with a Rust core. Free and open source.

[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
[![Website](https://img.shields.io/badge/Website-agentmux.ai-blue)](https://agentmux.ai)

## The Problem

Today's agents run inside terminal wrappers — raw stdout, no real identity, no shared context, and no way to talk to each other. You re-onboard a stranger every session, you can't see what an agent is actually doing mid-task, and long-lived provider CLIs sit there hogging memory the whole time.

- **Agents are afterthoughts.** A terminal wrapper hands you raw output, not a structured view of tool calls, reasoning steps, and file diffs.
- **Every provider is a silo.** Claude, Codex, and Gemini each run in isolation — no shared workspace, no interop, no way for one agent to hand off to another.
- **You're the integration layer.** Copy-pasting between windows, re-explaining the same context, babysitting heavyweight CLI processes.

## What AgentMux Does

AgentMux is an open-source agent operating environment. Run any agent as a first-class pane — with its own identity, native memory, and a structured view of tool calls and file diffs — and let agents drive the workspace itself through a local API. AgentMux owns the session state, so many provider CLIs run one-shot per turn instead of as long-lived processes, and one Rust core holds many agents at once.

Cross-platform (Windows, macOS, Linux). Rust backend (Tokio + Axum). CEF host (bundled Chromium). Apache 2.0.

- **Multi-provider agent panes** — Claude Code, Codex, Gemini, GitHub Copilot, Qwen, Kimi, OpenClaw, Pi, Mux Code and Antigravity as first-class providers, alongside terminal, editor, browser, file and media panes ([all widgets](#widgets)). Structured views of tool calls, reasoning, and diffs — not a terminal wrapper.
- **Universal pane tabs** — every pane is a generic container that can hold any mix of widgets as tabs, not just more of its own type. A pane's own tab strip *is* its header (one row, not a header-plus-strip stack), and its "+" opens a picker of every widget type — agent, terminal, browser, editor, sysinfo, swarm, drone, and the rest — so you build up an agent-plus-terminal-plus-browser workspace inside a single pane instead of endlessly splitting. Inspired by [cmux](https://github.com/manaflow-ai/cmux)'s `Workspace → Pane → Surface → Panel` model. See [Panes and tabs](#panes-and-tabs) below.
- **Agents drive the workspace** — Through MuxAPI, a running agent can open panes, rename tabs, navigate the layout, and message peer agents, using the tools of its built-in MCP server. Agents are operators, not passengers.
- **Interagent comms** — `SendMessage` routes one agent's output into another agent's input, so you can build hand-offs and reactive pipelines.
- **Swarm** — A live view of every agent and its subagents, background tasks and todos, including agents on other machines you're connected to. Follow one agent, or broadcast a message to several at once.
- **Accounts** — Sign-ins and keys for providers and services (Claude, Codex, GitHub, AWS, API keys and more), kept in Connectors → Accounts and bound to each agent, at most one per provider. They survive renames; switching an agent's account restarts that agent.
- **Bundles** — A collection of instructions, context files and skills, captured once and reused across agents (renamed from "presets"). MCP servers are

... (truncated)