The Linux distribution where every coding agent works as one team.

AgentUX ships Claude Code, Codex, OpenCode and Antigravity CLI already installed and wired together on an immutable Fedora base: one cockpit over all of them, an agent bus so they can talk to each other, and an orchestrator that takes an issue to a reviewed pull request.

Pre-release. Fedora Kinoite 44 with bootc, signed images, x86_64 and aarch64. What is tested so far

Bring every agent you already pay for

Running agents in parallel worktrees is solved. Getting agents from different vendors to work on the same task, and keeping track of them, is not. AgentUX does that part, and leaves inference with the providers: your own logins and API keys, no GPU required.

One interface for every vendor
Sessions, diffs and permission requests from any harness look the same in the cockpit, and land in one approvals inbox. The vendor's own TUI is always one key away.
Agents that talk to each other
An implementer can ask a reviewer from another vendor to check its diff, hand off a task, or escalate a decision to you, through a mediated bus where every message is logged.
From issue to reviewed pull request
The pipeline (plan, implement, test, review, open the PR) is declared in agentux.yaml and versioned with your repository. You are asked only for the decisions that are yours.
A system agents cannot break
An immutable, image-based OS with one-step rollback. Every task runs in its own git worktree, and gate checks can run in rootless Podman.

One platform, from the node to the organization

AgentUX is becoming a suite: local nodes execute the work, while a remote command center lets leadership see missions, teams and operations across the organization.

AgentUX OS
The immutable Linux station for running agents locally, with the cockpit, vendor CLIs and safe rollback.
AgentUX Core
The local runtime: worktrees, ACP sessions, approvals, gates, agent bus and pull-request automation.
AgentUX Command
The remote control plane for organizations, missions, governance, telemetry and multiple AgentUX nodes.
AgentUX Office
A visual command center where leadership can see departments, agents, missions and incidents as they unfold.

Install the role you need

One signed AgentUX image can become a workstation, a remote node, a Command server or an executive Office screen. The installer records the role and first boot activates only the selected platform capabilities.

Workstation
Local node plus the desktop cockpit for people building and reviewing work.
Node
Headless execution and registration for machines managed remotely.
Command
The optional organization control plane, isolated in rootless containers.
Executive
A lightweight remote Office view with no local orchestration stack.

What it looks like

Captured by CI from the real images booted in QEMU. The runs on screen use fake agents, so nothing is billed.

The AgentUX cockpit on the live ISO: a run board with two demo runs, one of them waiting for plan approval in the inbox on the right.
The cockpit: a run board for every project, and an inbox for what is waiting for you.
The boot splash: the AgentUX penguin swimming in place, its wake streaming back, with a wavy progress line.
The boot splash. The penguin swims while the system starts.
The AgentUX web installer on the live ISO.
The installer on the live ISO, in AgentUX colours.
First-boot setup over the AgentUX wallpaper.
First boot: create your account, then the agent CLIs install themselves.

How a run works

Hand AgentUX an issue or a prompt, from the cockpit or the aux CLI. This is what happens next.

  1. AgentUX creates a git worktree and a branch for the task.
  2. A planner writes a plan, and waits for your approval if the pipeline asks for it.
  3. An implementer harness works in the worktree. Its session streams to the cockpit.
  4. Tests and linters gate the result. Failures loop back to the implementer, up to a limit.
  5. A reviewer from a different vendor reviews the diff, and talks to the implementer over the agent bus when it needs to.
  6. AgentUX opens the pull request.

agentux.yaml, the default pipeline, shortened

roles:
  planner:     { harness: claude-code }
  implementer: { harness: claude-code }
  reviewer:    { harness: codex }   # another vendor

pipeline:
  - step: plan
    approve: true            # wait for you
  - step: implement
  - step: gate
    checks: [lint, test]
    on_fail: implement
    max_attempts: 3
  - step: review
    on_changes_requested: implement
  - step: pull_request

The parts

Harnesses over ACP
Each provider's own CLI, driven through the Agent Client Protocol, with headless modes as the fallback. AgentUX never handles your credentials. ADR 0002
The agent bus
agentuxd exposes the bus to every harness as an MCP server: post_message, request_review, handoff, get_run_state, ask_human. The daemon routes every message and caps turns per exchange, so two agents cannot loop forever. ADR 0004
The orchestrator
A workflow engine in agentuxd runs the pipeline as a persisted state machine, with budgets per run and approvals where you ask for them. ADR 0003
The cockpit and aux
A Tauri desktop app over every session the daemon runs, and its terminal twin, which works over SSH. Any session can switch to its vendor's TUI on the same conversation. ADR 0007
The OS
Fedora Kinoite 44 built with bootc: KDE Plasma, or a Hyprland tiling session in the second image. Immutable, with atomic updates and rollback, and every image signed with cosign. ADR 0010

Where it stands

AgentUX is pre-release (agentux-core 0.4.2 and agentux-desktop 0.4.5). Runs already go from prompt to pull request through real harness sessions. This is what has been checked, and what has not.

Verified in CI

  • Both variants build and pass a first-login smoke test on amd64 and arm64.
  • The Plasma image boots nightly under QEMU/KVM on x86_64 and takes a fake-agent run to a pull request.
  • The x86_64 live and installer ISOs are installed in a VM before they are published.
  • Image signatures are checked against the policy installed systems enforce.

Not verified yet

  • Full cross-vendor runs with real Claude Code, Codex and Antigravity sessions. OpenCode has been tried.
  • Booting the arm64 image on real hardware or a VM: CI's arm64 runners have no KVM.
  • Booting the Hyprland image. Its configuration is checked, not run.

The roadmap lists what is done and what comes next.

Documentation

The docs live with the code, on GitHub.

Boot it, look around, install when you want to keep it.