Fair source FSL-1.1 Linux
AgentBox runs a team of AI coding agents on one project. Each one gets a full Linux machine and a branch of its own, and a lead chat directs them. Shared project memory and token accounting come with it. It works with Claude Code, Codex and OpenCode.
AgentBox
chmod +x AgentBox-*-x86_64.AppImage && ./AgentBox-*-x86_64.AppImage
Or the .deb, the .pacman, or the CLI alone, from
the latest release.
One chat, four agents, one afternoon
You talk to one chat. It runs a team of agents, each in its own container on your computer, and tells you when they're done.
Point AgentBox at a git repository. Its chat, the project's lead, is ready straight away and costs nothing until you talk to it.
agentbox add ~/src/my-app
The lead creates the agents the job needs and briefs each one with what the project already knows.
agentbox chat my-app "Paginate the reminders list"
Each agent gets an Incus container, a git worktree on agentbox/<name>, its own network, Docker, a browser and a desktop. If an agent is stuck on a decision, the lead answers it or passes it to you.
agentbox shell my-app/agent-01
A finished agent hands the lead its summary and pull request. Screenshots, recordings and test reports wait in its Media tab. Destroy the agent and its branch stays.
agentbox diff my-app/agent-01
Desktop app ─┐ ┌─ agent-01: Incus container ── worktree on agentbox/agent-01
├── unix socket ── daemon ──┼─ agent-02: Incus container ── worktree on agentbox/agent-02
agentbox CLI ┘ (HTTP) state.db └─ lead: the project chat, on your machine, over MCPIsolation
Every port is free in every agent, so each one runs the dev server on its usual port and its own Postgres in Docker. Agents are capped at a share of the host's CPUs, so one agent's build can't freeze your machine. You can snapshot an agent, restore it, fork a new one from it, or pause it with its memory intact.
One timeline
The app talks to each tool through ACP adapters, with the tool running inside the agent's machine. Tool calls, diffs, plans, approvals and subagents all land in one timeline, and you can message an agent while it works. Pick the model, effort and permissions per agent. If you'd rather use the tool's own CLI, agentbox shell attaches to it.
Watch and take over
Every agent has Chromium on a desktop of its own. The agent drives the browser through Playwright and the desktop with a real mouse and keyboard. You watch it live in the Desktop tab, and when you click Take control, it's yours.
Project memory
Each project keeps its decisions, discoveries, open issues, task graph, artifacts and every agent's finish report in SQLite, searched with full-text search. No model call and no embeddings. Each new agent starts with the slice its task needs, under a token budget you set.
Token accounting
A ledger records every turn of every chat: tokens, estimated cost and context size, per model. Each Claude account's five-hour and weekly limits show in the top bar. Chats compact at a window you set, and screenshots and searches run in subagents, so a long task doesn't keep getting more expensive per step.
Agents keep screenshots, recordings with the cursor and keystrokes shown, test reports and logs in a Media tab, so you can review their work without starting them.
Any port an agent listens on is one URL away: <port>.<agent>.<project>.localhost:7777.
For mobile projects, each agent can have its own emulator with its own app data, and you can use it from the app.
Set up one agent with dependencies installed and a database seeded, then save it as the base that new agents start from.
Pick an account per project or per agent. The lead can spread new agents across accounts by how much of each one's limits is left.
Run AgentBox on a VPS as well as your PC and use both from one app. A VPS's agents keep working while your PC is off.
Give an API key to a project or to a single agent. It's sealed at rest, and the API and the app can never read it back.
Every project and agent in one place, ordered by what needs you first. Light and dark themes, or follow your desktop's.
Running agents in parallel usually means picking a compromise. Either they share your machine, or they run on somebody else's.
Parallel agents, one machine
Isolated, but on someone else's computer
Isolated, on your own computer
Agents are Incus containers, and Incus runs on Linux, so Linux comes first. Ports to macOS and Windows are in progress but haven't shipped.
x86_64. Arch, Debian 12 or 13, Ubuntu 22.04 or newer, and Fedora. AppImage, .deb, .pacman, or the CLI alone.
Port in progress. Agents are Incus containers, which need a Linux VM underneath.
Port in progress, through WSL2.
It's the same idea: you describe the work and autonomous agents do it in full development environments, with a browser, a terminal and a desktop you can watch. The difference is that the machines are containers on your own computer, the agents are the CLIs you already use, and you can run many of them at once on one project.
The repository, the agents' machines, the project memory and the token ledger all stay on your computer. The only thing that leaves is what any coding agent sends: the context the model reads goes to your AI provider, under your own account.
Claude Code, Codex and OpenCode, or just a shell. You sign in with your own accounts. You can add several Claude accounts and choose one per project or per agent.
Incus system containers boot in seconds and share the host's kernel, yet each one is a full Linux machine with systemd, Docker, its own network and its own users. You can run several at once on a laptop. Each running agent uses 1 to 2 GB of memory.
Linux on x86_64, 8 GB of memory at the very least (more for several agents at once), 4 cores, 30 GB of free disk, and a user who can run sudo. The app's Settings walk through installing Incus and downloading the agents' base image, without a logout.
Yes. The daemon and the agentbox CLI do everything the app does, which is how it runs on servers and in scripts. A hub connects a VPS and your PC so you can use both from the app or a phone's browser.
Yes. AgentBox is free to use, self-host and change under the Functional Source License (FSL). The one thing it rules out is selling a competing product or service built from it, and each version becomes Apache-2.0 two years after its release. You pay only for your AI tools' usage, and the token ledger shows where it goes.
Download it, add a repository, and tell the lead what you want.