agent-fleet

Agent Fleet is a self-hosted console for running AI coding agents as a fleet. Its pitch is that you can close your laptop and the agents keep working: a team shares agents from one browser console, and members check progress or send the next instruction from Discord, Slack or a phone instead of sitting in front of a terminal. Supported agent CLIs named in the README are Claude Code, Codex CLI, GitHub Copilot CLI, Antigravity CLI, Cursor CLI, Kiro and OpenCode. Each team member gets an isolated per-user environment — a Docker container with cgroup CPU and memory quotas, or a bubblewrap sandbox in the Docker-less native edition — with a persistent home directory and its own git working copies. Sessions are started, followed and steered from the browser. Because it is self-hosted, one company runs one deployment on its own infrastructure, so credentials, source code and conversations stay inside that organization. The same core runs both on a single Linux host with Docker Compose and on AWS ECS. Console capabilities described in the repository: - A launch dialog that selects the agent CLI, model, reasoning effort, start mode, and whether to run in a fresh git worktree or an existing working copy. - A chat mirror in which the agent's questions, plans and permission prompts appear as cards that can be answered in place. - Split panes: the chat mirror, a live terminal attached to a shell session via a PTY, and working-tree changes side by side; panes can pop out into their own tabs. - Git tooling in the console: a commit graph beside the diff of the selected commit, plus staging and committing, per working copy and per worktree. - A usage tab with a stacked per-feature token chart over 24h / 7d / 30d, KPI tiles for tokens, calls, cache reads, API-equivalent cost and unmeasured calls, and breakdowns by feature, agent and model. Calls that report no tokens are counted separately rather than as zero. The interface is available in English or Japanese, switched per user in Settings; screenshots exist for both. Deployment options are documented as a twenty-minute decision: compose (the default, a team on one Linux host with Docker), native (no Docker, single user, e.g. WSL2 or a personal Linux box), ecs / ecs-ec2 (AWS with task-level isolation), and ec2-single (AWS, small team, Compose on one VM). Released bundles pull pinned images from GHCR and are published to a separate distribution repository; command procedures live alongside what they operate (deploy/compose, deploy/native, deploy/aws/ecs). Building from the tree involves copying .env.example, generating secrets such as AF_MASTER_KEY, building the per-user workspace image, and running docker compose. Caddy handles TLS via Let's Encrypt, and sign-in uses the control plane's own OAuth. The documentation is split by reader: guide/ explains how to do things and ships into every workspace container, opened from the Console's "User guide" entry; docs/ explains how the code works and why. The source is organized into workspace/ (the agent and its image), control-plane/ and console/, with a local development stack started through deploy/local/run-dev.sh (local = Docker, wsl, native, reset). The README also fixes terminology: a Workspace is the persistent container environment for one user with a home volume and running processes; a Working copy is the working directory of a git repository cloned inside a Workspace; a Session is the logical unit of a conversation, its settings and execution state, tied to a working copy. Sessions do not imply a terminal — Codex, OpenCode, Copilot, Cursor and Kiro default to a managed execution method driven from the chat view (Codex and OpenCode run on a shared runtime with no per-session CLI process), while Claude, Antigravity and plain shell / SSM sessions use a terminal. The project is licensed under Apache License 2.0, and the README notes that publishing the source of a credential-handling tool so each company can audit the crypto and isolation implementation is part of the pitch. Contribution guidance is in CONTRIBUTING.md, and vulnerability reports and the threat model are in SECURITY.md.