Ateam Agent Workbench
Coordinate parallel coding agents in isolated Git worktrees from one desktop board.
Per-dimension scores and reasoning
The repository shows per-task git worktree isolation, working-directory-scoped git operations, abort-on-divergence behavior, private Tailscale deployment guidance, and constrained CI permissions. It also describes material effects such as commits, pushes, merges, VPS creation, and SSH-key generation, supporting moderate scores for least privilege, transparency, and recovery. Deductions apply because Auto/YOLO bypass modes, root-level setup, and curl-to-shell installation are present, while the supplied files do not establish confirmation gates for every high-impact action. Privacy is only summarized in the README; the referenced PRIVACY.md and implementation evidence for credential storage, log redaction, key lifecycle, or third-party agent data handling are absent. Dependencies use a specified Bun version, frozen-lockfile CI, and some exact versions, but no vulnerability audit, artifact checksum, dependency review, or installer integrity control is shown. Copyright and licensing identify Clawnify, although publisher identity remains unverified and unknown.
The README, tests, and workflows are broadly consistent about supported agents, session restoration, worktree isolation, and build commands. Tests cover shell quoting, failed or shifted CLI output, and graceful fallback from model-generated tagging; the installer is described as producing a readiness report, and release automation verifies artifact contents and upload presence. Deductions reflect partial rather than end-to-end coverage: the supplied evidence does not substantiate the complete desktop process, remote connection, VPS provisioning, or full task lifecycle. CI explicitly omits lint, and release comments document prior missing server assets. No runtime execution was performed.
The sources thoroughly identify users and scenarios across desktop, mobile, local and remote execution, multiple agent CLIs, and self-managed or provider-created hosts. Requirements, Status, Roadmap, and agent-specific session constraints establish useful capability boundaries, while command tests demonstrate agent-specific invocation behavior. Deductions apply because several boundaries depend on changing external CLI behavior, Auto mode lacks a detailed risk and suitability policy, and environment guidance is concentrated on Apple Silicon macOS, Ubuntu, Bun 1.3, and named CLIs without broad architecture, distribution, offline, or compatibility coverage.
The README has strong information architecture spanning installation, development, testing, remote deployment, the git safety model, current status, and roadmap. Installation guidance is unusually concrete, and the complete GPL-3.0-or-later text plus dual-licensing statement justifies full license credit. Deductions apply because there is no dedicated FAQ, formal known-issues inventory, compatibility matrix, or migration guide. Releases and isolated version incident notes exist, but no systematic changelog is shown. Clawnify is named as copyright holder and commercial-license contact, yet maintainers, support expectations, security reporting, and sustained update responsibility are not defined, and publisher provenance is unverified.
The kanban board, changes view, parallel terminals, session search, remote execution, and integrated git workflow provide directly usable outputs. Coordinating established coding agents in isolated worktrees offers clear marginal value over manually managing terminals, branches, and worktrees. The documentation provides a rough one-gigabyte-per-session estimate and example VPS sizing and says planned integrations have no paywall. Deductions apply because agent subscriptions, hosting, bandwidth, storage, concurrency, and maintenance costs are not comprehensively quantified, and static evidence cannot establish the actual productivity gain.
Major claims are traceable to concrete README sections and receive partial corroboration from command, session, and tagging tests plus CI and release workflows. Status and Roadmap clearly separate implemented features from planned work, while tests distinguish unknown state, no result, and failure, supporting strong fact-versus-inference separation. Deductions apply because stronger claims such as local-only operation, absence of analytics, signing/notarization, comprehensive git safety, and desktop boot verification are asserted without the referenced privacy file, core implementations, complete test output, or release attestations in the supplied material. This assessment did not execute the software.
- Auto/YOLO mode passes permission-bypass flags to agents; verify each agent's effective permissions and use isolated accounts and recoverable branches before enabling it.
- The documentation recommends remote curl-to-shell installers and includes root-level steps; pin an audited revision, inspect downloaded content, and avoid directly executing a mutable main-branch script.
- Ateam can commit, push, merge, create VPS instances, generate SSH keys, and inspect several agents' session records. Do not assume adequate confirmation, secret storage, or log redaction without reviewing the missing implementation evidence.
- CI omits lint, and workflow comments record previous missing server release assets; independently verify the lockfile, release artifacts, and remote installation path for the target revision.
- Open-source use is governed by GPL-3.0-or-later; organizations unable to meet its obligations require a separate commercial license.
What does this agent do, and when should you use it?
Ateam is a local-first desktop application for orchestrating Claude Code, OpenCode, and Codex coding sessions in parallel. It creates a dedicated Git branch and worktree for each task, then organizes the work across a kanban board with states including Backlog, In Progress, Needs You, Review, and Done. Its repository contains an Electron and React desktop client, a local SQLite layer, a safe Git engine, an agent registry, a transport-independent RPC protocol, a headless server, and an Expo/React Native iOS client. Users can run multiple PTY terminals per task, inspect file lists and side-by-side diffs, and perform commit, push, update, and GitHub PR merge operations. Execution can remain on the Mac or move to a Linux host reached over SSH or WebSocket, with the desktop or phone acting as the control surface.
A user registers a project, optionally initializes Git for a plain folder, and creates a task by selecting Claude Code, OpenCode, or Codex and entering the first instruction. Ateam creates the task worktree at <repo>/.ateam/worktrees/<slug>, assigns it a dedicated branch, and launches the chosen claude, opencode, or codex CLI in a PTY. Hook-derived status feeds the kanban board, while Mission Control presents multiple tasks and their sessions in a read-only grid. The Changes view aggregates diffstat, lists affected files, and renders GitHub-style side-by-side diffs before built-in commit, push, update, and merge actions. Merge uses gh pr merge, then updates local main either by directly fast-forwarding its ref when it is not checked out or by running merge --ff-only in main's own worktree; divergence causes an abort. The application also provides extra agent sessions and shell tabs, image paste and drag-and-drop into terminals, transcript-backed session search, and worktree cleanup informed by age, PR state, ahead count, and dirty state. packages/server supplies the headless engine and daemon used by desktop and iOS clients over SSH or WebSocket.
- A developer assigning separate features to Claude Code, OpenCode, and Codex can keep every task in its own worktree so the sessions do not contend over one checkout.
- A maintainer handling several fixes can track task state, inspect diffs, commit changes, push branches, and merge GitHub PRs from one board.
- An engineering lead supervising long-running coding sessions can use the Mission Control grid to watch multiple tasks and their terminals together.
- A Mac user who wants compilation, tests, and agent sessions to run elsewhere can place the engine on a Linux box and control it over SSH through Tailscale.
- A user who needs mobile access to remote development work can install the persistent WebSocket service and connect through the iOS client.
- A developer with accumulated task worktrees can evaluate cleanup candidates using merge state, idle age, ahead counts, and uncommitted changes.
What are this agent's strengths and limitations?
- Each task receives a dedicated branch and worktree, and mutations are scoped with git -C <worktree>, preventing one session from changing another session's checkout.
- PR merges happen remotely through gh pr merge, and local main is updated only through a safe fast-forward path that aborts on divergence.
- The desktop interface combines kanban status, multiple PTY sessions, GitHub-style diff review, Git operations, and worktree cleanup.
- It supports three distinct coding CLIs—Claude Code, OpenCode, and Codex—and can execute either locally on a Mac or on a remote Linux host.
- The Git engine and database layer are unit-tested, and the Electron main process is boot-verified with its native modules.
- The packaged desktop client targets macOS; source development also brings Bun 1.3, Electron native-module rebuilds, and host-architecture considerations.
- GitHub identity, PR merging, and other GitHub operations depend on an authenticated gh CLI, limiting suitability for workflows hosted entirely outside GitHub.
- Users must separately install and authenticate at least one supported coding CLI; Ateam does not provide the underlying model account or subscription.
- Remote and mobile operation requires a Linux host, SSH configuration, network connectivity, and the documented setup recommends Tailscale; the phone cannot start the daemon over SSH as the desktop does.
- Transcript-to-task generation, cross-worktree session continuity, and Linear, Slack, and GitHub Issues integrations through MCP are roadmap items rather than current capabilities.
- The dual-license model uses GPL-3.0-or-later for open-source use; organizations unable to comply need a commercial license from Clawnify.
How do you install or deploy this agent?
For the packaged application, download the signed and notarized DMG from https://github.com/clawnify/ateam/releases/latest/download/Ateam-macos.dmg, drag Ateam into Applications, and open it. Install Git 2.31 or later and gh, then confirm GitHub authentication with gh auth status. At least one of claude, opencode, or codex must be available on PATH.
To run from source:
brew install oven-sh/bun/bun
bun install
bun run --filter @ateam/desktop devSource development requires Bun 1.3 or later. bun install runs the apps/desktop postinstall hook that rebuilds better-sqlite3 and node-pty for Electron. On Apple Silicon, the lifecycle script must run through a native Arm64 shell.
To install the engine on a remote Linux host:
curl -fsSL https://raw.githubusercontent.com/clawnify/ateam/main/packages/server/scripts/install.sh | bashConfigure the host's Git user identity, authenticate GitHub with gh auth login, and install and sign in to at least one supported agent CLI. For persistent iOS access, set ATEAM_WS_ADDR to the host's own Tailscale IP on port 8787 and install the service:
export ATEAM_WS_ADDR=100.x.y.z:8787
curl -fsSL https://raw.githubusercontent.com/clawnify/ateam/main/packages/server/scripts/install.sh | bash -s -- --serviceHow do you use this agent?
Open Ateam and register a Git project; for a plain directory, the application can optionally initialize a repository. Create a task, choose Claude Code, OpenCode, or Codex, and enter the initial instruction. Ateam creates the task-specific branch and worktree and starts the selected CLI in a PTY session. Add further agent sessions or ordinary shell tabs as needed, and monitor progress through the board or Mission Control. Review the aggregate diffstat, file list, and side-by-side changes before using the built-in commit, push, update, and merge workflow; GitHub identity and operations come from the authenticated gh CLI. For remote execution, install the engine on the Linux host, configure SSH, add its alias to ~/.ssh/config, and select that host from Ateam's connection switcher.
How does this agent compare with similar options?
Compared with running Claude Code, OpenCode, or Codex independently in separate terminals, Ateam adds a shared task board, one-worktree-per-task isolation, centralized change review, and a guarded merge workflow. It does not replace those CLIs or act as a model provider; it coordinates them and can relocate their execution to a remote Linux host.