Up.computer Coding Workbench
Turn chat discussions into background coding tasks and orchestrate multiple agents from one desktop app.
- Source repo
- krl-gr/upcomputer
- Stars
- ★ 87
- Last updated
- 2d ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 44/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Developers who want to switch among installed agents such as Codex and Claude Code from one desktop interface.
- Not a fit if
- Teams that require a signed Windows installer
- Users who need an official Homebrew, winget, or AUR install path
- Users unwilling to install and authenticate a coding agent
- Source review
- 44/100 · Major gaps
What does this agent do, and when should you use it?
Up.computer is an open-source desktop control plane for coding agents, bringing Codex, Claude Code, Cursor, Grok Build, OpenCode, Antigravity, Pi, and agents from the ACP Registry into one app. Users discuss an idea in chat and create a task for a worker agent to handle in the background; each agent uses the user's own subscription, and chats stay on the computer. The app groups work into project threads, supports tasks and scheduled automations, and lets users review changes and diffs before committing, pushing, publishing repositories, or opening pull requests. It includes a local backend for desktop projects and can pair a browser or another desktop app over a local network or Tailscale. The project is a fork of T3 Code, with an Electron desktop shell, a Node.js server, and a React/Vite web interface; releases target macOS, Windows, and Linux.
After installing and authenticating at least one supported coding agent, users enable or add agents in Settings → Providers. They discuss work in chat and create tasks; saved task agents pick them up by status and tag, and tasks can also be created on a schedule. Agents work in project threads, while the app presents changed files and diffs and offers actions to commit, push, publish repositories, and open pull requests. It also gives agents a managed browser and native desktop observation and control, and supports pairing browsers or other desktop apps over a local network or Tailscale.
- Developers who want to switch among installed agents such as Codex and Claude Code from one desktop interface.
- Individuals or teams who want to turn development discussions into background tasks assigned to saved task agents by status or tag.
- Code maintainers who need to inspect changed files and diffs before accepting agent work.
- Contributors who want to commit, push, publish a repository, or open a pull request from the agent workflow.
- Developers who need agents to use a managed browser or want to pair another client over a local network or Tailscale.
How do you install or deploy this agent?
Download the app from up.computer/download or the latest GitHub release. Builds are available for macOS (Apple Silicon and Intel), Windows, and Linux. macOS builds are signed and notarized; the Windows installer is not yet signed and SmartScreen warns on first launch. Installed apps update themselves from GitHub releases. The app runs coding agents that you install, so install and authenticate whichever agents you plan to use. To run from source, the documented requirements are Vite+ vp, Node 24.13.1+, Git, and at least one supported coding agent installed and authenticated:
curl -fsSL https://vite.plus | bash
vp install
vp run devHow do you use this agent?
On first use, connect or enable agents in Settings → Providers. The README gives these login commands:
codex login
claude auth loginYou can also install OpenCode and run opencode auth login, or follow the relevant agent's setup instructions. Codex and Claude are enabled by default; Cursor, Grok, OpenCode, Antigravity, and Pi are disabled by default. Cursor can be connected in Settings → Providers → Cursor or configured with CURSOR_API_KEY. Once an agent is ready, describe the work in chat and create a task for a worker to pick up in the background. Review its file changes and diffs, then use the app's Git and publishing actions as needed.
What are this agent's strengths and limitations?
- Orchestrates multiple installed coding agents in one desktop app and supports adding agents from the ACP Registry.
- Saved task agents can pick up tasks by status and tag, and tasks can be created on a schedule.
- Includes change and diff review plus actions to commit, push, publish repositories, and open pull requests.
- Supports local projects and pairing other clients over a local network or Tailscale.
- You must install and authenticate at least one supported coding agent; each agent uses your own subscription.
- Running from source requires Vite+, Node 24.13.1+, Git, and a coding agent, so setup involves multiple steps.
- The Windows installer is not signed and triggers a SmartScreen warning on first launch.
- Homebrew, winget, and AUR are not official install paths yet; the mobile app is not published.
How does this agent compare with similar options?
Up.computer is a fork of T3 Code. It keeps the upstream runtime, provider orchestration, contracts, desktop infrastructure, and release plumbing while developing a different product direction, UI, branding, and workflow. The README does not provide a feature comparison with other products.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Up.computer Coding Workbench This agent | 44 · Major gaps | Desktop appFree + model costs | ★ 87 | 2d ago | TypeScript | Codex · Claude Code |
| Atlas | 50 · Major gaps | Desktop appFree + model costs | ★ 9.6k | today | Rust | Codex · Claude Code |
| stagewise | 25 · Major gaps | Desktop appFreemium | ★ 6.8k | 21d ago | TypeScript | — |
| Code Assistant | 70 · Some gaps | Desktop appFree + model costs | ★ 183 | today | Rust | Claude.ai · OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
permission-modes.md documents Supervised/Auto-accept/Auto/Full access modes, so a confirmation mechanism exists, but the initial default is Full access, which is deducted. environment-auth.md shows a fairly detailed scoped-session model (DPoP, PKCE, separate MCP audience), yet it explicitly admits filesystem:read reaches absolute paths outside a project and that signed asset URLs are bearer credentials that keep working after revocation until expiry, so least privilege scores 1. Sensitive data handling notes that values marked sensitive are not redisplayed after saving, but storage details are not shown, so 2. Dependency security rests only on package.json and CI with no audit/SBOM/lockfile policy evidence, so 1. External effects (commit, push, publish, open PR, browser and desktop control) are described without default limits or recovery, so 1. Rollback is limited to t3 uninstall, t3 doctor and update notes, with no mechanism to undo agent actions already taken externally, so 1. Source attribution is explicit: README and LICENSE state the T3 Code fork lineage and retain upstream copyright, so 2.
The docs contradict each other: README says package-manager installs are not official, while docs/user/install.md still offers brew/winget/AUR under the T3 Code name; install.md is titled and written as T3 Code, conflicting with the Up.computer branding, so self-consistency is deducted to 1. Dependency availability requires Node 24.13.1+, Vite+ installed via a curl pipe, pnpm 11, and no CLI on Intel Macs, so 1. Failure messages exist in places (PATH, SmartScreen, update failure) but there is no systematic error-code or diagnostics reference, so 1.
Audience and scenarios cover desktop, web, mobile, remote, WSL and background service, so 2. Capability boundaries are stated clearly (projects are not filesystem sandboxes, mobile not published, relay not deployed, package managers unofficial), so 2. Trigger precision: task agents pick up by status and tag and scheduled tasks exist, but exact trigger conditions and conflict handling are not specified, so 1. Environment fit covers macOS/Windows/Linux, Tailscale, SSH and WSL, so 2.
Information architecture has a README and docs tree, but install.md still uses the T3 Code name and links to t3.codes and the pingdotgg repo, conflicting with the README download path, so 1. Install notes exist but contradict each other (official download vs package managers), so 1. Naming stability is deliberately mixed: internal @t3tools/*, T3CODE_* and the t3 CLI are kept for upstream compatibility, which is intentional but leaves external naming inconsistent, so 1. Examples and FAQ are thin, limited to scattered commands, so 1. Known limitations are well covered (unsigned Windows installer, no Intel Mac CLI, mobile unpublished, relay undeployed), so 2. LICENSE is a complete MIT text with both copyright holders, so 2. No CHANGELOG or versioning policy is present, so 0. Maintenance responsibility has CONTRIBUTING and a security contact, but publisher identity is unverified, so 1.
Output usability is clear: a unified GUI for multiple coding agents, diff review, task orchestration and usage tracking, so 2. Marginal value is limited because this is a fork of T3 Code whose runtime and orchestration come from upstream, with the increment mainly branding and a workflow layer, so 1. Cost-benefit: free and open source with bring-your-own subscription is a plus, but setup requires several provider CLIs, Vite+ and Node 24, so 1.
Most README claims (signed and notarized builds, self-update, provider support) lack in-repo build or release evidence; only the CI workflows partially corroborate that tests and builds exist, so 1. Cross-source corroboration is limited: CI and package.json agree, but install.md and README contradict each other, weakening trust, so 1. Fact/inference separation is mostly present (supported vs unpublished vs undeployed), but claims such as every chat staying on your computer sit uneasily with remote access, the relay and Cursor cloud usage reads without being flagged, so 1.
- The default permission mode is Full access, letting agents run commands and edit files without confirmation; switch to Supervised for first use.
- docs/user/install.md still offers brew/winget/AUR and the t3.codes install script under the T3 Code name, conflicting with the README download path; verify install sources carefully.
- The docs state projects are not filesystem sandboxes, filesystem:read reaches absolute paths outside a project, and signed asset URLs keep working after revocation until expiry.
- Publisher identity is unverified and the repository is a fork of T3 Code; confirm how release builds relate to upstream.
- The repo vendors upstream reference code under .repos/alchemy-effect that is not part of this product and should not be counted in the assessment.
- The Windows installer is unsigned, so SmartScreen warns on first launch.
- There is no CHANGELOG or versioning policy; verify upgrade behavior against GitHub Releases yourself.