Prometeu
Run Claude Code, Codex, Cursor CLI and Antigravity side by side in one desktop app, each task in its own git worktree with built-in diff review.
- Source repo
- prometeucorp/prometeu
- Stars
- ★ 65
- Last updated
- today
- License
- MIT
- Primary language
- Rust
- FA score
- 57/100 · Major gaps
At a glance
- How it runs
- Works with
- Portable with changesCodex · Claude CodeChatGPT (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A solo developer handing several refactors to different agents at once and wanting each to land in its own worktree before reviewing and committing one by one.
- Not a fit if
- Windows users who need Claude Code or Cursor CLI; Windows currently supports Codex only
- Teams treating worktrees as security sandboxes or requiring per-tool approval by default
- Source review
- 57/100 · Major gaps
What does this agent do, and when should you use it?
Prometeu is a free, MIT-licensed Tauri desktop app that gathers several coding-agent CLIs into a single window. It drives the claude, codex, cursor-agent or agy binaries you already installed and authenticated locally, spawning each task in an automatically created git worktree so parallel agents do not edit the same checkout. A visual Git panel handles unified or split diffs, review notes, staging, commits, history and conflict resolution, while built-in editor, terminal, browser preview, MCP servers, plugins and skills round out the workspace. A Tasks view filters Linear and GitHub issues and pull requests to start workspaces, conversations persist and can be resumed, and local token and cost history is visible under Settings / Analytics. macOS on Apple Silicon and Linux x86_64 are the primary targets; Windows x64 is experimental and Codex-only through WSL. The software is free, but model usage is billed by whichever provider account you bring.
Prometeu reads a local Git repository or plain folder, creates a workspace and a fresh worktree for the task, then launches the chosen agent CLI (claude, codex, cursor-agent or agy) as a process, streaming back output, tool calls, clarifying questions and plan approvals. A desk view lets conversations sit side by side; you can send follow-ups, stop an agent and resume it later. The Git view reads changes in the worktree, renders unified or split diffs, and supports review notes, staging, commits, history browsing and merge-conflict resolution. Settings attach MCP servers and plugins/skills, the sidebar opens standalone terminal sessions kept across restarts, and the Tasks view filters Linear and GitHub issues and pull requests to launch workspaces and track PR status. Settings / Analytics filters local consumption by period and workspace, showing token composition and explicitly flagging provider-reported costs and missing measurements. Optional cloud collaboration shares end-to-end encrypted conversations and lets companion devices continue them while execution stays on the owner's Mac.
- A solo developer handing several refactors to different agents at once and wanting each to land in its own worktree before reviewing and committing one by one.
- An engineer who normally uses Claude Code but wants to run the same task through Codex for comparison without juggling terminal windows.
- A team that pulls tasks straight from Linear or GitHub issues, runs an agent on them and then checks PR status in the same place.
- A reviewer who insists on a visual diff with notes before accepting agent output rather than trusting raw terminal scrollback.
- Developers on macOS and Linux who want one multi-agent workflow with an English and Portuguese interface.
- Someone who wants to share a conversation with a colleague on another device while keeping code execution on their own machine.
How do you install or deploy this agent?
macOS on Apple Silicon: download the dmg, open it and drag Prometeu to Applications; in-app updates are available.
# macOS: download Prometeu_aarch64.dmg from Releases, then drag it to ApplicationsLinux x86_64: download the AppImage into a directory your user can write to, make it executable and run it.
chmod +x Prometeu_x86_64.AppImage
./Prometeu_x86_64.AppImageWindows x64 (experimental): install WSL with a default distribution containing Git and an authenticated Codex CLI, run the installer, then start Prometeu from the Start menu. The installer is unsigned and updates require installing the next version manually.
# Windows: prepare WSL first, then run Prometeu_x64-setup.exeOn every platform you need at least one agent CLI installed and signed in on your shell: Claude Code, Codex, Cursor CLI (via agent login) or Antigravity CLI.
How do you use this agent?
Connect an agent first: make sure claude, codex, cursor-agent or agy is available and authenticated in your shell; Cursor and Antigravity use their own CLI login. Then add a project by opening a local Git repository or folder — no Prometeu account is needed for local work. For the first task, create a workspace, pick a model, keep the new-worktree option enabled for isolation and send the task; open a second workspace to work in parallel, then review changes in the Git view before committing.
# Sanity-check that the agent CLIs are reachable and signed in
claude --version
codex --version
cursor-agent login
agy --version # requires 1.2.7 or newerNote that agents execute tools without per-tool approval by default with your local user's file and process access; worktrees separate changes but are not sandboxes. Usage analytics are opt-in and off by default since v0.34.0 and can be toggled under Settings → General → Privacy → Share usage data.
What are this agent's strengths and limitations?
- Makes one-git-worktree-per-task the default flow, so parallel agents never collide on the same checkout.
- Brings four different agent CLIs — Claude Code, Codex, Cursor CLI and Antigravity — into one shell while keeping each vendor's own login and quotas.
- Ships a real Git review loop in-app: unified/split diffs, review notes, staging, commits, history and conflict resolution.
- MIT-licensed with no mandatory account for local work, and analytics are opt-in with prompts, transcripts and paths explicitly excluded.
- Built on Tauri for macOS and Linux with an AppImage and in-app updates, plus experimental Codex support on Windows via WSL.
- Ships no model of its own: you must install and authenticate at least one agent CLI, and model usage is billed by that provider.
- Windows is Codex-only through WSL, with an unsigned installer, no in-app updates and no managed login or cloud sync.
- Agents run tools without per-tool approval by default and worktrees are not sandboxes, which is risky for untrusted projects.
- The Cursor CLI path lacks hub plugins, skills and per-turn token/cost meters, so feature coverage differs from Claude Code and Codex.
- Optional sharing uses trust-on-first-contact keys with automatic adoption of changed keys and, per the docs, no forward secrecy.
How does this agent compare with similar options?
The README's explicit comparison target is the plain CLI workflow: juggling terminals, creating worktrees by hand and reading diffs as terminal text. Prometeu's pitch is folding those three into one desktop app; no competing GUI product is named in the repository.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Prometeu This agent | 57 · Major gaps | Desktop appFree + model costs | ★ 65 | today | Rust | Codex · Claude Code |
| Agentbox | 51 · Major gaps | Desktop appFree + model costs | ★ 362 | today | JavaScript | Codex · Claude Code |
| Skiff | 40 · Major gaps | Desktop appFree + model costs | ★ 19 | today | Rust | Codex · Claude Code |
| Parallel Code | 53 · Major gaps | Desktop appFree + model costs | ★ 1.1k | today | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
README and SECURITY.md honestly state agents run with the local user's file/process permissions, worktrees are not sandboxes, and tools execute without per-tool approval by default; this transparency is good but means least privilege and user confirmation are weak (1 each). Data-flow transparency is solid: opt-in PostHog telemetry off by default with content exclusion, ~/.prometeu storage, E2EE TOFU and no forward secrecy documented (2). Sensitive-data handling has clear boundaries but no key-management detail (2). Dependency security: CI pins action SHAs, uses --locked and npm ci, but package.json uses caret ranges rather than exact pins (2). External effects: agents can run arbitrary tools, write files and spawn subprocesses with only worktree isolation and no approval gate or sandbox (1). Rollback: worktree isolation, diff review before commit, updater accepts only newer versions (2). Source attribution: MIT copyright present, but publisher identity is unverified; signing/verification exists in release.yml (1).
Self-consistency is good: README, SECURITY.md, package.json version 0.39.0 and CI/release flows agree, and release verifies tag matches package.json version (2). Dependency availability: engines pin Node 24.14.0 and npm 11.9.0, CI builds on macOS/Linux/Windows, all deps are public (2). Failure messages: CI scripts have explicit error messages, but runtime error handling is barely visible in the provided files (1).
Audience and scenarios are clear: parallel agents, multi-repo, diff review, task integration (2). Capability boundaries are well documented: provider matrix, platform limits, Windows Codex-only, Cursor feature gaps (2). Trigger precision: as a desktop app there is no explicit trigger mechanism; agent launch is manual selection with no described control over automatic triggering or privilege escalation (1). Environment fit: macOS/Linux/Windows(WSL) install notes and CI coverage (2).
Information architecture is good: clear README sections and rich docs references (2). Install notes are detailed with three-platform steps and Cursor CLI setup (2). Naming stability: package name, version and tag rules are consistent (2). Examples and FAQ: quick start and demo exist, but no standalone FAQ and the demo uses simulated data (1). Known limitations are well disclosed: Windows experimental, unsigned installer, E2EE limits, worktree not a sandbox (2). License: full MIT text (3). Versioning/changelog: CHANGELOG.md referenced and release flow enforces it (2). Maintenance responsibility: SECURITY.md says only latest release is maintained with no guaranteed response deadline, and publisher identity is unverified (1).
Output usability: diff review, staging, commits, conflict resolution, terminal and editor integration serve real workflows (2). Marginal value: parallel worktree management and visual review over plain CLI is a clear gain (2). Cost-benefit: free MIT, bring-your-own provider account, optional telemetry, transparent costs (2).
Claim traceability: README claims point to docs contracts and ADRs (2). Cross-source corroboration: CI workflows, package.json scripts and README descriptions corroborate each other (2). Fact/inference separation: README explicitly labels the demo as simulated data and defers provider differences to a matrix (2).
- Agents execute tools with the local user's permissions and no per-tool approval by default; worktrees isolate Git changes but are not security sandboxes—use only on trusted projects.
- Publisher identity is unverified and the Windows installer is unsigned; verify signatures and provenance before downloading or updating.
- Shared conversations use TOFU keys with automatic adoption of changed keys and no forward secrecy; be cautious sharing sensitive content.
- package.json dependencies use caret ranges, so build reproducibility depends on the lockfile; verify it.