Cloudroom
Run coding agents on your Mac or in dedicated cloud sandboxes, where they can keep working after you close your laptop.
- Source repo
- davidondrej/cloudroom
- Stars
- ★ 1.4k
- Last updated
- 1d ago
- License
- Apache-2.0
- Primary language
- TypeScript
- FA score
- 49/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free tier plus a paid hosted plan
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Individual developers who want a coding agent to keep working after closing a Mac laptop can run its thread in a cloud sandbox.
- Not a fit if
- Teams requiring an entirely local, cloud-free workflow
- Teams that want to use the app without signing in
- Source review
- 49/100 · Major gaps
What does this agent do, and when should you use it?
Cloudroom is a desktop app for AI coding agents: start agents locally on your Mac or give each cloud thread its own sandbox and Git branch. The repository contains gui, a desktop app built on BB; core, a Rust service that runs agents and saves their history; and skills shipped with every Cloudroom agent. It works with tools including Claude Code, Codex, Pi, OpenCode, Hermes, Cursor, and Devin, and agents can use your own Claude and ChatGPT logins. Cloud agents can reach your Mac under configurable Off, Read-only, Ask first, or Full access settings, and Teleport moves a running thread between Mac and cloud. You can download the desktop app or self-host core on Ubuntu 24.04; supported desktop platforms are Apple Silicon Macs running macOS 13+, Linux x64, and Windows 11.
From the desktop app, users start a coding agent locally or create a cloud thread with its own sandbox and Git branch. Each cloud sandbox runs Cloudroom core alongside the agent, and the app talks directly to core. Agents use the user's Claude or ChatGPT login, while thread history is saved on the machine hosting it. Users can set the cloud agent's Mac access to Off, Read-only, Ask first, or Full, and use Teleport to move a running thread, including its chat, model, and effort, between Mac and cloud. The repository also documents core's HTTP API, storage, harnesses, session lifecycle, diagnostics, and dashboard API.
- Individual developers who want a coding agent to keep working after closing a Mac laptop can run its thread in a cloud sandbox.
- Teams separating concurrent agent tasks can use the dedicated sandbox and Git branch created for each Cloud thread.
- Developers who want cloud agents to reach files on their Mac can choose Off, Read-only, Ask first, or Full access.
- Developers already using Claude or ChatGPT can let agents use their own service logins.
- Teams that want control of the agent runtime and history storage can deploy core on their own Ubuntu 24.04 Linux machine.
- Developers moving long-running work between local and cloud environments can use Teleport to move the active thread.
How do you install or deploy this agent?
Download the desktop app from the download page; the README does not provide a command-line installation procedure for it. The documented self-hosted core installation targets Ubuntu 24.04:
How do you use this agent?
After installing self-hosted core, connect the desktop app with your server URL and token. The README does not give a command to generate either value. Building core from source requires Rust 1.89+:
What are this agent's strengths and limitations?
- Each cloud thread gets its own sandbox and Git branch, which isolates agent work.
- Teleport moves a running thread between Mac and cloud with the same chat, model, and effort.
- The Rust core can be self-hosted without a Cloudroom account.
- Mac access for cloud agents has four documented settings, from Off to Full.
- The documented self-host install targets Ubuntu 24.04; other Linux distributions are not specified here.
- The README does not state hosted Cloudroom pricing or the limits of its free offering.
- Agents run without approval prompts; the README recommends a separate machine for each trusted user.
- Connecting the desktop app to a self-hosted server requires a server URL and token, but this README excerpt does not explain how to obtain them.
How does this agent compare with similar options?
Cloudroom supports both local and cloud agent runs and moving threads between them; the README does not compare it with other products on features or price.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Cloudroom This agent | 49 · Major gaps | Desktop appFreemium | ★ 1.4k | 1d ago | TypeScript | Codex · Claude Code |
| Codync | 51 · Major gaps | Desktop appFree | ★ 1.9k | today | Rust | ChatGPT · Codex · Claude Code · Claude.ai |
| Rovai AI | 49 · Major gaps | — | ★ 247 | today | Rust | Codex · Claude Code |
| amux Agent Control Plane | 77 · Good | Self-hosted serviceFree + model costs | ★ 554 | 1d ago | Rust | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
README states agents run as a separate Linux user, API requires a secret token, and listens only on localhost; tests confirm non-loopback binding needs explicit permission and token auth, so least_privilege scores 2. However README says agents run without approval prompts, so user_confirmation is 1. Data flows (subscription logins, Mac access, cloud sandbox) are only high-level, data_flow_transparency 1. Tests show token/key redaction, 0600 file modes, and log scrubbing, sensitive_data_handling 2. Dependency security only shows --locked and pinned action SHAs, no audit/vuln scanning, dependency_security 1. external_effects and rollback lack substantive coverage, both 1. Apache-2.0 with BB MIT NOTICE gives clear attribution, source_attribution 2.
Tests cover idempotent requests, restart recovery, SSE replay, graceful shutdown, self_consistency 2. Dependency availability only cites Rust 1.89+ and an install script, no locked dependency manifest evidence, dependency_availability 1. Error messages, diagnostic IDs, redaction and failure states appear in tests, failure_messages 2.
README names macOS/Linux/Windows and local/cloud scenarios, audience_and_scenarios 2. capability_boundaries only scattered mentions, 1. trigger_precision has no agent trigger/routing mechanism, 1. environment_fit has platform and self-host notes, 2.
README is well-structured with links to multiple docs, information_architecture 2. Install steps (download, self-host script, source build) are explicit, install_notes 2. naming_stability has no version/naming policy evidence, 1. examples_and_faq only has example commands, no FAQ, 1. known_limitations only one security sentence about no approval prompts, 1. LICENSE is full Apache-2.0, license 3. versioning_changelog only external changelog link, 1. maintenance_responsibility: publisher unverified, no maintenance commitment, 1.
Output usability backed by API/session/diagnostics docs, output_usability 2. marginal_value differentiates via local+cloud+teleport, 2. cost_benefit has no cost/resource quantification, 1.
claim_traceability only README assertions without per-claim citations, 1. cross_source_corroboration tests partially echo README but no systematic cross-check, 1. fact_inference_separation docs do not separate fact from inference, 1.
- README states agents run without approval prompts; isolate trust boundaries before production use.
- Publisher identity is unverified; maintenance responsibility and update path are unclear.
- Dependency security only shows locked builds and pinned action SHAs, no vulnerability scanning evidence.
- Data flows and rollback are undocumented; external effects of cloud/local switching are unquantified.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
gui-v141 Cloudroom v141 (macOS Apple Silicon, Linux x64, Windows x64)
Release notes excerpt
Cloudroom v141 Runaway threads can't sink your Mac. For Apple Silicon Macs on macOS 13 or later, Linux x64, and Windows x64. Changes - If one thread eats more than half your Mac's memory, Cloudroom stops it before your Mac freezes. - After a crash or reboot, a thread that was running no longer restarts on its own. - Moving a big chat to Cloud is much faster, and the progress shows megabytes and time left. - A Teleport that times out now retries by itself instead of waiting behind Retry. - Sending the same message twice to an idle thread no longer starts a second turn. …View on GitHub ↗ -
core-v0.1.96 Core 0.1.96
Release notes excerpt
- Moving a big chat to Cloud is much faster: Core no longer copies it line by line. - Cursor threads no longer fill the sandbox disk with chat copies. - Codex starts more reliably on a new sandbox, and says why if it can't list models. - Mac access stops jobs from an old account after a failed sign-out. Linux builds of Cloudroom core at 0442e1b1d7f1, tested by CI. VMs verify them by SHA-256. Want hosted Cloudroom instead? [Get started](https://www.cloudroom.dev).View on GitHub ↗ -
gui-v140 Cloudroom v140 (macOS Apple Silicon, Linux x64, Windows x64)
Release notes excerpt
Cloudroom v140 Crashed Cloud agents come back. For Apple Silicon Macs on macOS 13 or later, Linux x64, and Windows x64. Changes - A Cloud agent that crashed now restarts as soon as you message it. - Messages you send to a stopped Cloud thread are saved first, so none get lost while its sandbox wakes. - Cloud subagents show as starting the moment you spawn them, and can message their parent and siblings. - An idle cloud browser now pauses until an agent needs it again. - Setting up a worktree again never wipes your uncommitted work. …View on GitHub ↗