CompozyOS
A local-first operating system for AI agents, enabling them to collaborate, persist, and share memory.
- Source repo
- compozy/compozy
- Stars
- ★ 2.8k
- Last updated
- today
- License
- MIT
- Primary language
- Go
- FA score
- 52/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformChatGPT · Codex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Developers wanting to use multiple agents (Claude Code, Codex) on the same project, coordinating tasks and sharing context.
- Not a fit if
- Teams on Windows: only macOS and Linux are supported
- Production users: v0.3 is beta and v0.2.15 is deprecated, migration required
- Users wanting zero learning curve: new commands and config.toml differ from v0.2
- Source review
- 52/100 · Major gaps
What does this agent do, and when should you use it?
CompozyOS is a local-first operating system for AI agents that manages sessions, tasks, loops, memory, permissions, automation, and tools through a single daemon. It offers multiple control surfaces (Web shell, CLI, HTTP/SSE, UDS, MCP, native tools) so humans and agents interact with the same runtime state. Key features include durable work that survives terminal closure, unified state across interfaces, human oversight with approvals and permissions, local-first SQLite storage, an extension system, and the Compozy Network protocol. The v0.3 beta is currently in development, with the previous v0.2.15 deprecated.
CompozyOS starts a daemon with compozy daemon start, which resolves workspaces, applies permissions and runtime policy, coordinates ACP-compatible agent CLIs (Claude Code, Codex, Gemini CLI, Cursor), and persists events and resource state. Users create sessions via compozy session new --agent general --name first-run, manage resources with commands like compozy agent list, compozy skill list, and compozy extension list. The extension system allows creating tool providers via compozy extension init and invoking them with compozy tool invoke. It supports task schemas (Task Schema v2) and loops, as well as workflow memory for agents to inspect and propose memory updates.
- Developers wanting to use multiple agents (Claude Code, Codex) on the same project, coordinating tasks and sharing context.
- Users needing long-running automated jobs that continue even after closing the terminal.
- Teams that want a persistent project memory shared across sessions and agents.
- Integrators looking to manage multiple agents through a unified HTTP/SSE or MCP interface.
- Organizations requiring audit trails, approvals, and permission checks for agent activities.
How do you install or deploy this agent?
For the v0.3 beta, use the official installer (macOS/Linux): curl -fsSL https://compozy.com/install.sh | sh, or install via npm: npm install -g @compozy/cli@beta, or via Go with a specific tag: go install github.com/compozy/compozy@<release-tag>. You can also build from source: git clone https://github.com/compozy/compozy.git && cd compozy && go build -o ./bin/compozy ..
How do you use this agent?
After installation, run compozy install to initialize, then compozy daemon start. In your repository directory, run compozy session new --agent general --name first-run to create a session. Use compozy --help for commands, and compozy doctor -o json to check the environment. Configuration can be viewed with compozy config show -o json.
What are this agent's strengths and limitations?
- Supports multiple agent CLIs via ACP, enabling multi-agent collaboration.
- Local-first, storing data in SQLite without cloud dependencies.
- Multiple control surfaces (Web, CLI, HTTP/SSE, UDS, MCP) for flexible integration.
- Extensible via a code-first extension system with SDKs for Go and npm.
- Human oversight features (approvals, permissions) enhance security.
- v0.3 is in beta and may be unstable; migration from v0.2 is required.
- New commands and configuration (config.toml) differ from v0.2, requiring learning effort.
- Only supports macOS and Linux, no Windows.
- Compozy Network features rely on network, adding potential external dependencies.
- Documentation and examples are limited; extension development may have a learning curve.
How does this agent compare with similar options?
Compared to using individual agent CLIs directly, CompozyOS provides multi-agent orchestration and persistent state. Compared to cloud platforms like OpenAI's console, it emphasizes local control and privacy.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| CompozyOS This agent | 52 · Major gaps | CLIFree + model costs | ★ 2.8k | today | Go | ChatGPT · Codex · Claude Code |
| AGNT Agent OS | 52 · Major gaps | Desktop appFree + model costs | ★ 554 | today | JavaScript | Codex · Claude Code · OpenAI API · Claude API |
| AI DevOps Framework (aidevops) | 65 · Some gaps | CLIFree + model costs | ★ 411 | today | Shell | OpenAI API · Claude API |
| EpicStaff Agent Workflow Platform | 57 · Major gaps | Self-hosted service | ★ 359 | 10d ago | Python | — |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows permission model (workspace boundaries, permission settings), user confirmation (approvals), data flow transparency (inspectable activity), sensitive data handling (secret redaction, token protection), external effects (network protocol, extensions). But dependency security is partial (Sigstore verification present, but no dependency vulnerability scanning evidence), rollback is partial (migration guide present, but no explicit rollback mechanism), source attribution is partial (publisher unverified).
Self-consistency is good (docs align with code), dependency availability is partial (CI present but no dependency availability guarantees), failure messages are partial (doctor command present, but no specific failure message examples).
Audience and scenarios are clear (developers, agents), capability boundaries are clear (extensions, skills, marketplace), trigger precision is partial (CLI commands present but no detailed trigger conditions), environment fit is good (local-first, multi-platform).
Information architecture is good (clear documentation structure), install notes are detailed (multiple installation methods), naming stability is partial (v0.3 beta, names may change), examples and FAQ are present (quick start, guides), known limitations are explicit (beta warning), license is clear (MIT), versioning changelog is partial (migration guide present but no detailed changelog), maintenance responsibility is partial (contributing guide present but maintainers not explicit).
Output usability is good (structured output, CLI reference), marginal value is high (unique features), cost-benefit is partial (no performance or resource usage data).
Claim traceability is partial (docs present but no specific implementation references), cross-source corroboration is partial (CI and tests present but no independent verification), fact-inference separation is partial (docs distinguish facts and inferences, but not explicitly labeled).
- Publisher identity is unverified; proceed with caution.
- v0.3 is beta; may be unstable.
- Dependency security is partial; check for vulnerabilities.
- Rollback mechanism is unclear; backup before upgrading.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
0.3.0-beta.35 - 2026-10-18 ♻️ Refactoring - Remove 12 long-tail builtin providers behind a one-release compatibility shim (#778) ⚡ Performance Improvements - Reduce desktop startup and live transcript overhead (#784) - Cut daemon idle cost and keep the host responsive under agent load (#797) - Reduce concurrent session query and projection overhead (#795) 🎉 Features - Add session folders, manual drag order, and session context menus (#786) - Give session titles their own line in the session list (#790) - Adding antigravity acp (#773) …View on GitHub ↗