Compozy

CompozyOS is a complete runtime environment built around AI agents, with the goal of building in advance the system itself that agents need to work continuously, so that users no longer have to write their own scripts, cron jobs, and glue code. It is described as "a complete environment for creating, automating, and supervising agent work, without scripts, plugin chains, or orchestration frameworks."

Purpose and Positioning

CompozyOS is aimed at scenarios that require AI agents to execute tasks continuously and reliably. Traditional approaches usually involve assembling loops, triggers, cron, permissions, approvals, and observability yourself, and maintaining glue code between them; CompozyOS makes these core objects within a single runtime, letting humans and agents use the same primitives.

Main Capabilities

1. Autonomy kernel: tasks atomically claim runs via ClaimNextRun, hold a lease and must send heartbeats. If a crash occurs, the lease expires according to its TTL and the run automatically returns to the queue, to be taken over by the next idle agent. Only an agent holding the claim token can send heartbeats or complete the run; sessions cannot access runs that do not belong to them. Tokens are hashed before entering the event ledger, and the original value never leaves the daemon, so there is no duplicate execution.

2. Shared queue: both operators and agents write to the same task_runs SQLite table.compozy task create(performed by humans) and coordinator agents (performed by agents) use the same primitives and the same audit trail.

3. Permission narrowing: a child cannot expand the parent's permissions. Lineage, TTL, and permission scope are part of the generated contract and are enforced in code, not written in prompts.

4. Persistent sessions: agent work can outlive the terminal lifecycle. Persistent sessions retain the timeline, edits, and runs; they can be paused and resumed, and the context is still there.

5. Observability: every run leaves a trace. Traces, event counts, and tool usage all come from the same records written by the runtime, with no additional instrumentation needed.

6. Automation: persistent scheduling supporting Cron and Webhooks, allowing work to be scheduled periodically and sessions to be triggered from external events, with every run being persistent and inspectable.

7. Extensibility: the daemon is extensible at every seam where it is needed, with no need to write plugins, and the contract is just ordinary files. This includes: typed dispatch lifecycle hooks (sessions, prompts, tools, permissions, autonomy; hooks can reject or narrow but cannot bypass); drop-in SKILL.md skill packages (Markdown with YAML frontmatter, supporting built-in defaults, global and workspace scopes); Cron, Webhook, and event triggers; extensions can be installed locally or from a marketplace, packaging skills, hooks, tools, and MCP servers, distributed as zip files or via a GitHub registry. Everything the daemon loads can be compozy skill view、compozy hooks list and compozy extension list inspected.

Usage

CompozyOS starts built-in CLI and provider integrations as persistent sessions, sharing tools, explicit permissions, with history retained on the local machine. It is currently in beta, and installation methods include: a verified installer on macOS or Linux, the npm beta channel, or a pinned-version Go distribution; Homebrew will return with the stable release. The workflow is roughly: start the daemon (a local process, detached by default, exposing CLI, HTTP/SSE, and Web UI); run it in the repository you want to manage, and CompozyOS will infer and register the workspace from the current directory; then start a real persistent session. The documentation provides an install-to-run quick start, a CLI reference, and architectural explanations of hooks, skills, automation, and extensions.

Use Cases

Suitable for developers and technical users who want agents to keep working without maintaining glue logic such as retry scripts, monitoring loops, temporary guardrails, triggers and scheduling, permission wrappers and approval prompts, log scraping, and ad hoc dashboards. It runs the agents users are already using, and the text mentions Claude Code, OpenClaw, and Hermes, among others.

Limitations

The current experience is fairly technical: it requires a terminal, a local daemon, and a configuration file. The official documentation explicitly states that this is in the beta stage and that Homebrew will have to wait for the stable release. Specific pricing, costs, a full platform list, performance results, or usage steps not mentioned in the materials are not provided.