mrkeyoor.com_
Thu 13 Aug 15:38 UTC
LLM Toolsevaluationupdated 13 Aug 2026

compozy

CompozyOS is a local control system for running AI coding agents as durable sessions, scheduled loops, and cooperating workers instead of disposable terminal commands. It gives Claude Code, OpenClaw, Hermes, and other compatible agent tools shared state, permissions, memory, approvals, automation, and a browser interface.

Verdict

CompozyOS is one of the more thoughtful attempts to turn coding agents into supervised, durable infrastructure, especially for developers already juggling several CLIs. Its daemon-owned state, explicit permissions, reusable resources, and multiple control surfaces solve real problems. Start with a noncritical repository and expect migration churn until v0.3 is stable; this is promising operational software, not yet a boring appliance.

Setup3/5One binary starts quickly, but agents and policy need real setup
Docs5/5Detailed guides cover runtime, migration, CLI, extensions, and policy
Community4/5Fast beta releases and current issue activity around a young project
Maturity2/5The current line is beta and makes a hard break from v0.2

Who it’s for

Developers already using multiple coding-agent CLIs who need their sessions and artifacts to survive terminal closures.
Technical operators who want scheduled or webhook-triggered agent work with inspectable state and approval controls.
Teams willing to define reusable agents, skills, task files, and Loops around a local daemon.
Tool builders who want the same agent runtime exposed through CLI, HTTP, server-sent events, Unix sockets, MCP, and extension SDKs.

Who it’s NOT for

Conservative production teams that require a stable current line: the README labels v0.3 beta, while v0.2.15 is deprecated and receives only critical fixes.
Existing v0.2 users seeking an in-place upgrade: the migration guide calls v0.3 a hard cut, with incompatible databases, commands, extension contracts, and review behavior.
People expecting a hosted service with no local operations: CompozyOS centers on a home-scoped daemon, local SQLite state, installed agent CLIs, and operator-managed provider credentials.
Visual workflows that depend on pasting screenshots or mockups into chat: image attachments are still an open feature request in issue #366.
Codex users who cannot tolerate a failed session becoming inaccessible in the normal UI: issue #365 documents a beta.13 ACP resume failure where the transcript remained readable only through CLI recap.

Setup reality

The core install is smaller than the product's scope suggests: install one beta binary, bootstrap its home, start the daemon, and create a session from a repository. A useful deployment still requires at least one supported agent CLI, its provider authentication, an agent definition, workspace permissions, and decisions about which tasks may run unattended. Existing v0.2 installations face a manual migration rather than an upgrade. Remote access, automation, custom extensions, multiple peers, and carefully scoped secrets each add another layer after the local trial works.

The missing control plane around coding agents

CompozyOS starts from an increasingly familiar mess. A developer can launch Claude Code or another agent in seconds, but continuous work needs much more than a prompt. Someone must preserve sessions, isolate workspaces, schedule jobs, route approvals, hold credentials, record artifacts, expose status, and recover from failures. Teams often build this layer from shell scripts, terminal multiplexers, cron entries, and conventions that only one person understands.

CompozyOS puts those responsibilities into a home-scoped daemon. Sessions, tasks, Loops, memory, permissions, and agent activity live beyond any single terminal or browser connection. A CLI, web interface, HTTP and server-sent events, Unix sockets, MCP, and native tools all read the same daemon-owned state. That shared source of truth is the project's best idea. It avoids the common failure where a web dashboard says one thing while a background process has moved on.

The project calls itself an operating system for agents. That is inflated terminology, but the scope is closer to an application runtime than a simple wrapper. It manages lifecycle, persistence, automation, extension trust, and communication between agents. It does not provide the underlying intelligence. You still install and authenticate a supported ACP-compatible CLI such as Claude Code, OpenClaw, or Hermes.

A small local core with substantial responsibilities

The first-run path is reasonable. Install the beta binary through the verified shell installer, npm's beta tag, an explicit Go release tag, or source. Run compozy install, start the daemon, enter a repository, and create a session using an agent definition. State is local by default and backed by SQLite unless a configured provider or extension owns an external boundary.

That is enough for an evaluation, not a complete operational setup. Provider credentials and each CLI's own home-directory policy still matter. Agents live in global or workspace directories, and each can carry MCP configuration. Global settings, workspace overrides, explicit flags, permissions, and runtime rules form a hierarchy that operators need to understand before permitting unattended changes. The doctor and structured JSON output are helpful because automation can verify state rather than scrape friendly text.

Remote access is deliberately separate through Compozy Gateway. That keeps a local daemon from becoming network-visible by accident, but it also means remote supervision is another service and security boundary to configure. Peer discovery and delegation through Compozy Network add more possibilities, along with more identities and receipts to inspect. Start local and single-user before opening those surfaces.

Loops, sessions, and extensions fit real work

A session is durable agent conversation and activity. A Loop is reusable automation that can run on a schedule, webhook, or trigger with no terminal attached. Authored task files remain Markdown with typed frontmatter, then the runtime imports them into durable tasks for Loop execution. This gives a team reviewable files in git and inspectable runtime state instead of hiding the whole workflow in a database.

The extension model is similarly deliberate. Extensions declare versioned resources and behaviors, while the daemon retains discovery, trust, lifecycle, and hook control. Go and npm SDKs match daemon versions, and generated manifests reduce drift between code and packaging. The same runtime can expose tools through MCP, which lets another agent manage sessions and resources without a custom one-off adapter.

Compozy Network lets agents discover peers, exchange typed messages, delegate work, and close tasks with receipts. That is more accountable than vague agent-to-agent chat. Still, teams should treat every automated handoff as policy design: which agent may call which tool, where it may write, when a person must approve, and what evidence counts as completion. CompozyOS supplies objects for those answers, not the answers themselves.

The v0.3 migration is a reset

The largest buying warning is not hidden. The README says v0.3 is beta, v0.2.15 is deprecated, and the older line receives only critical fixes. The migration guide goes further, calling v0.3 a hard cut rather than an in-place compatibility release. Runtime databases, command names, extension contracts, and review behavior do not carry forward automatically.

Many old commands have new shapes or no equivalent. TUI run attachment, bulk run purging, task validation, provider-backed review fetching, PR polling, auto-push behavior, and several parallel or recovery flags were removed or deferred. No live configuration migrator ships in the beta. Old extension SDK users must port their work, while programs importing the v0.2 Go library need to pin it or redesign against current interfaces. This is honest documentation, but it makes CompozyOS a risky foundation for a workflow that cannot absorb change.

The active issues are concrete reminders of beta status. Issue #365 describes a Codex ACP process failure that leaves a valid transcript readable through CLI recap but makes normal UI reopening return an internal error. Issue #370 reports inconsistent skill resolution for extension-published agents. Image attachments are proposed in issue #366, so a user cannot yet expect ordinary paste, drag, or picker handling for screenshots and mockups.

Health is excellent, maturity is not

The repository was pushed on August 13, 2026. Beta.13 was published on August 11 after a rapid sequence of v0.3 prereleases, while the latest non-prerelease remains v0.2.15 from July 17. The six-item open count combines issues and pull requests. Several issues were created or updated on August 13, which shows immediate project activity rather than a quiet queue.

Documentation is unusually thorough for such a young system. The README explains ownership boundaries, installation channels, config precedence, extensions, skills, and the daemon model. Separate installation, provider, gateway, protocol, CLI, and migration material answers the next layer of questions. The migration guide's explicit list of removed behavior is particularly valuable.

CompozyOS is worth a serious trial when agent operations already feel like infrastructure. Its architecture recognizes the unglamorous parts that make autonomous work safe and observable. The current beta is best used on disposable or well-backed-up workspaces with explicit approvals. Wait for v0.3 stable if your priority is compatibility and predictable upgrades rather than early access to a stronger runtime model.

Alternatives

ProjectWhat it isPick it when
OpenHandsA platform for running software-development agents with a web interface and sandboxed execution.pick this instead when you want an integrated coding agent experience more than a control plane for several existing CLIs.
LangGraphA library for building stateful, controllable agent graphs in application code.pick this instead when you want to program the orchestration logic yourself and embed it inside a Python or JavaScript service.
AutoGenA framework for constructing multi-agent applications and conversational workflows.pick this instead when your main goal is developing a custom multi-agent application rather than operating existing coding-agent CLIs.

What people are saying

  1. [github-trending] compozy/compozy

Sources

  1. CompozyOS README
  2. CompozyOS v0.3 migration guide
  3. CompozyOS 0.3.0-beta.13 release
  4. Codex ACP session resume issue
  5. Image attachment feature request
  6. Extension agent skill resolution issue