CompozyOS documentation
What CompozyOS is, how sessions, memory, permissions, automation, and Compozy Network fit together, and where to start for your job.
CompozyOS is an agent operating system for real work. It runs the agent CLIs you already use — Claude Code, OpenClaw, Hermes, and others — as daemon-owned sessions with durable history, workspace-aware memory, permissions, tasks, Loops, automation, and coordination.
The useful part of the operating-system label is the connection between those parts. A prompt can continue after a viewer disconnects, its events stay inspectable, workspace policy still applies, and people or agents can operate the same state through structured surfaces.
Choose your path
The docs serve four jobs. Start with yours — every path stays inside the same system, so the deeper pages are always one link away.
Start and steer agent work
Install the runtime, start your first session, and watch work finish. Sessions stay durable and inspectable — nothing lives in a throwaway terminal tab.
Build on the runtime
Extensions, hooks, skills, tools, and SDKs plug into daemon-owned registries and public contracts. Anything the UI can do, an agent can do through CLI, HTTP, or UDS.
Implement compozy-network/v0
The wire format for agent-to-agent coordination: envelopes, message kinds, delivery, and receipts — implementable without adopting Compozy internals.
See what ships today
Real use cases and runnable examples, each grounded in a shipped mechanism and labeled with honest maturity. No roadmap dressed as product.
One system, not a stack of tools
Everything in these docs shares one daemon-owned state: a task can start a session, permissions bound it, memory follows the workspace, and another agent can continue the work without rebuilding context by hand. Compozy Network — peer discovery, typed messages, lifecycle-bearing work — is a subsystem of that same runtime, documented under its own group below with the protocol spec nested for implementers.
# no Docker, no Postgres — one local binary
compozy daemon startHow the docs are organized
One sidebar, eight groups, ordered by journey: start, do, understand, automate, coordinate, extend, operate, look up. Task-oriented content sits above concepts; the generated references collapse to two rows until you need them.
Getting started and migration — install, run the quick start, meet your first agent.
Work that runs without you: jobs, triggers, and webhooks, goal-seeking Loops, and chat bridges.
Using the network from Compozy, with the compozy-network/v0 protocol spec nested one level deeper for implementers.
How the system grows: the marketplace, skills, tools, resources, extensions, and hooks.
Running the daemon well: operations and every
configuration surface, from config.toml to file locations.
Generated from the runtime, never paraphrased: the CLI reference and the API reference.
The problem Compozy solves
Most agent CLIs are excellent at doing work inside one session, but weak at everything around the session:
- the run disappears when the terminal closes
- there is no clean replay or audit trail
- each project needs its own context, defaults, and guardrails
- coordinating multiple specialized agents turns into custom glue code
- operators end up stitching together scripts, logs, and dashboards that never become a product
Compozy fills that gap with a local-first runtime that keeps work durable and observable without asking you to abandon the agent CLIs you already use.
What the CompozyOS runtime adds
Run real agent CLIs as durable work
Compozy does not reimplement agent logic behind API wrappers. It orchestrates real agent CLIs as managed sessions, so Claude Code stays Claude Code, OpenClaw stays OpenClaw, and Hermes stays Hermes.
That matters because you keep the native tools, context behavior, and interaction style of each agent instead of flattening everything into one abstraction layer.
Keep sessions durable, replayable, and auditable
Each session becomes a durable unit of work. While it is active, another client can acquire an
attach lease with compozy session resume. After it stops, the public lifecycle is terminal, but
its status, recap, event history, and transcript remain available for review.
Instead of losing the context after a crash, restart, or handoff, Compozy keeps the operational record around the work itself.
Standardize behavior per workspace
Compozy layers configuration, memory, and skills around the workspace, so each project can carry its own defaults without forcing one global setup to fit every team and repository.
That gives operators a cleaner way to keep agent behavior consistent without rebuilding prompts and context on every run.
Operate through shared control surfaces
The same daemon-owned state is available through the CLI, HTTP/SSE, UDS, and the web UI. Managed
agents use policy-filtered native tools, and trusted external MCP clients can receive a
workspace-bound Host API projection through compozy mcp serve. Sessions, memory, tasks, automation,
bridges, and observability stay in one control plane instead of separate stores.
Opt into Compozy Network when agents need to coordinate
Compozy Network is the coordination subsystem inside CompozyOS. Local is the default; a session or task run participates only when its execution is explicitly Live. Live participants can discover peers, exchange typed messages, and track lifecycle-bearing work in a channel.
The wire boundary remains independently implementable even though Network is part of the integrated system. Agents can publish capability catalogs so peers discover outcomes instead of internal tool names.
Under the hood, Compozy runs as a single local daemon that manages sessions, persists events, and talks to compatible agents through a standardized protocol layer. The operating model matters more than the internals — head into Sessions when you want the package-level view.
Why Compozy vs. other models
The AI agent ecosystem is crowded. The useful question is not "which project exists," but "which operating model are you actually buying into?"
CompozyOS combines local agent execution, durable state, workspace policy, memory, tasks, Loops, automation, extensibility, an OS shell, and optional Live coordination. That integrated operating model is a different lane from a channel-native assistant, hosted gateway, standalone agent framework, or tool protocol.
| Operating model | Best fit | Where CompozyOS differs |
|---|---|---|
| Personal assistant or channel-native helper | You want one assistant persona inside chat tools and user-facing channels. | CompozyOS focuses on durable work, shared state, policy, and coordination across agents. |
| Hosted or multi-tenant gateway | You need a centralized hosted control plane and shared tenancy. | CompozyOS is local-first and self-managed on the machine doing the work. |
| Agent framework | You want to build custom reasoning graphs or orchestration from application code. | CompozyOS manages existing agent CLIs and also owns deterministic Loops, tasks, and runtime state. |
| Tool protocol only | You need an agent to call tools and read resources. | MCP integrates into CompozyOS; tool access is one part of a system that also owns execution, state, and governance. |
| Agent operating system | You want sessions, memory, tasks, permissions, automation, coordination, and presentation connected. | This is CompozyOS's primary lane. |
Most alternatives optimize for one of these lanes: assistant surface, hosted gateway, framework for custom orchestration, or protocol for tool access. CompozyOS connects those operational concerns around one daemon-owned state model, while Compozy Network supplies the optional coordination boundary for explicitly Live work.
When Compozy is the right choice
Compozy is a strong fit when you:
- run real agent CLIs and need them to behave like durable work instead of ephemeral chat
- want replay, auditability, and observability without bolting on external control planes
- need project-specific memory, policies, and behavior without rebuilding prompts every time
- need people, managed agents, and trusted integrations to operate the same runtime state
- expect specialized agents to coordinate through explicit, bounded Live participation
When another lane may fit better
- Choose a channel-native assistant product if the main job is user-facing help inside chat tools.
- Choose a hosted gateway if the main job is centralized multi-tenant operations.
- Choose an agent framework if the main job is building custom orchestration logic from the ground up.
- Choose a protocol-only layer if the problem is tool connectivity, not runtime control or coordination.
CompozyOS exists for the gap between those categories: an integrated, extensible operating system around real agent work, with a coordination wire boundary that does not have to stay Compozy-only.
What Compozy is not
- Not an agent framework. Compozy does not ask you to rebuild your agents as chains, graphs, or wrapper abstractions.
- Not an API wrapper. The point is to keep real agent CLIs intact, not to flatten them into thin model calls.
- Not a hosted control plane. Compozy is local-first. The runtime lives on your machine and keeps data close to the workspace.