The core is a REST API that stores work items with structured state. Agents read and write state. The system auto-generates history. Artifacts are first-class objects. Claims prevent collisions.
Cloud by default. That's not a convenience — it's a requirement. Local state fragments the moment you have more than one agent or one device. Shared state only works if it's always reachable.
Your command center. See every work item across every project. Who's claimed what, what's blocked, what's complete. Override state, approve actions, add items manually.
Claude Code, OpenClaw, Hermes, custom agents — anything that can make HTTP requests. Pull-based: agents claim work before acting, release when done. Each agent has a registered identity.
LLM-powered coordination. Detects duplicate work items across agents. Suggests priority ordering. Summarizes execution context for long-running items. Consumes the same API as your agents.
Work items, artifacts, events, claims, agents, projects. No LLM calls. No inference. Every mutation auto-generates a structured event. Fast, reliable, self-hostable. The open-source heart.
Your state, your data. On Cloud Threadron, we host it — reliable, backed up, always accessible from any agent or device. Single-node, you run it in Docker on your own machine.
The atomic unit. Carries goal, current_state, next_action, blockers, outcome_definition, and confidence. Structured state — not a title and a status field.
Auto-generated on every mutation. 11 structured types: status_change, observation, decision, blocker, artifact_added, claim, handoff, and more. Actor attribution on every event.
First-class attached objects. Types: branch, PR, commit, file, plan, doc, terminal_output. Not buried in notes — queryable, linkable, part of the structured record.
Agents claim a work item before acting. Claim carries claimed_by, expires_at, and intent. Prevents two agents from starting the same work. Released when done or on timeout.
Registered identities. Each agent has a name, capabilities, and last_seen timestamp. Every event is attributed to an agent, human, or the system itself.
Organizational hierarchy. Projects group related work items. Domains scope visibility and defaults. Rarely changes — provides structure without overhead.
Every agent follows the same pattern: claim before working, update state while working, release when done. The state lives in the cloud — any agent, any device, picks up the same thread.
Agents connect via MCP server (native tools for Claude Code, OpenClaw, and other MCP-compatible agents) or via REST API (copy skill.md into any agent's system prompt). Setup guides →
Real workflows. Real agents. State that never breaks.