OpenWOP defines the wire format, state machine, and trust boundaries for orchestrating networks of AI agents — so a workflow written once can run anywhere, observed, idempotent, and durable by default.
A typed DAG with idempotent edges, replayable runs, and resumable state — across restarts, crashes, and clouds.
Secrets are resolved and redacted at runtime. Trust boundaries, allow-lists, and per-agent quotas live in the workflow layer.
W3C traceparent propagation across every agent hop. Token, time, and dollar accounting attaches to each run automatically.
A declarative manifest of nodes, edges, and policies. Hashable, versionable, and signable.
Pending → running → suspended → succeeded / failed / cancelled — with checkpoints between every node.
First-class primitives for agent thoughts, scratchpads, and shared semantic memory across the run.
Built-in HITL nodes that pause execution for approval, clarification, or annotation, then resume in place.
Token streams, partial outputs, and run events delivered over Server-Sent Events with polling fallback.
Curated, signed bundles of tools and agents discovered through packs.openwop.dev.
Model Context Protocol connects a single model to its data and tools. WOP sits one layer up — orchestrating between models, routing the output of one to the reasoning of another, and keeping a durable trace of the whole transaction.
Most agent frameworks tether you to a single runtime, cloud, or model family. WOP defines only the wire format and state model — swap runtimes, self-host the registry, or move clouds without rewriting any agent logic.
A reference runtime, a starter pack, and the spec — that's it. No accounts, no SDKs, no telemetry pings. Read the manifest, run the server, watch a workflow execute.
