Everything in one open-source platform
A visual canvas that's genuinely agentic — agents call their own tools while the engine orchestrates branching, recovery, state, and human approvals around them.
A workflow engine that does the hard parts
OpenBook's executor walks a graph and actually branches, recovers, pauses, and carries state — the things most AI-workflow tools only pretend to do.
Run only the matching branch; prune the rest.
Route a failed node to recovery, not a dead run.
set_variable writes ${vars.x} for everything downstream.
Manual, schedule, webhook, channel, and events.
Describe it; the AI builds the graph
The builder inventories your actions, creates the agents the workflow needs, wires the nodes with correct data-piping, self-verifies, and returns a readiness checklist.
A pipeline, not a one-shot prompt.
Provisions the agents a workflow needs, with tools.
Pick the builder's model and system prompt in settings.
Static checks + up to two repair rounds.
Agents that remember, on the models you choose
Multi-model agents run their own plan→act→observe loop and remember across runs by default — long-term memory is core, no plugin to install.
memory_* tools on every agent by default.
One shared pool of capabilities to draw from.
Pause for approval/input/choice; resume on decision.
Anthropic, OpenAI, or local — per agent.
An open ecosystem, plus any API
Install plugins for email, channels, images, and RAG — or call any REST API with the generic HTTP action's auth, retries, pagination, secret store, and SSRF guard.
Actions, providers, channels, and UI.
http.request for any REST API, safely.
Your infra, your data, your keys. MIT.
Write a plugin: manifest + handlers.
Build your first agentic workflow today.
OpenBook is free and open source. Self-host it in minutes, or read the docs to see how far the engine goes.