Once upon a workflow, your agents came to life
OpenBook is the open-source platform where AI agents and workflows step off the page — with real conditional branching, error paths, human-in-the-loop, long-term memory, and an AI builder that writes the whole tale for you.
$ git clone https://github.com/bookbaghq/openbook$ cd openbook && npm install && npm run devAdmin opens at localhost:3000/bb-admin · MIT-licensed · your keys, your box
How OpenBook works
No glue code, no vendor lock-in. Describe what you want, let the AI wire it, then run it on your own box.
Describe it
Tell the AI builder your goal in plain English. It inventories your actions and skills and plans the workflow.
It builds the graph
OpenBook provisions the agents it needs, wires the nodes with correct data-piping, and self-verifies the result.
Run & approve
Fire it on a schedule, webhook, or event. Approve the risky steps, and watch every run in a full execution trace.
Watch a real workflow run
A scheduled trigger fires the Content Radar agent; it scores and drafts items, writes them to memory, then pauses for your approval before the digest goes out. This is the actual engine — branching, traces, schedules, and human-in-the-loop.
A workflow engine that actually does the hard parts
Most “AI workflow” tools draw a pretty graph but can't branch, can't recover from a failed step, and can't pause for a human. OpenBook's executor does all three — and more. Like any good clockwork of wonder, every gear has a purpose.
Conditional branching
Real if/else routing that prunes the untaken path — not every node fires. Expressions with comparisons, &&/||, in, and helpers.
Error paths
Route a node's failure to a recovery branch instead of killing the run. ${error} is right there to handle it.
Variables
Capture customer/order/session data once with set_variable; read ${vars.x} anywhere downstream.
Human-in-the-loop
Pause a run for approval, free-text input, or routing. Resolve from an inbox; the decision flows on as ${input.decision}.
Triggers that fire
Manual, schedule (cron), webhook, channel message, and named events — all start real runs.
Core long-term memory
Every agent remembers across runs by default — no plugin to install. Backed by a workspace memory store.
Describe it. The AI builds the graph.
Tell the builder what you want in plain English. It inventories your actions, creates the agents the workflow needs, wires the nodes with the correct data interpolation, statically verifies the result, and returns a readiness checklist of anything you still need to configure.
- Plans → provisions agents → synthesizes the graph → self-verifies & repairs
- Knows every node type, output shape, and the interpolation scope
- Returns a readiness report — what's wired vs. what to configure
- Tune the model and system prompt in Builder settings
Every morning at 7am, pull AI-engineering news from RSS,
Reddit and Hacker News, have an agent score each item 0-100,
assign a topic and write 3 hook ideas, then email me a ranked
"[Content Radar]" digest via the resend plugin.
Agents, actions, memory, and an open plugin ecosystem
Every agent in OpenBook remembers across runs by default — no plugin to install, no separate vector service to babysit. Give them tools from one shared pool of Actions and Skills, and let them call their own mid-run while the graph orchestrates the rest.
AI builder
Describe a workflow in plain English; the builder creates the agents it needs, wires the graph with correct data-piping, self-verifies, and hands you a readiness checklist.
Actions & Skills
Code-backed Actions (with intent) and markdown Skills register into one pool. Agents draw from it; activating a plugin adds more.
Plugin ecosystem
Email, chat channels, image generation, RAG, a full HTTP-action connector for any REST API — install, configure, go.
Visual + agentic
A drag-and-drop canvas that is also genuinely agentic: agents call their own tools mid-run while the graph orchestrates the rest.
Multi-model
Bring Anthropic, OpenAI, or local models. The model registry is provider-agnostic; pick per agent.
Self-hosted & private
Run it on your own box. Your data, your keys, your network. MIT-licensed, no vendor lock-in.
Cost control and human-in-the-loop, built in
Give every agent a model and a monthly budget, and keep a hand on the wheel. When a step is risky or a budget runs low, the run pauses for a human instead of charging ahead — and every decision is written to an immutable audit log.
- Per-agent model + monthly token budget; pause-on-cap, never a silent overspend
- Approve, reject, edit the input, reassign, pause, or resume any run
- Immutable audit log — who approved what, when, and the exact payload
- Bring Anthropic, OpenAI, or local models — pick one per agent
About to email the [Content Radar] digest — 24 ranked items to team@acme.com. Approve before it sends?
Every decision is logged — who approved, when, and the exact payload.
Your infrastructure. Your data. Your keys.
OpenBook is MIT-licensed and built to self-host. There's no required cloud, no per-run billing, and no black box — the whole platform is yours to run, audit, and extend. Raise the drawbridge whenever you like.
git clone https://github.com/bookbaghq/openbook
cd openbook
npm install
npm run dev # admin at http://localhost:3000/bb-admin
The agentic capabilities other tools skip
OpenBook is the open-source alternative to closed agent platforms and linear automation tools alike.
| Capability | OpenBook | Typical tool |
|---|---|---|
| Open source & self-hostable | Often closed / SaaS-only | |
| True conditional branching | Varies | |
| Error-path routing | ||
| Human-in-the-loop (pause/resume) | Sometimes | |
| Agents with long-term memory built in | ||
| AI builds the workflow for you | Rare | |
| Generic HTTP connector (any API) | Varies |
Frequently asked questions
Every great tale begins with a single page.
OpenBook is free and open source. Self-host it in minutes, describe your first workflow, and watch the story write itself.
$ git clone https://github.com/bookbaghq/openbook$ cd openbook && npm install && npm run dev