Digitorn
Digitorn
Digitorn Builder

Design AI agents by talking.
Wire them on a canvas.
Ship in one click.

Tell the builder what your agent should do. It generates the YAML, validates it against the live runtime, and shows it to you on a canvas you can edit by hand. Both views write the same file, so you never choose between power and ease.

Try Digitorn BuilderBrowse templates
Edit modes
2
Canvas · YAML · same state
View lenses
5
Architecture · Security · Perf · Runtime · Sequence
Lifecycle lanes
9
Inputs to Outputs, in order
Undo depth
200
Coalesce-aware history
Why it matters

YAML is power. Most people want a conversation.

Digitorn apps are declarative: every agent, tool, hook, and trigger fits in one YAML file. That makes them auditable, portable, and easy to ship. It also makes the first step harder than it has to be when you do not yet know the schema.

The builder solves both ends. Newcomers describe the agent and get a working app without reading the spec. Power users get a canvas that turns the same YAML into a graph they can rewire, inspect, and stress-test. Switch tracks any time, the file stays in sync.

What the Builder does for you
  • Talks to you in plain language
    Asks targeted questions until it understands the goal, never makes you guess what to type.
  • Shows the app as a graph
    Agents, modules, hooks, channels, capabilities. Every edge is a real wire in the runtime.
  • Validates before you deploy
    The runtime compiles every change. Errors surface inline with the offending node highlighted.
  • Wires credentials safely
    Slots prompt for the right scope, never leak secrets into the YAML, audit trail by default.

Two ways in, one source of truth

The conversational track and the canvas track are not parallel tools that drift apart. They write the same YAML, every keystroke, every drag.

CONVERSATIONDescribe the appin plain languageCANVASDrag, drop, editvisual graphBUILDERGenerate · Validategrounded in 3 RAG indexesONE STATEapp.yamltwo-way boundlive validatedundo / redoDEPLOYone clickEither track edits the same YAML. The other view updates instantly.
The canvas

Not a flowchart. A real model of the runtime.

Most visual editors render boxes for show. The Digitorn canvas mirrors what the runtime actually does, swimlane by swimlane, in the same order the loop executes them. If a node is wrong on the canvas, the agent is wrong in production.

Lifecycle swimlanes
01Inputs
User message, trigger fire, channel webhook
02Palette
Slash commands resolve before the brain sees the message
03Behavior
Classifier injects directives, per-tool checks fire later
04Middleware
.before() and .after() wrap the LLM call
05Agents
Brain decides, streams text, calls tools, spawns specialists
06Capabilities
HITL gates pause the loop on risky tools
07Tools
Module actions execute with start and end hooks
08Hooks
Transversal interceptors, fired around tools and turns
09Outputs
SSE, preview, output channels (slack, email, webhook)

These nine lanes are the runtime, not a metaphor. Inputs land, commands resolve, behavior injects directives, middleware wraps, agents decide, capabilities gate, tools run, hooks intercept, outputs ship. Click any node to jump to its exact YAML location.

Five lenses on the same graph

Most graph editors give you one view. The Digitorn canvas gives you five, each tuned for a different question. Toggle in one click, never lose your place.

Lens
Architecture
Full graph, every node visible. The default view.
Lens
Security
Highlight risky tools, HITL gates, missing approvals. Dim everything else.
Lens
Performance
Focus on brains, fallbacks, sub-agents, costly tools. Spot the hot paths.
Lens
Runtime
Animated turn walkthrough. Watch one fictional run unfold step by step.
Lens
Sequence
UML-style diagram. Inputs and outputs of one runtime turn, top to bottom.

Two-way binding, with ripple rename

Edit the YAML by hand, the canvas updates immediately. Drag a node, the YAML rewrites at the right path. Rename an agent, every reference in hooks, channels, capabilities, and sub-agents follows.

Undo and redo are 200 steps deep with smart coalescing, so a burst of small edits collapses into one reversible step. No surprise diffs at commit time.

YAMLagents: - id: helper modules: [{web: [search]}] brain: model: claude-haiku-4-5 credential: anthropic_main system_prompt: | Answer concisely. Cite sources.CANVASuserhelperwebINSPECTOR · agent.helpermodel: claude-haiku-4-5tools: web.searchedit YAML → canvas updatesedit canvas → YAML updatesRename an agent in either view, every reference rewires automatically.

What sets this canvas apart

Twelve concrete features that competitors do not ship together. Every one of them is wired through to the live runtime, not a mock.

Conversational interview

Tell the builder what you want. It asks the right follow-up questions and stops when it has enough to ship.

Bidirectional canvas

Drag, drop, connect. Or write YAML directly. Either side wins, the other follows in real time.

Lifecycle swimlanes

Nine canonical lanes mirror the runtime loop, top to bottom. You always know what runs next.

Security lens

Highlights risky tools, missing approvals, ungated destructive actions. Audit before you ship.

Performance lens

Brain providers, fallbacks, sub-agents, expensive tools. Spot the cost paths at a glance.

Story mode

Animated narrated walkthrough of one fictional turn through your app, derived from the YAML.

Sequence diagram

UML-style timeline of one runtime turn, top to bottom. Great for design reviews and docs.

Beginner mode

Humanises every label. context_builder.ask_user becomes Ask the user. Toggle once, read forever.

Command palette

Cmd-K to find any node, jump to it, focus, pin. The whole app is one keystroke away.

Live validation

Every change recompiles against the runtime. Errors anchor to the exact node, not a line number.

Smart undo

200-step history with coalescing. Bursts of edits collapse into single reversible commits.

Adaptive density

Comfortable on 30 nodes, compact at 80, list above. The canvas resizes to the app.

What you can build with the Builder

Five archetypes cover the bulk of what teams actually ship. The builder recognises which one you are describing and starts you on the right rails.

Archetype

Conversational assistant

A chat agent with tools, memory, and a clear scope. Slack copilots, support helpers, internal Q&A.

Archetype

Scheduled monitor

A cron-driven agent that watches something and reports. Job boards, dashboards, news beats, KPI alerts.

Archetype

Webhook processor

Triggered by events. New PR, new ticket, new payment. Reads, decides, posts back, logs everything.

Archetype

Document pipeline

Ingest, extract, summarise, file. RAG-friendly, validation-aware, bulk-friendly with rate limits.

Archetype

Multi-agent research

A coordinator dispatches specialists, joins their findings, writes a report. Fan-out, join, synthesise.

Quickstart

Open the Builder, describe your agent, deploy in five minutes.

# 1. install the runtime
curl -sSL https://digitorn.ai/install | sh

# 2. start the daemon
digitorn daemon

# 3. open the builder in your browser
digitorn open digitorn-builder
Newsletter

Get the next post in your inbox.

Engineering notes from the Digitorn team. No marketing, no launch announcements, no "10 prompts that will change your life". Just the things we write that we'd want to read.

One-click unsubscribe. We never share your address. Powered by our own infrastructure, not a tracker.