Node Agent

"Visual programming powered by AI — 39 nodes, zero boilerplate, runs when you’re ready."

39 Built-in Nodes
11 Node Categories
Save Flows & Reuse
Any Agent Handoff
FrankenCoder Node Agent showing the node library, a procedural galaxy generator flow on the canvas, and the Class Design node settings panel

Describe the task. The agent builds the flow.

The Node Agent is a three-panel visual canvas: a node library on the left, the flow canvas in the center, and a properties panel on the right. Tell the agent what you want built — a web scraper, a report pipeline, a full application — and it designs the flow, drops in the correct nodes, wires the connections, and produces a runnable program.

You can build or edit manually at any point. Drag nodes from the library, connect them with Bezier wires, configure each node in the properties panel, and run. The agent builds on top of the same canvas you use. There is no separate "AI mode".

39 Nodes. 11 Categories.

Every node connects to the FrankenCoder daemon for LLM calls, terminal access, web, YouTube, and file I/O. Custom node types can be registered with their own fields and handle configuration, and they persist across sessions.

Flow Control
Start, End, Delay
Triggers
Manual Trigger, Schedule (cron), Webhook, File Watch (glob + debounce)
AI / LLM
AI Node — any configured model, full prompt/temperature/system prompt control
Web / IO
Web Fetch (GET/POST), Web Scrape (DOM extraction)
YouTube
Search, Transcript, Batch Transcripts, Clip Finder, YouTube Report (daily/topic/creator modes)
Data Processing
Transform (JS), Filter, Merge, Text Chunker, Template, Join, Reduce, Extract (JSON path/regex), Variable, Parse JSON, Sort
Orchestration
If-Condition, Loop (10k iteration cap)
Project Setup
Folders, Dependencies (npm / pip / cargo / go / composer / maven / gradle)
Design
Class Design — UML blueprint with file path, language, properties, methods, purpose, and implementation instructions
System
Command (shell), Read File, Write File, Lint, Log, File List (glob), Env Var
Error
Catch — fallback routing and rethrow

Design programs in UML. The agent writes the code.

The Class Design node is not a template filler. It is an intent-driven code generator. Describe the class: its file path, language, properties, methods, purpose, and implementation requirements. At run time the agent reads that spec and writes a complete, idiomatic implementation directly to disk — decorators, imports, dependency injection, business logic, error handling, all of it.

Chain a Folders node to create the project structure first. Chain a Dependencies node to install packages. Chain as many Class Design nodes as the project needs. The agent scaffolds and writes the entire application in one flow run.

Class Design Node — Example
Class Name
GalaxyRenderer
File Path
galaxy-generator/src/renderer.ts
Language
TypeScript
Purpose
WebGL canvas renderer for star fields. Manages shader programs, buffer geometry, and per-frame animation loop. Stateless except for canvas context.
Methods
init(), render(timestamp: number), resize(w: number, h: number), dispose()
Not a template writer

The Class Design node produces real code, not filled-in stubs. The agent writes the file itself — no downstream write node required.

Flows are assets. Save, load, reuse.

Flows are named and stored with full metadata: last modified, node count, canvas state including zoom and pan position. Both you and the agent can save and load them. Export to JSON to share or back up. Import to restore. Preset flows ship out of the box.

web scraping data pipelines report generation project scaffolding scheduled tasks file processing YouTube research any repeatable task

The agent runs it. And hands off when needed.

Execution is transparent. Every node gets its own status: pending, running, passed, failed, needs_human, or skipped. The execution log at the bottom is timestamped and expandable — click any entry to highlight the corresponding node on the canvas. Validation runs before execution, catching orphan nodes, cycles, and a missing start node before anything runs.

The Node Agent can hand off to any other FrankenCoder agent mid-flow. In the demo above: after writing all source files and installing dependencies, the Node Agent calls the Browser Agent to open the generated application and confirm it runs.

Node Agent → Browser Agent → Cyber Agent → Planner → Any Agent

Sequential or parallel branch execution. Stop button halts mid-run. The agent can flag any node as needs_human and pause for your input before continuing.

"Describe a procedural galaxy generator. The Node Agent designs the flow, scaffolds the project, writes every source file, installs the dependencies, and calls the Browser Agent to confirm it runs. Zero manual steps."

Build Your First Flow

Describe a task. Watch the agent wire it up. Run it.

Get FrankenCoder Back to all features →