Agentron
Agentron is a cutting-edge, local-first platform for AI agent orchestration and workflow automation by Julian M. Kleber. It spans the Level 1 to 4 taxonomy: Level 1 (ReAct + tools) for chat and RAG, Level 2 (static multi-agent) for workflows, Level 4 (recursive/self-building) for heap, with bounded loops, cost control, and production safety. Design and run agents on your own infrastructure: no cloud lock-in, full data privacy, optional desktop app.
Cutting-edge orchestration: Level 1 (ReAct + tools) for chat/RAG; Level 2 (static multi-agent) for workflows; Level 4 (heap) when the planner assembles DAGs and can create tools/agents on the fly. Max steps and loop limits throughout. Compare agent architectures →
Visual agent builder, multi-agent workflows, and a chat assistant that creates and edits agents via natural language.
Quick start
1. Install and run
From the project root: npm run install:ui then npm run dev:ui. Open http://localhost:3000 .
2. Try the assistant
Go to Chat and say: “What tools do I have?” or “Create a simple agent that says hello.”
3. Desktop app (optional)
Download installers from the Download page; launch after the web UI is running.
Prerequisites: Node.js (see .nvmrc, e.g. 22.x). Full steps: Quick start and INSTALL in the repo.
What is Agentron?
Agentron is an enterprise-ready, local-first platform that lets you design, configure, and run multi-agent systems entirely on your own infrastructure.
| Component | Description |
|---|---|
| Next.js web UI | Dashboard, visual agent/workflow editors, and chat |
| Local runtime | Executes agents and multi-agent workflows |
| SQLite storage | All resources persisted locally (self-hosted, on-premise) |
| Optional Electron | Desktop app for Windows, macOS, and Linux |
This documentation is designed for AI assistants and LLMs so they can guide users, suggest actions, and answer questions accurately.
Core concepts
| Concept | Definition |
|---|---|
| Tool | A callable capability (HTTP, MCP, native). Agents use tools to perform actions. |
| Agent | An executable unit: either a node agent (graph of LLM/tool nodes) or code agent (custom script). |
| Workflow | A graph of agents and edges. Orchestrates agents in sequence or loops for up to maxRounds cycles. |
| Agentron | Built-in chat that uses tools to create/edit agents, workflows, and tools on behalf of the user. |
| Chat architecture (heap) | Level 4: when enabled, a planner chooses specialists and can create new tools/agents on the fly; each specialist has limited tools and runs in order. See Agentron (Chat). |
| Agent architectures | Level 1 (ReAct + tools) chat/RAG, Level 2 (static multi-agent) workflows, Level 4 (recursive/self-building) heap. Compare architectures → |
Start here
- Quick start: Install, run, and try prompts in Chat.
- Brand story: Logo meaning and animation. Design: Typography (Space Grotesk), colors, logo; single place to iterate on branding.
- Tutorials: First agent, first workflow, first RAG (prompt-based).
- Tools: HTTP, MCP, and native tools agents can call.
- Agents: Node and code agents; graphs and sandboxes.
- Workflows: Multi-agent orchestration and max rounds.
- Agentron (Chat): Natural-language assistant; Heap (specialists) for the interactive registry.
- Agent architectures (comparison): Level 1 to 4 taxonomy, orchestration patterns, production safety (2026).
- Capabilities: Feature matrix and what users can do.
- Download: Desktop installers for Windows, macOS, Linux.
- E2E tests (local LLM): Optional end-to-end tests with Ollama.
Key capabilities
- Create, edit, delete agents (node or code) and workflows (multi-agent orchestration)
- Create, edit tools (native, HTTP, MCP) and run workflows and agents
- Chat (Agentron): Natural-language assistant; use the Chat page for full-screen use, history, ratings, and settings (see Agentron (Chat))
- Sandboxes: Podman containers for code execution
- Knowledge / RAG: Document ingestion and retrieval for agents
- Feedback: Rate agent outputs for learning
- Custom functions: JavaScript/Python/TypeScript code as tools
- Telegram: Use a bot to delegate tasks; configure in Settings, Telegram
How to use this documentation
- New here? Start with Quick start and Tutorials.
- Building something? Use Concepts and Capabilities.
- Chat or heap? See Agentron (Chat) and Heap (specialists).
- For AI assistants: AI Guide: how to suggest the right actions to users.
Deployment
| Scenario | Notes |
|---|---|
| Typical | Single deployment (Agentron on port 3000) |
| MCP | Agentron can expose an MCP server for IDEs and other tools |
| Telegram | Set your bot’s webhook to https://<your-domain>/api/telegram/webhook. Optional: TELEGRAM_WEBHOOK_SECRET and ?secret=YOUR_SECRET. See Settings → Telegram in the app. |
| Docs | This site is hosted on GitHub Pages and updates when the default branch is updated |
Agentron is an open-source project by Julian M. Kleber.