Fully local · Zero API · No Ollama

DevonAI

Your personal agent that runs entirely on your machine.
Single binary. Fully offline. Zero cloud dependency.

Get Started View Architecture

DevonAI — Local Agent

❯ devonai

Loading model dev0.1-q4_k_m.gguf ...

Model ready · context 8192 · local quantized model

──────────────────────────────

you › Analyze the project structure and suggest improvements

Thinking...

→ tool: fileTool.list + terminalTool

Devon › Scanned 24 files. Suggest refactoring engine init in src/core...

Built for real local work

Not another cloud chatbot. A complete agent system that runs on your own machine.

Interface

Terminal-native UI

TypeScript + React + Ink. Status bar, streaming chat, thinking indicators, and a fixed input box — all driven by the event bus.

Inference

In-process local models

Loads GGUF models directly inside the Node.js process with node-llama-cpp. No Ollama daemon. No external services.

Agent

ReAct loop

Think → tool call → local execution → result back to the model. Multi-turn until the task is done or a limit is hit.

Safety

Sandboxed tools

Terminal commands are filtered for high-risk patterns. File tools are restricted to the project root with path traversal protection.

Setup

Automatic model download

On first launch the agent checks for a local model and downloads a lightweight quantized checkpoint from Hugging Face if needed.

Runtime

Event-driven core

A global EventBus connects the model engine and the UI. Tokens, tool calls, and results update the interface in real time.

Three core modules

From interface to engine to tools — every layer is designed for local-first operation.

Module 1 · Terminal UI

A modern terminal interface with a status header, streaming chat history, thinking indicators, and a fixed input box. Everything is driven by the event bus so the UI stays in sync while the agent thinks or runs tools.

View structure

            // src/ui/App.tsx

            import { Header, ChatStream, InputBox } from './components';

            import { eventBus } from '../core/eventBus';

            export function App() {

              const [messages, setMessages] = useState([]);

              // Listen to core events and update UI

              useEffect(() => { ... }, []);

              return (

                <Box flexDirection="column">...

              );

            }

Module 2 · Local model engine

node-llama-cpp loads GGUF models in-process. On first run the agent checks for a local model and downloads a lightweight quantized checkpoint if needed. streamPrompt provides efficient token streaming with sensible defaults.

How download works

            // src/core/engine.ts

            import { getLlama, LlamaChatSession } from 'node-llama-cpp';

            export async function streamPrompt(prompt: string) {

              const session = new LlamaChatSession({ ... });

              for await (const chunk of session.prompt(prompt)) {

                eventBus.emit('token', chunk);

              }

            }

Module 3 · Agent loop & tools

A ReAct loop coordinates thinking and tool use. A robust JSON parser recovers tool calls from imperfect model output. terminalTool blocks dangerous commands; fileTool stays inside the project root.

Tool design

            // High-risk command filter

            const BLOCKED = [

              /rm\s+-rf\s+[\/~]/,

              /mkfs/, /dd\s+if=/,

              /chmod\s+-R\s+777\s+\/

            ];

            if (BLOCKED.some(r => r.test(cmd))) {

              throw new Error('Blocked dangerous command');

            }

Why DevonAI

How it differs from common setups.

Capability	DevonAI	Ollama + extras	Cloud API
Fully offline	Yes	Yes	No
No extra daemon	Yes	Needs Ollama	Needs network
Single-binary feel	Yes	Multiple parts	Account + key
Data never leaves the machine	Yes	Yes	No
Built-in tool sandbox	Yes	DIY	Limited
Terminal-native UI	Yes	—	—
Automatic model management	Yes	Yes	—