Open Source
Local-First AI Agent

Rust-powered AI agent. Zero runtime dependencies.
100+ LLM providers, Discord bot, voice I/O.

An AI agent that belongs on your desktop, not in the cloud. Single binary, zero dependencies, built with Rust for performance and reliability.

Core Capabilities

Everything you need to run AI agents on your own infrastructure.

100+ Providers

OpenRouter, OpenAI, Anthropic, Google AI, GitHub Copilot, Ollama, Groq, DeepSeek, xAI, AWS Bedrock, Azure, and many more.

Web UI

Modern chat interface with workspaces and tool visualization.

30+ Built-in Tools

File ops, code execution, grep/glob, web fetch, LSP, and more.

Voice I/O

Whisper STT + Piper TTS with browser fallback.

Skills System

Extend with custom integrations. Installable .oskill bundles.

OAuth Login

Sign in with GitHub Copilot, Google, or OpenAI Codex. No API keys needed.

Visual Workflow Editor

Node-based drag-and-drop automation. Build complex pipelines visually with conditions, loops, and branching logic.

Discord Bot

Slash commands, per-channel sessions, real-time tool progress, and thinking indicators.

Jobs Scheduling

Schedule reminders, recurring tasks, and daily briefings with cron expressions or natural language.

Benchmarks

Measured with in-repo runtime benchmarks (debug vs release, provider-free workloads).

~543ms
Startup to Ready

Release avg over 10 runs to first successful /api/auth/status (p50 ~549ms)

~13.68MB
Ready RSS

Resident memory sampled immediately after server readiness

~22.66MB
Idle RSS

Resident memory after 2s idle settle window

Latest run: 2026-04-08, 10 release iterations on a Windows dev machine. Method: cargo run --release --bin osagent-bench -- --profiles release --iterations 10 --skip-build.

Metric
OSAgent
Native Agents
Web Platforms
Memory (idle RSS)
~22.66MB
500MB-1.5GB+
1GB-2GB+
Runtime Dependencies
Zero
Node.js / Bun
Python + Docker
Startup to ready
~543ms
100-200ms
Seconds
Install Size
~50MB binary
npm + 100+ deps
2-5GB image
Response Compression
Brotli + Gzip
None
Varies
DB Optimization
WAL + Cached Stmts
Basic
ORM Pool
Binary Optimization
Fat LTO + Strip
N/A (interpreted)
N/A (interpreted)

"Making it feel like a useful assistant integrated into your daily routine. Not just another tool you open occasionally, but something that actually helps with real tasks."

OSAgent Design Philosophy

Whether it's coding, research, automations, or just having a conversation while you're working. The dream is an agent that feels like it belongs on your desktop.

Architecture

Built with Rust for performance and reliability. Runs at http://localhost:8765 with optional GUI launcher. Read our research on self-optimizing prompts for more.

Super-duper fast

OSAgent keeps the system compact: Rust on the backend, a static web UI on the frontend, local persistence by default, and optional subsystems only when you need them.

Axum + Tokio serve the app locally at localhost:8765
SQLite stores sessions, messages, and agent state on-device
Voice and search stay optional so the base install stays lean

Runtime

Tokio 1.35

Port

localhost:8765

Config

~/.osagent/

License

MIT

Core stack
Rust Tokio Axum SQLite Vanilla JS Whisper Piper MeiliSearch

Integrations

Connect to your existing tools and platforms.

OpenRouter
OpenAI
Anthropic
Google AI
GitHub Copilot
OpenAI Codex
Ollama
Groq
DeepSeek
xAI
AWS Bedrock
Azure

Frequently Asked Questions

Common questions about OSAgent, local-first AI agents, and self-hosting.

What is OSAgent?

OSAgent is a personal AI agent built with Rust. It started as a way to bring coding agent-style tools into a lighter, more personal assistant context. Think of it as the helpful agent that lives on your machine, helping with coding tasks, research, automating stuff via Discord, or just being a voice away when you're hands-busy. Zero runtime dependencies, ~50MB binary, starts in milliseconds.

How is this different from Cursor or Windsurf?

Those are great coding-focused editors. OSAgent is more of a general personal assistant that happens to be really good at coding tasks too. The main differences: it's a single binary (not an Electron app eating 500MB), it runs Discord and voice interfaces out of the box, and it's designed to integrate into your daily workflow rather than being a dedicated IDE. Also: you own it completely, runs locally, no cloud dependency.

What do I need to get started?

You bring your own model. Options: GitHub Copilot or OpenAI Codex via OAuth (sign in with your existing account), any OpenRouter/Anthropic/etc API key, or Ollama for fully local models. Download the binary, run it, open localhost:8765. That's it. No Docker, no Node, no Python.

Can I run it fully offline?

Yes. Point OSAgent at a local Ollama instance and you're off-grid. Great for privacy-sensitive work or just avoiding API costs. The binary itself is tiny (~50MB) so you can even run it from a USB stick if needed.

What models do you recommend?

Depends on the task. GLM 5.1 is great intelligence to price ratio, tends to lose its mind occasionally but gets tuned regularly. Minimax M2.7 is like GLM but cheaper and more reliable, solid all-rounder. For heavy lifting like planning or debugging: GPT 5.4 High. For raw speed and execution: GPT 5.3 Codex - it will dump 17 file writes before you blink if it knows what to do.

How does Discord integration work?

Drop your bot token in the config, and OSAgent becomes a Discord bot. Per-channel or per-user sessions, slash commands, real-time tool progress, thinking indicators while it works. Same binary as the web UI. No separate deployment.

What's the goal for OSAgent?

Making it feel like a useful assistant integrated into your daily routine. Not just another tool you open occasionally, but something that actually helps with real tasks, whether that's coding, research, automations, or just having a conversation while you're working. The dream is an agent that feels like it belongs on your desktop, not in the cloud.

How much does it cost?

Free and open source under MIT. The binary has no ads, no subscriptions, no paid tiers. You just pay for whatever API keys you decide to use. If you run Ollama locally, the cost is literally zero.

Ready to build?

Deploy your own AI agent in minutes.

Get Started on GitHub