A self-improving AI orchestrator that runs 100% locally.
Single Go binary. 40+ tools. 20+ LLM providers. Multi-agent coordination.
Persistent memory with a knowledge graph. Browser automation. Computer control.
And it gets better at its job over time.
🚀 Quick Start · ✨ Features · 📊 Scorecard · 🏗️ Architecture · 🤝 Contributing
Most AI assistants are cloud-locked chatbots that wait for you to type something. Sofia is different.
- 🏠 Local-first — No cloud dependency. Your data stays on your machine.
- ⚡ Single binary — Written in Go. No Python, no Docker, no Node.js. Just
make build. - 🧠 Self-improving — Evolution engine analyzes past performance and optimizes itself.
- 🤖 Multi-agent — Spawn, coordinate, and retire agents dynamically. A2A protocol built-in.
- 🔐 Security-first — 35+ prompt injection defenses, PII detection, AES-256-GCM encryption.
- 🌐 20+ LLM providers — OpenAI, Claude, Gemini, DeepSeek, Grok, and more. Automatic fallback.
git clone https://github.com/grasberg/sofia.git
cd sofia
make deps && make build
./build/sofia onboard # Initialize config & workspace
./build/sofia gateway # Start the web UI
# Open http://127.0.0.1:18795 → Models tab → add your API key → start chattingPrerequisites: Go 1.26+ (download)
Register domain names, publish web pages, read/edit files, run bash commands, manage Google Services (Gmail/Calendar) — all without human intervention.
Tiered memory system — short-term, long-term, episodic, and semantic (knowledge graph). Automatic consolidation and strategic forgetting keep memory efficient.
Delegate tasks to parallel agents. AgentArchitect creates and optimizes new agents on the fly. A2A mailbox messaging with broadcast.
OpenAI, Anthropic (Claude 4.5), Gemini, DeepSeek, Grok, MiniMax, Moonshot, Qwen, GitHub Copilot, and more. Automatic fallback chains with exponential backoff.
126+ built-in skills with expert personas. Automatic skill creation, refinement, and knowledge distillation. ClawHub marketplace for community skills.
5-phase EvolutionEngine: observe → diagnose → plan → act → verify. Post-task evaluation, prompt self-optimization, and code self-modification.
Autonomous computer control via screenshots and vision-LLM. Control mouse and keyboard on macOS and Linux.
Playwright-based web browsing. Click, fill forms, take screenshots, extract text — all autonomously.
Input validation, budget management, prompt injection defense (35+ patterns in 6 languages), PII detection, and action confirmation for high-risk operations.
📋 More Features
- Smart Tool Management — Semantic tool matching via embeddings, performance tracking, dynamic tool composition
- Lightning-Fast Execution — Caching of tool definitions, prompts, and connection pooling for low latency
- Autonomy & Proactivity — Long-term goals, context-aware triggers, proactive suggestions, self-initiated research
- MCP Client — Model Context Protocol support for external MCP servers and tools
- Gateway Mode — Built-in support for Telegram, Discord, Email, Web, and CLI
- Plan & Execute — Structured task planning with step-by-step tracking and sub-plans
- Shared Scratchpad — Key-value storage for agent-to-agent communication
- Cron Scheduling — Independent task scheduling with at/every/cron patterns
- Modern Web UI — Brutalist design with CRT effects, real-time updates, and file uploads
Evaluated by automated code analysis across 12 capability dimensions.
| Category | Score | Highlights |
|---|---|---|
| Multi-Agent Orchestration | 9/10 | Dependency-aware scheduling, semantic delegation, dynamic agent spawning, A2A messaging |
| Tool Use & Execution | 9/10 | 40+ tools, embeddings-based filtering, parallel execution, circuit breaker |
| Context & Memory | 9/10 | Knowledge graph, 4-layer memory, strategic forgetting, prompt cache |
| Planning & Reasoning | 9/10 | Hierarchical plans, doom loop detection, evaluation loop, auto-checkpoint |
| Safety & Guardrails | 9/10 | 35+ injection patterns, PII detection, secret scrubbing, AES-256-GCM |
| Provider Abstraction | 9/10 | 20+ providers, fallback chains, Bayesian quality ranking, retry with jitter |
| Channel Integration | 8/10 | Telegram, Discord, Email, Web, CLI with shared retry logic |
| Observability & Ops | 9/10 | SQLite audit logging, distributed tracing, health/metrics endpoints |
| Self-Improvement | 8/10 | 7-phase evolution loop, SafeModifier, performance-driven agent retirement |
| Skills & Extensibility | 8/10 | 126 skills, 40 agent templates, ClawHub marketplace, lazy loading |
| Evaluation & Testing | 8/10 | LLM-as-judge scoring, A/B comparison, 5 benchmark suites |
| Scheduling & Automation | 8/10 | 3 schedule types, heartbeat, goal-driven autonomy, context triggers |
| Avg: 8.6 |
┌─────────────────────────────────────────────────────────┐
│ Channels (Input) │
│ Telegram · Discord · Email · Web UI · CLI │
└──────────────────────────┬──────────────────────────────┘
│
┌──────▼──────┐
│ Gateway │ ← HTTP/WebSocket/REST
└──────┬──────┘
│
┌────────────┼────────────────┐
│ Orchestrator │
│ ┌─────────┴──────────┐ │
│ │ Evolution Engine │ │
│ │ (Self-Improvement) │ │
│ └────────────────────┘ │
│ │
│ ┌──────────┐ ┌──────────┐ │
│ │ Agent 1 │ │ Agent 2 │ │
│ │ (A2A) │ │ (A2A) │ │
│ └─────┬────┘ └─────┬────┘ │
└────────┼─────────────┼────────┘
│ │
┌────────┴─────────────┴────────┐
│ Tool Layer │
│ File · Shell · Web · Browser │
│ GitHub · Google · Bitcoin │
│ Cron · Memory · Plan · MCP │
└───────────────┬────────────────┘
│
┌─────────▼─────────┐
│ LLM Providers │
│ 20+ with fallback │
└───────────────────┘
| Sofia | AutoGPT | CrewAI | LangChain | OpenDevin | |
|---|---|---|---|---|---|
| Runs 100% locally | ✅ | ❌ | ❌ | ❌ | ❌ |
| Single binary | ✅ | ❌ | ❌ | ❌ | ❌ |
| 40+ built-in tools | ✅ | Limited | Via plugins | Via plugins | Limited |
| Knowledge graph | ✅ | Basic | ❌ | ❌ | ❌ |
| Self-improving | ✅ | ❌ | ❌ | ❌ | ❌ |
| Multi-agent orchestration | ✅ | Basic | ✅ | Basic | ❌ |
| Browser automation | ✅ | ❌ | ❌ | ❌ | ❌ |
| Computer use | ✅ | ❌ | ❌ | ❌ | ❌ |
| 20+ LLM providers | ✅ | Limited | Limited | ✅ | Limited |
| Security guardrails | ✅ | ❌ | ❌ | ❌ | ❌ |
| Language | Go | Python | Python | Python | Python |
~/.sofia/
├── config.json # LLM providers, model selection, preferences
├── workspace/
│ ├── IDENTITY.md # Who Sofia is
│ ├── SOUL.md # Personality & behavior rules
│ ├── USER.md # User preferences
│ ├── skills/ # 126+ skill definitions
│ └── agents/ # 40+ agent templates
├── db/ # SQLite databases (memory, audit, goals)
└── logs/ # Execution logs
- Plugin SDK — Third-party tool development kit
- Voice Interface — Speech-to-text and text-to-speech
- Mobile Companion — iOS/Android control app
- Team Mode — Multi-user collaboration
- Cloud Sync — Optional encrypted cloud backup
- Marketplace — Community skills and agents marketplace
We love contributions! See CONTRIBUTING.md for guidelines.
Whether you are fixing a bug, adding a feature, improving docs, or sharing a skill — every contribution matters.
Found a vulnerability? Please see SECURITY.md for responsible disclosure.
Sofia is released under the MIT License.
⭐ Star us on GitHub · 🐛 Report a Bug · 💬 Join the Discussion
Made with ❤️ by the Sofia community