Overview

AgentZero is self-growing AI agent infrastructure built entirely in Rust. Describe what you need in plain English — the system creates tools, defines agents, assembles swarms, and remembers what works. Every session makes the next one smarter.

Why AgentZero

Self-improving. Every session is recorded as a trajectory. The system tracks what works, auto-fixes failing tools, auto-improves successful ones, detects tool gaps, and learns from its own performance history. Month 2 is measurably smarter than month 1.
Natural language tools. Describe a missing tool in conversation. The system creates it (shell, HTTP, LLM, composite, or compiled WASM strategy), registers it immediately, and persists it forever. Tools evolve automatically based on quality metrics.
NL agent definitions. Say “an agent that reviews my PRs daily” and the system derives name, system prompt, keywords, allowed tools, and schedule automatically. Agents are stored encrypted and auto-routed to by keywords.
Autonomous swarms. Give a goal, get a parallel agent DAG. Each node gets only the tools it needs via per-node tool hints. Sandboxed execution with conflict detection and dead agent recovery. Durable inter-agent messaging with delivery guarantees.
Deploy anywhere. A single native binary runs on Linux, macOS, and Windows — 8 platform targets, Docker multi-arch. 4.5 MB embedded binary. Cost-aware model routing sends simple queries to cheap models automatically.
Secure by default. Encrypted storage, allowlists, policy gates, audit logging, secret redaction, Unicode injection defense, context file scanning, and emergency stop. Fail-closed from day one.

Features

Self-Improving

Session trajectories, auto-evolution, insights engine, tool gap detection. The system learns from every run and gets measurably better over time.

NL Tools & Agents

Describe tools and agents in plain English. Five execution strategies including compiled WASM. Auto-fix and auto-improve with quality tracking.

Autonomous Swarms

Goal decomposition into parallel agent DAGs. Durable message queues, A2A streaming, multi-turn tasks, and agent discovery.

Catalog Learning

Successful tool combos recorded as recipes. Failure patterns detected and fed back. Cost-aware routing sends simple queries to cheap models.

Small & Fast

4.5 MB embedded binary. 4-phase context compression cuts token costs. Credential pooling avoids rate limits. Prompt caching for Anthropic.

Secure by Default

Encrypted storage, deny-by-default tools, Unicode injection defense, context file scanning, checkpoint recovery, audit trails.

At a Glance


Language	100% Rust
Crates	16 workspace crates
Binary Size	4.5 MB embedded / 18 MB full
Platform Targets	8 (Linux x86/ARM/musl, macOS Intel/ARM, Windows)
Client SDKs	Python, TypeScript, Swift, Kotlin
CLI Commands	38+ subcommands
Providers	37+ (OpenAI, Anthropic, Google, Ollama, Candle local inference, and more)
Memory	SQLite (default, encrypted), Turso/libsql (optional)
Tools	58+ built-in (file I/O, shell, networking, browser, delegation, memory, git, cron, hardware, autopilot, and more)
MCP	First-class Model Context Protocol — each MCP tool registered individually with real schema
Channels	25+ (Telegram, Discord, Slack, Matrix, Email, WhatsApp, SMS, and more)
Plugins	WASM sandbox with integrity verification
Security	Encrypted storage, allowlists, OTP gates, Unicode injection defense, context file scanning, audit trail, secret redaction, estop
Multi-Agent	Delegation, swarm coordination, pipelines, durable message queues, A2A streaming, persistent named agents
Self-Improving	Trajectory recording, insights engine, auto-evolution, tool gap detection, session summarization
Cost Optimization	Complexity-based model routing, 4-phase context compression, credential pooling, prompt caching, pre-execution cost estimation
Autopilot	Autonomous agent loop — proposals, cap gates, missions, triggers, reaction matrices
API Docs	Interactive Scalar docs at /docs on every gateway deployment
License	MIT / Apache-2.0

Quick Start

# Build from source
cargo build -p agentzero --release

# Run interactive onboarding
cargo run -p agentzero -- onboard --interactive

# Set your API key
export OPENAI_API_KEY="sk-..."

# Send a message
cargo run -p agentzero -- agent -m "hello"

# Check system health
cargo run -p agentzero -- doctor models
cargo run -p agentzero -- status

See the full Installation guide and Quick Start for details.