SELF-GROWING AI AGENT INFRASTRUCTURE IN RUST

One binary.
Secure by default.
Infinite agents.

Production-grade AI agents that create their own tools, assemble into swarms, and get smarter with every run. 4.5 MB. 12-layer security. Pure Rust.

Say "summarize this video" — it decomposes the goal, creates missing tools, executes in parallel, and remembers what works. Next time, it's faster. Encrypted at rest, signed in transit, sandboxed at every layer. 🦀

4.5MB

Embedded Binary

8

Platform Targets

4

Client SDKs

58+

Built-in Tools

Built for production

Self-Growing

Create a tool on Monday, it's there on Friday. Define an agent once, it runs forever. Every capability you add is encrypted and persisted across sessions.

Natural Language Tools

Describe a tool in plain English. The system creates it (shell, HTTP, LLM, or composite), registers it immediately, and persists it forever.

NL Agent Definitions

Say 'an agent that reviews my PRs daily' — name, system prompt, keywords, tools, and schedule are derived automatically.

Autonomous Swarms

Give a goal, get a parallel agent DAG. Each node gets only the tools it needs. Sandboxed execution with conflict detection.

Small & Fast

4.5 MB embedded binary. Millisecond cold starts. Async Tokio runtime with streaming output.

Secure by Default

Defense-in-depth: encrypted storage, path canonicalization, DNS rebinding protection, PII redaction, prompt injection detection, credential leak guards, and audit logging. Fail-closed from day one.

Plugin Sandbox

WASM plugins run in a bounded sandbox with fuel metering, memory caps, and Ed25519 signature verification required in production.

Multi-Platform Channels

Telegram, Discord, Slack, and 20+ more. One agent, every platform. Add a channel with a single config line.

37+ LLM Providers

OpenAI, Anthropic, Google, Ollama, Candle local inference, and more. Swap providers with a config change — no code required.

Why AgentZero

Everything you need to run AI agents in production — nothing you don't.

Self-Growing Runtime

Every tool you create and every agent you define persists encrypted. The system you use in month 2 has everything you built in month 1 — nothing is lost.

NL Tool Creation

Describe a missing tool in conversation. The system creates it (shell, HTTP, LLM, composite), makes it available immediately, and persists it across sessions.

Catalog Learning

When a tool combo solves a task, it's saved as a recipe. Next time you ask something similar, those exact tools are selected automatically.

Single Binary

One static binary, zero runtime dependencies. 4.5 MB embedded, 18 MB full. Download and run.

Rust Memory Safety

No GC pauses, no null pointers, no data races. Compile-time guarantees eliminate entire classes of bugs.

Security First

12-layer defense-in-depth: XChaCha20-Poly1305 encryption, mTLS, HMAC request signing, PII/injection guardrails, credential leak detection, MCP binary attestation, and declarative YAML security policies.

37+ Providers

OpenAI, Anthropic, Google, Ollama, Candle local inference, and 30+ more. Swap providers with a config change.

8 Platform Targets

Linux (x86/ARM/musl), macOS (Intel/Apple Silicon), Windows. CI-tested on all three OS families.

MCP + A2A

Model Context Protocol for tool interop. Google A2A for agent-to-agent communication. Standards-first.

Autonomous Swarms

Give a goal in plain English. The system decomposes it into a parallel agent DAG, each node sandboxed with only the tools it needs.

Client SDKs

Thin HTTP/WebSocket SDKs for Python, TypeScript, Swift, and Kotlin. Any platform that can make HTTP calls can control AgentZero.

Interactive API Docs

Every gateway ships Scalar API docs at /docs. Explore and test all 40+ endpoints from the browser.

12-layer defense-in-depth

The most comprehensive security model in any AI agent framework.

File I/O

Path canonicalization, symlink resolution, hard-link detection, sensitive file blocking (40+ patterns including .env.*, .kube/config, *.pem)

Shell Execution

Context-aware quote parser, explicit command allowlist, backtick/dollar always blocked, output truncation

Encryption at Rest

XChaCha20-Poly1305 AEAD with random nonces, 256-bit keys, file permissions 0o600, atomic writes

Network Security

DNS rebinding protection, private IP blocking (RFC1918/link-local/carrier-grade NAT), constant-time token comparison

Credential Leak Guard

Pattern + Shannon entropy detection for 11+ secret formats, boundary isolation between local and remote channels, configurable patterns

LLM Guardrails

PII redaction (email, phone, SSN, API keys) and prompt injection detection enabled by default in audit mode

WASM Plugin Sandbox

Fuel metering, 256 MB memory cap, default-deny network/filesystem, Ed25519 signature required in release builds

Gateway Security

mTLS client certificate verification, HMAC-SHA256 request signing with replay protection, TLS enforced in production mode

MCP Attestation

SHA-256 binary hash verification before spawning MCP server subprocesses

Autonomy & Delegation

3-tier autonomy (ReadOnly/Supervised/Full), policy intersection for sub-agents, per-tool rate limits

Declarative Policies

YAML per-tool security policies with granular egress, filesystem, and command rules

Audit & Redaction

Structured audit events, automatic secret redaction in errors/logs/panics, path sanitization in LLM-facing messages

Read the full security documentation

Install in seconds

One command. One binary. No dependencies.

$ curl -fsSL https://raw.githubusercontent.com/auser/agentzero/main/scripts/install.sh | bash

Or install with Cargo: cargo install agentzero

How it works

01

Install

cargo install agentzero

Single binary, no runtime dependencies.

02

Onboard

agentzero onboard --interactive

Interactive wizard writes agentzero.toml.

03

Run Agent

agentzero agent -m "hello"

Send a message through the agent loop.

04

Start Gateway

agentzero gateway

HTTP/WebSocket gateway with chat, SSE streaming, and pairing.

Modular architecture

Self-Growing

  • NL tool creation (persists forever)
  • NL agent definitions (auto-routed)
  • Recipe learning (compounds over time)
  • Export/import tools between instances

Swap Anything

  • Provider trait (37+ LLMs)
  • MemoryStore trait (SQLite, Turso)
  • Tool trait (58+ built-in + dynamic)
  • Channel trait (25+ platforms)

Fully Autonomous

  • Goal decomposition to agent swarms
  • Per-node tool filtering
  • Sandboxed parallel execution
  • Dead agent recovery
  • WASM plugin sandbox

One binary. Every platform. Unbreakable security. 🦀

4.5 MB, 12 security layers, 37+ LLM providers, 58+ tools. Agents that build themselves and never forget. Deploy it and walk away.