THE PERSONAL AI AGENT
YOU ACTUALLY OWN.
One lightweight Rust binary on your machine, with your keys or your local models. It reaches 30+ channels, asks before risky actions, and can stamp every tool call with a cryptographic receipt. No cloud seat. No subscription.
70+ LLM Providers
30+ Channels
Most AI assistants are a seat you rent on someone else's machine. ZeroClaw is a binary you own on yours.
It uses less memory than a browser tab and runs on hardware as small as a Raspberry Pi — the same agent on your workstation, your homelab, and boards it can actually touch (GPIO on Pi, STM32, Arduino, ESP32).
Running in under a minute
Install the binary, point it at a provider, connect a channel. The one-liner grabs a prebuilt binary — a measured 14 seconds — and compiling from source is strictly opt-in.
Install ZeroClaw
One command downloads the prebuilt native binary — no Rust toolchain needed. Homebrew and Docker work too.
Run Quickstart
Create your first agent end-to-end: pick a provider (local Ollama or your own key) and connect a channel in one guided pass.
Start Agent Mode
Run interactive chat or one-shot prompts with tool + memory support.
Run Runtime Services
Start gateway/daemon for channel and webhook workflows, then validate with status/doctor.
An agent with shell access you can audit
ZeroClaw is security-first by default: supervised autonomy, OS-level sandboxes, command policy — and opt-in receipts that make every tool call provable.
Supervised autonomy
The default mode asks before medium-risk actions and blocks high-risk ones. Autonomy is something you grant, not something you hope to contain.
OS sandboxes
Landlock, Bubblewrap, Seatbelt, or Docker contain what the agent can touch — enforced by the operating system, not by a prompt.
Command policy
Explicit allowlists and workspace scoping decide what runs and where. Medium-risk actions wait for your approval; high-risk ones are blocked outright.
Tool receipts
Flip on receipts and every successful tool call returns with an HMAC-SHA256 tag. A run the model claims but never made — or a result it invented — produces a missing or invalid receipt.
YOLO mode exists for trusted dev boxes. It is opt-in — never the default.
Built for Production
Security, portability, and low overhead — capabilities in the shipping binary, not the pitch deck.
Secure by Design
Pairing, strict sandboxing, explicit allowlists, and workspace scoping are enabled by default.
Fast Cold Starts
The core agent starts in milliseconds — cold-start speed is an architectural constraint, not a marketing claim.
No Lock-In
OpenAI-compatible provider support plus custom endpoints and pluggable integrations.
Lean Runtime
Uses less memory than a browser tab: a small core, extended by the components you choose.
Trait-Driven Core
Providers, channels, tools, memory, and tunnels are swappable with config changes.
Deploy Anywhere
One binary-first workflow across ARM and x86, from a Raspberry Pi to a cloud VM — plus Docker.
What can you build with ZeroClaw?
From a personal assistant on your phone to unattended automation in your homelab — private AI on your terms.
Personal AI Assistant
Your own AI that runs locally—Telegram, Discord, WhatsApp, and more, 24/7.
Bring your own API keys or run Ollama offline; one agent config syncs across every channel you enable.
Code Generation
Generate, review, and refactor code without exposing your codebase.
Data Analysis
AI-powered insights and summaries on your data, completely private.
Document Processing
Summarize, extract, and transform documents privately.
Secure by Design
Pairing, sandboxing, allowlists, and workspace scoping by default.
Fast Cold Starts
Single native binary—starts in milliseconds, no JVM or heavy runtime.
Ship one artifact per platform; cold boots stay predictable so agents feel instant in scripts and CI.
No Lock-In
OpenAI-compatible providers plus custom endpoints and pluggable integrations.
Task Automation
Schedules, webhooks, and chained actions—run agents unattended inside your guardrails.
Trigger from cron, HTTP, or channel events; every step stays inside your sandbox and allowlists.
Lean Runtime
Uses less memory than a browser tab — small core, pluggable components.
Research Assistant
Search, distill, and summarize sources — sessions stay private.
Content Creation
Generate blog posts, social media, and marketing copy on your hardware.
Customer Support Bot
Private AI frontline on Slack, Discord, and email—triage before humans step in.
Connect FAQs, hand off to your team, and keep transcripts off third-party analytics by default.
Development Roadmap
Curated from the live GitHub milestones. No dates — things ship when they're ready.
The v0.8 Runtime
The big rewrite is behind us; the v0.8.x series ships at a regular cadence
- Multi-agent runtime — many named agents from one daemon, each isolated
- zerocode terminal UI
- Security hardening — per-agent tool allowlists, sandboxing, tool receipts
- A2A discovery — agents describe and discover one another through the gateway
- Skills expansion — extra registries, typed slash-commands, plugin loading
Plugins + Operations
Easier to extend and operate seriously
- A fuller WASM plugin platform
- Richer operator management surfaces
- Easier ways to install, manage, and inspect plugins, skills, and tools
Durable Automation + Access
Automation you can supervise, resume, and audit
- A more durable SOP control plane — auditable procedures and approval gates
- Persisted run state and shared event fan-in across channels and triggers
- Pluggable authentication (OIDC, ssh-key, and more)
- Per-user, per-agent authorization and per-user isolation
v1.0
A stable 1.0 that's sandboxable and interoperable with other agents — the foundation for enterprise and commercial use
- Stable — an API/config/security contract you can build on
- Sandboxed — strong isolation per agent
- Swappable — agents discover and call each other (A2A)
- Enterprise-ready — authentication, authorization, auditability
Questions, answered
What is ZeroClaw?
ZeroClaw is a personal AI assistant you own — an agent runtime that runs as a single Rust binary on your own machine, with your own keys or your own local models. It talks to 70+ LLM providers, reaches 30+ channels like Telegram, Discord, Matrix, and email, and acts through tools such as shell, browser, HTTP, and hardware.
Is ZeroClaw free?
Yes. ZeroClaw is open source, dual-licensed MIT OR Apache-2.0. There is no subscription and no hosted seat — you pay only your own LLM provider costs, or nothing at all if you run a local model with Ollama.
How lightweight is it?
ZeroClaw is one native binary with no runtime to install — no Node, no JVM, no Python environment. The core agent starts in milliseconds and uses less memory than a browser tab; that's an architectural constraint the project holds itself to, not a marketing line. It runs on ARM and x86, down to a Raspberry Pi.
Does my data leave my machine?
The agent always runs on your machine. With a local model (like Ollama), the model runs there too, so nothing needs to leave the box. With a hosted provider, only the prompts you send go to the provider you chose, using your own key — there is no ZeroClaw cloud in between.
How is it secured?
Security is the default posture, not an add-on: autonomy starts in supervised mode (medium-risk actions need approval, high-risk are blocked), OS sandboxes (Landlock, Bubblewrap, Seatbelt, or Docker) contain what it can touch, and command allowlists and workspace scoping decide what runs. You can also enable cryptographic tool receipts, which stamp every successful tool call with an HMAC tag so a fabricated run or invented result is detectable.
What can it connect to?
30+ channels — including Telegram, Discord, WhatsApp, Slack, Signal, iMessage, email, Matrix, and webhooks — and 70+ LLM providers, from local Ollama, LM Studio, llama.cpp, and vLLM to Anthropic, OpenAI, or any OpenAI-compatible endpoint, with fallback routing between them.
Own your agent tonight.
One command, one binary, and an agent that answers to you — and only you.