ZeroClaw

ZeroClaw is an open-source AI agent runtime written in Rust and shipped as a single static binary of a few megabytes. It wires roughly twenty LLM providers to more than thirty messaging channels and a sandboxed tool layer, letting engineering teams run persistent autonomous agents on hardware they own rather than inside a vendor's cloud.

ZeroClaw is an open-source AI agent runtime written in Rust and distributed as a single static binary, positioned explicitly as a lightweight alternative to Node.js-based agent frameworks such as OpenClaw. Its architecture is deliberately three-layered: a provider layer that speaks to Anthropic, OpenAI, Ollama and roughly twenty other LLM endpoints with configurable fallback chains; a channel layer with more than thirty inbound adapters including Discord, Telegram, Matrix, email, voice, webhooks and a CLI; and a tool layer that executes shell commands, browser automation, HTTP calls, hardware I/O and custom Model Context Protocol servers. Persistent memory is SQLite-backed with vector search for semantic recall. The stated default is supervised autonomy — medium-risk operations require approval and high-risk operations are blocked — enforced through workspace boundaries, command allowlists, OS-level sandboxes (Landlock, Bubblewrap, Seatbelt or Docker) and cryptographic tool receipts written for every action. Published performance figures are a 3.4 MB binary, sub-10 ms startup on a 0.8 GHz core and under 5 MB resident memory, which is what lets it run continuously on a Raspberry Pi or comparable low-cost hardware; GPIO, I2C, SPI and USB support extend that to STM32, Arduino and ESP32 boards. Development is fast-moving: the zeroclaw-labs repository carries roughly 32,500 stars and 440 open issues, is dual-licensed MIT or Apache-2.0, and shipped v0.8.1 through v0.8.4 between June and August 2026 — a two-to-three week cadence in which each release lands 150 to 380 commits, recently adding an SOP engine, a WebAssembly plugin host and A2A agent discovery.

Category

AI Agents & Orchestration

Pricing

Free, Open source

Target Market

CTOs, Enterprise Developers, Platform Engineers, DevOps Engineers, Security Engineers

Deployment

Self-hosted, Open-source, Edge-first

Founded

2026

• ✓
  Provider-agnostic LLM layer

  Speaks to Anthropic, OpenAI, Ollama and roughly twenty more endpoints with fallback chains, so a provider outage or price change does not require a rewrite.

• ✓
  30+ channel adapters

  Discord, Telegram, Matrix, email, voice, webhooks and CLI all reach the same agent, so one runtime serves every place your team already talks.

• ✓
  Sandboxed tool execution

  Shell, browser, HTTP and hardware tools run behind Landlock, Bubblewrap, Seatbelt or Docker with workspace boundaries and command allowlists limiting blast radius.

• ✓
  Supervised autonomy with approval gates

  Medium-risk operations require human approval and high-risk operations are blocked by default, inverting the usual opt-in-to-safety posture of agent frameworks.

• ✓
  Cryptographic tool receipts

  Every action the agent takes is signed and recorded, giving an after-the-fact audit trail of exactly what an autonomous process did.

• ✓
  SQLite-backed memory with vector search

  Persistent semantic recall ships in the binary itself, so no separate vector database or managed memory service is required.

• ✓
  Hardware and edge I/O

  Native GPIO, I2C, SPI and USB support targets Raspberry Pi, STM32, Arduino and ESP32, extending agents into physical control loops.

• ✓
  MCP and ACP support

  Custom Model Context Protocol servers add tools, and Agent Client Protocol over JSON-RPC 2.0 exposes the agent inside IDEs and editors.

• •
  Private always-on assistant

  Runs locally on owned hardware so chat history, documents and API keys never transit a third-party agent platform.

• •
  Scheduled operational reporting

  Built-in cron and SOP engine generate recurring digests such as weekly tech-trend summaries and push them to Slack or WhatsApp.

• •
  Multi-channel team bot

  One deployment answers questions in Discord, Telegram and email, sharing memory and tool access across every channel.

• •
  Edge and homelab automation

  Sub-5 MB memory footprint lets an agent run permanently on a Raspberry Pi driving GPIO-attached sensors and relays.

• •
  Draft-and-approve content workflow

  Agent drafts social or blog copy and holds it at an approval gate so a human signs off before anything publishes.

• •
  Embedded coding agent

  ACP support surfaces the runtime inside an IDE, giving editors an agent that can run shell and HTTP tools under policy.

There is no licence fee and no paid tier — the project is dual-licensed MIT or Apache-2.0 and every feature ships in the binary. Real cost is inference tokens billed by whichever LLM provider you configure, plus the hardware you run it on, which the project claims can be as little as a $10 single-board computer. There is no vendor-hosted option, so the operational cost is your own engineering time.

This page was written from 5 sources, 5 on domains other than zeroclawlabs.ai.

1. 1.github.com — zeroclaw
2. 2.github.com — releases
3. 3.docs.zeroclawlabs.ai — introduction
4. 4.pinggy.io — zeroclaw lightweight openclaw alternative
5. 5.digital-loop.com — agent based ai my exeprience with zeroclaw