microclaw

MicroClaw

English · 简体中文 · हिन्दी · Español · العربية · Français · বাংলা · Português · Bahasa Indonesia · اردو

[!IMPORTANT] For production use, choose the stable branch. main moves quickly and may include breaking changes. One embeddable Rust agent core. Two products, one SDK.

Run MicroClaw Server, use the native Work app, or embed the same Agent Engine in your own Rust application.

Quick start · Rust SDK · Why MicroClaw · Capabilities · Documentation

MicroClaw is a self-hosted Rust agent platform with two product surfaces and an embeddable SDK. MicroClaw Server runs continuously for chat channels, Web, APIs, scheduling, and automation. MicroClaw Work is a native GPUI desktop application for local, workspace-centered agent work. Other Rust applications can use microclaw-sdk to run the same channel-independent Agent Engine, tools, policy, memory, skills, and runtime events.

It is designed for work that lasts longer than one request: multi-step tool use, resumable sessions, durable delivery, persistent memory, scheduled tasks, and governed extensions all run in the same runtime.

Read the MicroClaw Work product guide for the local task loop, platform support levels, native settings, safety boundary, and packaging model. The current stable delivery baseline is MicroClaw v0.6.1.

Quick start

Choose the path that matches what you are building:

Install the native MicroClaw Work desktop app on Apple Silicon macOS 13+:

brew tap microclaw/tap
brew install --cask microclaw-work

Linux x86_64/arm64 and Windows x86_64 portable previews are available from the v0.6.1 release. macOS remains the officially supported Work desktop platform while the preview builds complete platform acceptance.

To run MicroClaw Server, install on macOS or Linux:

curl -fsSL https://microclaw.org/install.sh | bash

On Windows PowerShell:

iwr https://microclaw.org/install.ps1 -UseBasicParsing | iex

Configure and start:

microclaw doctor
microclaw setup
microclaw start

Then open http://127.0.0.1:10961.

The latest release is v0.6.1. It keeps the Server, Work, and Rust SDK contract from 0.6 while making releases faster and deterministic through reusable Rust caches, an explicit artifact manifest, and current GitHub Actions runtimes. See the release notes and downloads.

For Homebrew, Docker, source builds, Linux compatibility, upgrades, and service installation, see the getting-started guide.

Why MicroClaw

• One core, two product surfaces. Server and Work share the same Agent Engine, provider layer, tools, memory, policy, and recovery model.
• An embeddable Rust SDK. Applications can use stable Agent, Run, event, control, and Worker contracts without depending on Server, Web, channel, or desktop UI code.
• A local-first native Workspace. Work shows project and Git branch context, supports native Workspace-file drag and drop, and keeps attachment access inside the shared runtime's folder guards.
• Execution that can continue. Sessions, safe tool boundaries, scheduled work, and outbound delivery survive process restarts.
• Provider freedom. Use native Anthropic or a broad set of OpenAI-compatible and local providers through one internal message model.
• Extensible by design. Add skills, MCP servers, plugins, hooks, tools, or channel adapters without replacing the core runtime.
• Operable on modest infrastructure. A Rust service and embedded SQLite keep deployment straightforward; no separate vector database or service mesh is required.
• Safety is part of execution. Tool risk gates, scoped grants, egress controls, sandboxing, redaction, and audit trails use shared enforcement points.

Reliability you can verify

Read the durable coworker guide and reliability proof report for the concrete recovery model. Every release publishes a machine-readable reliability scorecard as a release asset — reproduce it locally with scripts/ci/reliability_scorecard.sh.

How it works

Every message follows the same flow:

1. Resume session state and load relevant memory, skills, and runtime context.
2. Call the selected model using provider-neutral messages and tool definitions.
3. Execute tools through shared guardrails, hooks, and authorization checks.
4. Continue until completion, then persist the turn and deliver the result durably.

Server channel adapters translate ingress and delivery events only. Work projects the same runtime events into native GPUI state through microclaw-work-runtime and microclaw-work-app. Neither surface carries a separate agent loop or provider implementation.

Embed MicroClaw in Rust

microclaw-sdk exposes the shared run lifecycle to other Rust applications. Its full preset includes the configured Agent Engine, Skills, tools, MCP, memory, hooks, Subagents, and Local Worker support without pulling in the Server, Web console, concrete channel adapters, or Work UI.

[dependencies]
microclaw-sdk = { version = "0.6.1", features = ["full"] }

use microclaw_sdk::{FullRuntimeConfig, MicroClaw};

#[tokio::main]
async fn main() -> Result> {
    let config = FullRuntimeConfig::new("openai", "gpt-5", std::env::var("OPENAI_API_KEY")?);
    let microclaw = MicroClaw::configure(config).build().await?;
    let result = microclaw.agent("assistant").build()?.run("Plan this task").result().await?;
    println!("{}", result.final_text);
    Ok(())
}

Use MicroClaw::builder for YAML configuration or MicroClaw::configure for an entirely programmatic setup, inspect the Skill catalog, build an Agent with selected Skills, and consume ordered runtime events plus the terminal RunResult. See the compiling configured_skilled_agent example and the SDK quickstart, concept guide, Skills guide, and Worker guide. The public Rust packages are available on crates.io as microclaw-core, microclaw-engine, and microclaw-sdk.

Capabilities

Channels and providers

The runtime includes adapters for Telegram, Discord, Slack, Feishu/Lark, WeChat, DingTalk, QQ, WhatsApp, Signal, Matrix, IRC, Nostr, iMessage, email, and Web. Matrix is enabled by the full feature; availability and setup requirements vary by platform.

Anthropic has a native provider path. OpenAI, OpenAI Codex, OpenRouter, Ollama, Google, DeepSeek, local servers, and many other services use the shared OpenAI-compatible path. The generated provider matrix is the source of truth for current IDs, protocols, defaults, and models.

Install options

The prebuilt Linux binary has glibc and OpenSSL requirements. Read Getting started before installing on an older distribution.

MicroClaw Work is officially supported on Apple Silicon macOS. Linux and Windows portable previews are published automatically with each release while native platform acceptance continues. MicroClaw Server remains supported on all three platforms.

Documentation

Start with the documentation map. It separates everyday use, extension, operations, security, architecture, and contributor material so this page can remain a focused project entry point.

| Need | Canonical source