MesoClaw logo

MesoClaw

app-v0.0.10MITRust 2024

A lightweight, secure, local-first AI agent for desktop, CLI, and daemon

A miniaturized version of OpenClaw, built for desktop, CLI, and mobile (coming soon) and edge devices (coming soon). 15 built-in agent tools, 18 LLM providers, 3 messaging channels, persistent memory, scheduling, and a local gateway API. Under 50 MB idle RAM. Offline via Ollama.

terminal
$git clone https://github.com/sprklai/mesoclaw.git
$cd mesoclaw
$bun install
$bun run tauri dev
15 built-in agent tools
18 LLM providers supported
3 messaging channels
999 backend tests passing
Under 50 MB idle RAM
Under 3 second startup
4 Rust crates
88 gateway routes
10 platform targets
14 compile-time feature flags
15
Agent Tools
18
LLM Providers
999+
Backend Tests
<50 MB
Idle RAM

Built for real desktop AI work

Native Desktop App

Built with Tauri and Rust for a lightweight, secure native experience. ~20 MB binary, under 50 MB idle RAM.

Local Gateway API

88 HTTP and WebSocket routes exposing agent capabilities to local scripts, CLI tools, and browser extensions.

Persistent Memory

SQLite-backed memory with semantic search. Conversations, tool results, and context survive restarts.

15 Built-in Tools

Web search, sysinfo, shell, file read/write/list/search, patch, process management, learn, skill proposal, and more.

Task Scheduling

Cron-style and one-shot scheduled tasks. Agents run autonomously on your schedule, even offline via Ollama.

3 Messaging Channels

Telegram, Slack, and Discord integrations via feature-gated builds. Control your agent from any messaging platform.

Offline via Ollama

Run local models through Ollama for complete offline operation. No data leaves your machine.

18 LLM Providers

18 providers via rig-core: OpenAI, Anthropic, Google, Ollama, and more. Switch providers per-task.

Soul & Persona System

Configurable agent personality and behavior profiles. Define how your agent communicates and approaches tasks.

Skills System

Self-evolving agent capabilities through learned skills. Agents propose and acquire new skills over time.

Internationalization

Built-in i18n support via paraglide-js. Multi-language interface out of the box.

Plugin System

JSON-RPC 2.0 plugin architecture. Language-agnostic external plugins via stdio for extensibility.

Progressive User Learning

Agent learns your preferences with observations, confidence scoring, TTL, and privacy controls.

Where MesoClaw fits

FeatureMesoClawOpenClawPicoClawZeroClaw
LanguageRust + SvelteTypeScriptGoRust
InterfaceNative Desktop GUICLI / WebUICLICLI
RAM~50 MB>1 GB<10 MB<5 MB
Startup<3s>500s (edge)<1s<10ms
LLM Providers18 via rig-core5+12+Plugin-based
Channels3 built-in15+ (plugins)10+Plugin-based
Agent Tools15 built-in40+ (skills)Web + FilePlugin-based
DeploymentTauri binaryNode.js daemonSingle binarySingle binary
Tests999 total70% coverageMinimal1,017
Desktop GUIFull desktop appWeb Control UINoneNone

6-layer security defense

LayerDescription
Credentials
OS keyring with zeroize memory protection. No plaintext secrets on disk.
Autonomy
Three autonomy modes: Supervised, Autonomous, Strict. Configurable per-session.
Filesystem
Allowlist/blocklist path rules. Agents cannot access paths outside configured boundaries.
Injection
Prompt injection detection heuristics applied before every LLM call.
Rate Limiting
Per-provider and per-tool rate limits prevent runaway costs and abuse.
Audit Trail
Every tool invocation, LLM call, and file access is logged to a local audit database.

Autonomy Modes

Strict

Agent operates within tightly constrained boundaries. Minimal autonomy, maximum oversight.

Supervised

Agent proposes actions. User confirms or rejects each one before execution.

Autonomous

Agent executes autonomously within configured boundaries. Use with caution.

All data stays on your machine. No telemetry, no cloud sync, no account required. MesoClaw is app-v0.0.10, actively developed.

Get started in minutes

  1. 1Visit the releases page and download the installer for your platform.
  2. 2Run the installer (.dmg for macOS, .msi for Windows, .deb/.rpm for Linux).
  3. 3Launch MesoClaw from your applications menu.
  4. 4Configure your first LLM provider in Settings.
  5. Go to releases

Frequently asked questions

Getting Started

MesoClaw itself is free and MIT-licensed. However, you'll need API keys from LLM providers like OpenAI or Anthropic if you want to use cloud models. For a completely free setup, use Ollama to run open-source models locally at no cost.

MesoClaw stores credentials locally using encrypted storage. On first launch, the settings panel lets you add API keys for any of the 18 supported providers. Keys never leave your machine and are encrypted at rest.

MesoClaw is a standalone desktop agent, not a drop-in replacement for OpenClaw. While they share the Claw ecosystem philosophy, MesoClaw is optimized for single-user desktop workflows rather than OpenClaw's server-based multi-user setup. Your conversations and data would start fresh.

Practical Usage

Yes. MesoClaw supports cron-style and one-shot scheduled tasks. You can set up recurring jobs — like checking a website, summarizing emails, or running reports — and the agent executes them autonomously in the background, even with the GUI minimized.

Absolutely. MesoClaw lets you switch LLM providers per-task. Use a fast, cheap model for simple lookups, a powerful model for complex reasoning, or a local Ollama model when you're offline — all from the same interface.

The local Gateway API exposes 88 HTTP and WebSocket routes on your machine. Developers and power users can use it to integrate MesoClaw into scripts, CLI tools, browser extensions, or custom workflows — essentially turning MesoClaw into a local AI backend.

Trust & Control

MesoClaw's autonomy modes give you full control. In Strict mode, every action requires your approval. In Supervised mode, routine actions proceed automatically while sensitive ones (file deletion, shell commands) require confirmation. Autonomous mode is available for trusted, sandboxed workflows.

Yes. MesoClaw has filesystem access rules that let you whitelist or blacklist directories. Combined with prompt injection detection and rate limiting, you can tightly scope what the agent is allowed to touch.

Long-term

MesoClaw is distributed as a standalone binary. Updates are released on GitHub — you can download new versions directly or rebuild from source. Your local data (conversations, memory, settings) persists across updates.

MesoClaw is currently at v0.0.10 and under active development. It has 999+ backend tests and 4 Rust crates with 14 compile-time feature flags, indicating solid engineering foundations. Check the GitHub roadmap for upcoming milestones and stability targets.