GitHub - Marsstein/marsclaw-rs: The fastest AI agent runtime. Written in Rust. 5MB binary, 3MB RAM, <10ms cold start, zero dependencies.

Quick Start · Comparison · Features · Architecture · Configuration

5MB binary · ~3MB RAM · <10ms cold start · Zero CVEs · Zero dependencies

MarsClaw is a multi-agent AI runtime written in Rust. It connects to Claude, GPT, Gemini, and local models to help you code, automate tasks, and orchestrate multi-agent workflows — all from a single binary with no dependencies.

$ marsclaw "add error handling to main.rs"

> read_file
✓ read_file
> edit_file
✓ edit_file

Added error wrapping with anyhow to all three return paths in main().

── claude-sonnet-4 │ 1.2K in / 523 out │ $0.012 session ──

Quick Start

# Install
cargo install marsclaw

# Or download a binary
curl -sSfL https://marsclaw.dev/install.sh | sh

# Set your API key (pick one)
export ANTHROPIC_API_KEY="sk-ant-..."   # Anthropic
export GEMINI_API_KEY="..."             # Google Gemini
export OPENAI_API_KEY="sk-..."          # OpenAI
# Or use Ollama for free, fully offline — no key needed

# Interactive mode
marsclaw

# Single prompt
marsclaw chat "explain this codebase"

# Use with different providers
marsclaw -m claude-sonnet-4-20250514 "review this PR"
marsclaw -m gemini-2.5-flash "explain this code"
marsclaw -m gpt-4o "write tests for auth.rs"
marsclaw -m llama3.1 "refactor this function"    # Ollama, free

# Web UI
marsclaw serve --addr :8080

# Chat bots
marsclaw telegram        # Telegram bot
marsclaw discord         # Discord bot
marsclaw slack           # Slack bot

# Setup wizard
marsclaw init

Comparison

Performance & Footprint

	Claude Code	Aider	Goose	Cursor	OpenHands	MarsClaw
Language	TypeScript	Python	Python	Electron	Python	Rust
Install size	~200 MB	~150 MB	~120 MB	~400 MB	~2 GB	5 MB
Memory (idle)	~150 MB	~120 MB	~100 MB	~500 MB	~1 GB	~3 MB
Cold start	~3s	~2s	~2s	~5s	~10s	<10ms
Runtime deps	Node.js	Python + pip	Python + pip	Chromium	Docker	0
Single binary	No	No	No	No	No	Yes
Known CVEs	Inherits npm	Inherits pip	Inherits pip	Chromium	Docker	0

LLM Providers

	Claude Code	Aider	Goose	Cursor	OpenHands	MarsClaw
Anthropic (native)	Yes	Yes	Yes	Yes	Yes	Yes
OpenAI	No	Yes	Yes	Yes	Yes	Yes
Gemini	No	Yes	Yes	No	Yes	Yes
Ollama (offline)	No	Yes	Yes	No	Yes	Yes
OpenAI-compatible	No	Yes	Yes	No	Yes	Yes

MarsClaw supports 4 native providers + any OpenAI-compatible endpoint (Groq, Together, DeepSeek, Azure, vLLM, LM Studio).

Multi-Agent Orchestration

	Claude Code	Aider	Goose	Cursor	OpenHands	MarsClaw
Pipeline (sequential)	No	No	No	No	Yes	Yes
Parallel (fan-out)	No	No	No	No	No	Yes
Debate (adversarial)	No	No	No	No	No	Yes
Supervisor (coordinator)	No	No	Yes	No	Yes	Yes
Sub-agent delegation	No	No	No	No	No	Yes

Chat Channels

	Claude Code	Aider	Goose	Cursor	OpenHands	MarsClaw
Telegram	No	No	No	No	No	Yes
Discord	No	No	No	No	No	Yes
Slack	No	No	No	No	No	Yes
WhatsApp	No	No	No	No	No	Yes
Instagram	No	No	No	No	No	Yes

Platform Features

	Claude Code	Aider	Goose	Cursor	OpenHands	MarsClaw
Web dashboard	No	No	No	Yes	Yes	Yes (embedded)
MCP client	Yes	No	Yes	No	No	Yes
Persistent memory	No	No	No	No	No	Yes
Skills / prompt packs	No	No	No	No	No	Yes (5 + installable)
Scheduled tasks	No	No	No	No	No	Yes (cron)
Cost tracking	Yes	Yes	No	No	No	Yes (daily/monthly)
Credential scanning	No	No	No	No	No	Yes
Tool approval	Yes	No	No	No	No	Yes (per danger level)
Session persistence	No	Yes	No	No	Yes	Yes (SQLite)
Hook system	No	No	No	No	No	Yes
Offline mode	No	Yes	Yes	No	No	Yes (Ollama)
Self-hosted	No	Yes	Yes	No	Yes	Yes
Open source	No	Yes	Yes	No	Yes	Yes (Apache-2.0)

Features

LLM Providers

Connect to any major LLM provider — switch with a single flag:

Anthropic — Claude Opus, Sonnet, Haiku (native Messages API with streaming)
OpenAI — GPT-4o, GPT-4, o1 (+ any OpenAI-compatible endpoint)
Google Gemini — Gemini 2.5 Flash, Pro
Ollama — Llama 3, Mistral, CodeLlama, any local model (free, fully offline)
Any OpenAI-compatible API — Groq, Together, DeepSeek, Azure, vLLM, LM Studio

Multi-Agent Orchestration

Four built-in patterns for complex workflows:

Pipeline — chain agents sequentially, output of one feeds the next
Parallel — fan-out tasks to multiple agents, aggregate results
Debate — adversarial multi-round discussion between agents with a judge
Supervisor — coordinator agent delegates subtasks to specialist agents

Built-in Tools (7)

Tool	Description
`read_file`	Read files with line ranges
`write_file`	Create or overwrite files
`edit_file`	Surgical find-and-replace edits
`shell`	Execute shell commands with timeout
`list_files`	Recursive directory listing with glob
`search`	Content search across files (regex)
`git`	Read-only git operations (log, diff, status)

Channel Integrations (5)

Deploy your agent to any messaging platform:

Telegram — long-polling bot with /start, /clear, /help commands
Discord — Gateway WebSocket with real-time messaging
Slack — Socket Mode with event-driven responses
WhatsApp — Cloud API webhook (auto-mounts on serve)
Instagram — Messenger API integration

Platform

Web dashboard — embedded single-page UI, zero frontend build needed
Skills system — 5 built-in prompt packs (coder, devops, writer, analyst, compliance) + installable
Scheduler — cron expressions and interval-based task automation
MCP support — JSON-RPC 2.0 client for Zapier, n8n, filesystem, custom MCP servers
Persistent memory — episodic, semantic, and procedural memory backed by SQLite
Hook system — before/after tool calls, LLM calls, and error events
Security — credential scanning, path traversal guards, per-danger-level tool approval
Cost tracking — per-model pricing with daily and monthly budget limits
SQLite persistence — conversation history and sessions with zero config
Config — YAML file + MARSCLAW_* env var overrides

CLI Reference

marsclaw [OPTIONS] [COMMAND]

Commands:
  chat              Chat interactively or run single prompt
  serve             Start HTTP server + Web UI
  telegram          Run as Telegram bot
  discord           Run as Discord bot
  slack             Run as Slack bot
  whatsapp          Run WhatsApp webhook bot
  channels add      Connect a messaging channel
  channels list     Show configured channels
  channels remove   Remove a channel
  skills list       Show available skills
  skills install    Install a skill from URL
  skills use        Set the active skill
  init              Interactive setup wizard

Options:
  -c, --config      Config file path
  -m, --model       Override model (e.g., claude-sonnet-4-20250514)
  -v, --verbose     Debug logging
  -h, --help        Print help
  -V, --version     Print version

Architecture

src/                        10,500+ lines of Rust
  main.rs                   CLI entry point (clap)
  agent/                    Agent loop, context builder, sub-agent orchestrator
  bots/                     Telegram, Discord, Slack, WhatsApp + channel management
  config/                   YAML + env var configuration + setup wizard
  llm/                      4 providers (Anthropic, OpenAI, Gemini, Ollama) + cost + retry
  platform/                 Memory, hooks, MCP, scheduler, skills, security
  server/                   HTTP server (axum) + embedded Web UI + terminal REPL
  store/                    SQLite persistence (rusqlite)
  tool/                     7 built-in tools (read, write, edit, shell, list, search, git)
  types/                    Shared types and traits

Configuration

Config lives at ~/.marsclaw/config.yaml:

providers:
  default: anthropic
  anthropic:
    api_key_env: ANTHROPIC_API_KEY
    default_model: claude-sonnet-4-20250514
  gemini:
    api_key_env: GEMINI_API_KEY
    default_model: gemini-2.5-flash
  openai:
    api_key_env: OPENAI_API_KEY
    default_model: gpt-4o
  ollama:
    default_model: llama3.1

agent:
  max_turns: 25
  enable_streaming: true
  temperature: 0.0

cost:
  daily_budget: 10.0
  monthly_budget: 100.0

security:
  scan_credentials: true
  path_traversal_guard: true

# MCP servers
mcp:
  - name: n8n
    command: npx
    args: ["-y", "@anthropic/mcp-n8n", "--webhook-url", "http://localhost:5678"]

# WhatsApp webhook (auto-mounts on serve)
whatsapp:
  phone_number_id: "123456789"
  access_token: "EAAx..."
  verify_token: "marsclaw_verify"

# Scheduled tasks
scheduler:
  tasks:
    - id: daily-report
      name: "Daily Summary"
      schedule: "0 9 * * *"
      prompt: "Generate a daily summary of recent changes"
      channel: log

Environment variables override config: MARSCLAW_PROVIDER=ollama, MARSCLAW_MODEL=llama3.1, etc.

Contributing

git clone https://github.com/Marsstein/marsclaw-rs.git
cd marsclaw-rs
cargo build             # debug build
cargo test              # run 43 tests
cargo clippy            # lint
cargo build --release   # optimized 5MB binary

License

Apache-2.0

Name		Name	Last commit message	Last commit date
Latest commit History 13 Commits
assets		assets
src		src
.gitignore		.gitignore
Cargo.lock		Cargo.lock
Cargo.toml		Cargo.toml
LICENSE		LICENSE
README.md		README.md

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Quick Start

Comparison

Performance & Footprint

LLM Providers

Multi-Agent Orchestration

Chat Channels

Platform Features

Features

LLM Providers

Multi-Agent Orchestration

Built-in Tools (7)

Channel Integrations (5)

Platform

CLI Reference

Architecture

Configuration

Contributing

License

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

Quick Start

Comparison

Performance & Footprint

LLM Providers

Multi-Agent Orchestration

Chat Channels

Platform Features

Features

LLM Providers

Multi-Agent Orchestration

Built-in Tools (7)

Channel Integrations (5)

Platform

CLI Reference

Architecture

Configuration

Contributing

License

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages