A 24/7 local AI work assistant built with Go. Gateway, sessions, memory, and tool execution stay on your machine.

Language: 中文 | English

maxclaw is a local AI agent for developers and operators. Core value proposition: low memory footprint, fully local workflow, visual desktop/web UI, and fast onboarding.

• Go backend, resource-efficient runtime: single binary gateway + tool orchestration.
• Fully local workflow: sessions, memory, logs, and tool runs are stored locally.
• Desktop UI + Web UI: visual settings, streaming chat, file preview, and terminal integration.
• Out-of-the-box setup: one-command install and default workspace templates.

SEO keywords: Go AI Agent, local AI assistant, self-hosted AI agent, private AI workflow, desktop AI app, low-memory AI.

• Go-native agent loop and tool system
• Fully local execution path with auditable artifacts
• Desktop UI + Web UI + API on the same port
• executionMode=auto for unattended long-running tasks
• spawn sub-sessions with independent context/model/source and status callbacks
• Automatic task titles that summarize sessions without overwriting message content
• Monorepo-aware recursive context discovery (AGENTS.md / CLAUDE.md)
• Multi-channel integrations: Telegram, WhatsApp (Bridge), Discord, WebSocket
• Cron/Once/Every scheduler + daily memory digest

If you are familiar with OpenClaw, maxclaw follows similar local-first principles with a Go-first engineering focus:

• Local-first agent execution and private data boundaries
• Heartbeat context (memory/heartbeat.md)
• Memory layering (MEMORY.md + HISTORY.md)
• Autonomous mode (executionMode=auto)
• Sub-agent task split via spawn
• Monorepo context discovery for multi-module repositories

1. Install Go 1.24+ and Node.js 18+
2. Build: make build
3. Initialize workspace: ./build/maxclaw onboard
4. Configure: edit ~/.maxclaw/config.json
5. Run gateway: ./build/maxclaw-gateway -p 18890

Built binaries:

• ./build/maxclaw: full CLI (onboard, skills, telegram bind, gateway, ...)
• ./build/maxclaw-gateway: standalone backend for desktop packaging or headless use

All-in-one local dev startup:

make build && make restart-daemon && make electron-start

Common dev restart commands:

make dev-gateway
make backend-restart
make dev-electron
make electron-restart

curl -fsSL https://raw.githubusercontent.com/Lichas/maxclaw/main/install.sh | bash

Path: ~/.maxclaw/config.json

{
  "providers": {
    "anthropic": { "apiKey": "your-anthropic-key" }
  },
  "agents": {
    "defaults": {
      "model": "anthropic/claude-opus-4-5",
      "workspace": "/absolute/path/to/your/workspace",
      "executionMode": "auto"
    }
  }
}

OpenAI native models use the official openai-go SDK with default API base https://api.openai.com/v1. Anthropic native models use the official anthropic-sdk-go SDK with default API base https://api.anthropic.com.

Set agents.defaults.executionMode:

• safe: conservative exploration mode
• ask: default mode
• auto: autonomous continuation (no manual "continue" approval for paused plans)

MaxClaw implements a six-layer adaptive system that continuously learns from errors and user feedback:

Example - Automatic Error Recovery:

Error: 400 Bad Request (context too large)
↓
[Verification] Classifies as ContextOverflowError
↓
[Reflection] Compresses conversation to structured summary
↓
[Adaptation] Reduces context length by 30%, falls back to model with larger window
↓
[Persistence] Saves checkpoint for potential recovery
↓
[Evolution] Logs pattern: "Repeated context overflow in Go refactoring tasks"
↓
[Feedback] If user corrects, learns: "Prefer explicit type annotations in Go"

The system is cost-optimized: 70% of feedback detection uses regex rules (zero cost), only 15% require LLM semantic analysis (~$0.1/day). See docs/agent_lifecycle.md for details.

1. Build: make webui-install && make webui-build
2. Start: ./build/maxclaw-gateway -p 18890
3. Open: http://localhost:18890

• Architecture: ARCHITECTURE.md
• Operations: MAINTENANCE.md
• Browser runbook: BROWSER_OPS.md
• Full Chinese docs and all channel/config examples: README.zh.md