LightClaw: The Real Lightweight AI Assistant

LightClaw is a fast, local-first AI assistant inspired by OpenClaw and nanobot, packaged as a single Rust binary.

If you want agentic tooling, memory, and Telegram/Discord integration without a heavy runtime, lightclaw is built for that.

curl -fsSL https://lightclaw.dev/install.sh | bash
lightclaw configure

Supported platforms:

• Linux x86_64
• Linux ARM64 (Raspberry Pi 4/5, ARM servers)
• Linux ARMv7 (e.g. older ARM devices / e-readers)
• macOS x86_64
• macOS ARM64 (Apple Silicon)

Note: A Windows binary exists, but it is currently less stable and not as well-supported as the others.

• Single-binary deploy: ship one executable, no Python runtime.
• Tool-capable agent: file, shell, web, and scheduling actions.
• Telegram/Discord-native interface: high-performance polling built in.
• Local-first memory: vectors + metadata stored locally with SQLite.
• Rust reliability: strong typing, memory safety, and concurrency.
• Skills support: OpenClaw-style SKILL.md skills via activate_skill.

LightClaw includes long-term memory:

• Short-term chat history per session.
• Periodic summarization of recent conversation chunks.
• Semantic retrieval over stored memories.
• Privacy-first local storage (no external vector DB required).

Create ~/.lightclaw/config.json:

{
  "agents": {
    "defaults": {
      "provider": "openrouter",
      "model": "anthropic/claude-opus-4-5",
      "model_fallbacks": [
        "openai/gpt-4o-mini",
        "ollama/llama3.2"
      ]
    }
  },
  "providers": {
    "openrouter": {
      "apiKey": "sk-or-..."
    },
    "openai": {
      "apiKey": "sk-..."
    },
    "ollama": {
      "apiBase": "http://127.0.0.1:11434/v1"
    },
    "mistral": {
      "apiKey": "..."
    }
  },
  "channels": {
    "telegram": {
      "token": "YOUR_BOT_TOKEN",
      "allow_from": ["123456789"],
      "transcription": {
        "enabled": true,
        "provider": "openai",
        "model": "whisper-1",
        "language": "en",
        "max_bytes": 20971520,
        "diarize": false,
        "context_bias": "",
        "timestamp_granularities": ["segment"]
      }
    },
    "discord": {
      "token": "YOUR_DISCORD_BOT_TOKEN",
      "allow_from": ["123456789012345678"],
      "allowed_channels": ["123456789012345678"]
    }
  },
  "tools": {
    "web": {
      "search": {
        "provider": "firecrawl",
        "firecrawlApiKey": "fc-..."
      }
    }
  }
}

cargo build --release
./target/release/lightclaw

Cross-platform build script:

./scripts/build.sh

LightClaw uses an actor-like model with a central MessageBus:

• Agent: context handling and LLM orchestration.
• Telegram: chat input/output transport.
• Discord: chat input/output transport.
• Tools: executable capability modules.
• Memory: summary ingestion + retrieval loop.

All components run on a single async Tokio runtime.

lightclaw can discover and activate OpenClaw-style skills from:

• ~/.lightclaw/workspace/skills/*/SKILL.md
• ~/.lightclaw/workspace/.agents/skills/*/SKILL.md
• ~/.agents/skills/*/SKILL.md

When relevant, the model can call activate_skill to load the full instructions for a skill.

LightClaw includes a native skills command group backed by Rust APIs for ClawHub and skills.sh/source installs.

# Search across all backends (default: ClawHub + skills.sh)
lightclaw skills search "calendar"

# Search on skills.sh only
lightclaw skills search react --from skills

# Search on ClawHub only
lightclaw skills search react --from clawhub

# Install from ClawHub
lightclaw skills install weather --from clawhub

# Install from source (OpenClaw-compatible project layout)
lightclaw skills install vercel-labs/agent-skills --from skills

For --from skills, installs land in ./skills under the workspace.

LightClaw also includes a native service command group to manage a background daemon and stream logs with the same commands across platforms.

# Install and start a user-level service
lightclaw service install

# Check service state
lightclaw service status

# Restart/stop/start
lightclaw service restart
lightclaw service stop
lightclaw service start

# Stream runtime logs
lightclaw service logs -f

# Remove the service
lightclaw service uninstall

# Full uninstall wizard (service + optional files + optional binary)
lightclaw uninstall

By default, service commands target user-level services. Use --system for system-level service operations.

scripts/
  build.sh
  release.sh
  install.sh
  count_loc.sh
src/
  lib.rs          # Library crate root (app wiring / CLI runner)
  agent/          # Agent orchestration and core reasoning flow
  channels/       # Channel adapters (Telegram, Discord)
  cron/           # Scheduling types and persistent schedule storage
  memory/         # Summary, vector/file stores, retrieval logic
  skills/         # Skill manager, installer hub, and skills CLI commands
  tools/          # Tool implementations (fs, shell, web, send, cron)
  bus.rs          # Message bus for component coordination
  config.rs       # Config schema and loading
  configure.rs    # CLI setup flow for local configuration
  main.rs         # Thin binary entrypoint
  transcription.rs # Audio transcription integration

LightClaw is built on Rig, which provides:

• Provider abstraction across OpenAI/OpenRouter-style backends
• Structured tool calling
• Retrieval-friendly agent primitives

Contributions are welcome.

• Open an issue for bugs, regressions, or feature ideas.
• Open a PR with focused changes and a clear description.
• Keep changes lightweight and production-oriented.

_{run lighter, run faster, run everywhere}