RustClaw

A lightweight, memory-aware Discord AI assistant powered by multi-provider LLM APIs, implemented in Rust.

Inspired by NanoClaw and OpenClaw.

Features

• Discord Integration — Mention-based interaction with streaming responses
• Multi-Provider LLM — Anthropic, OpenAI, Gemini via Rig
• Vector Memory — usearch (F16 HNSW) + SQLite semantic memory with recent turns + similarity search
• Pluggable Embeddings — Local (fastembed, 384d) or Gemini API (768d, near-zero RAM)
• Sandboxed Execution — All commands run in isolated Debian Docker containers (Bun pre-installed)
• Tool Calling — Shell commands, web search, weather, Typst rendering, file sending, cron scheduler
• Owner/User Permissions — AI-aware permission system for safe multi-user operation
• Auto-Update — Daily binary updates via cargo-dist (systemd/launchd)

Quick Start

Install

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/shimaenaga1123/rustclaw/releases/latest/download/rustclaw-installer.sh | sh

Setup Service (optional)

curl -fsSL https://raw.githubusercontent.com/shimaenaga1123/rustclaw/main/setup-service.sh | bash

Build from Source

git clone https://github.com/shimaenaga1123/rustclaw && cd rustclaw
cp config.example.toml config.toml  # edit with your credentials
cargo run --release

Prerequisites

• Discord Bot Token (with Message Content Intent enabled)
• Anthropic-compatible API Key
• Docker running locally
• Brave Search API Key (optional)
• Gemini API Key (optional, for embedding provider)

Configuration

[discord]
token = "your_discord_bot_token"
owner_id = 123456789012345678

[api]
provider = "anthropic"  # or "openai", "gemini"
key = "your_api_key"
url = "https://api.anthropic.com/v1"
model = "claude-sonnet-4-20250514"

[brave]
# api_key = "optional_brave_key"

[storage]
data_dir = "data"

[commands]
timeout = 30

[model]
disable_reasoning = false

[embedding]
provider = "local"  # or "gemini"
# api_key = "gemini_key"    # omit to reuse [api].key
# model = "gemini-embedding-001"
# dimensions = 768

Switching embedding providers: Delete data/conversations.usearch when changing between local (384d) and gemini (768d) — dimensions are incompatible.

Tools

Tool	Description	Permission
run_command	Shell commands in Docker sandbox	All
send_file	Send workspace files as Discord attachments	All
typst_render	Render Typst markup (tables, math) to PNG	All
web_search	Brave Search web lookup	All
weather	Current weather and forecast	All
search_memory	Semantic search over past conversations	All
schedule / list_schedules	Create and list cron tasks	All
important_add / important_list / important_delete	Manage persistent key facts	Owner only (add/delete)
unschedule	Remove a scheduled task	Owner only
reset_container	Reset Docker sandbox	Owner only

Memory System

┌─────────────────────────────────┐
│  # Important Facts              │  ← All entries, always included
├─────────────────────────────────┤
│  # Recent Conversations         │  ← Last 5 turns (chronological)
├─────────────────────────────────┤
│  # Related Past Conversations   │  ← Top 10 semantic matches
└─────────────────────────────────┘

• Storage: SQLite (data/memory.db) + usearch index (data/conversations.usearch)
• Embedding model cache: data/models/ (local provider only, ~130MB)

Data Layout

data/
├── memory.db              # SQLite (conversations + important facts)
├── conversations.usearch  # Vector index (F16 quantized)
├── models/                # Embedding model cache (local only)
├── workspace/             # Docker sandbox mount
└── schedules.json         # Persisted cron tasks

Security

• Credentials: Discord token, LLM API key, and optional Brave/Gemini API keys are stored in plain text in config.toml. Keep this file private and never commit it to version control.
• Command Execution: The run_command tool executes arbitrary shell commands inside a Docker container. While sandboxed, the /workspace directory is bind-mounted from the host. The owner permission system mitigates risk, but treat the bot owner account as privileged.
• Data Storage: All conversation history (user input + assistant responses) is stored locally in data/memory.db. No data is sent to third parties beyond the configured LLM/embedding API providers.
• Network Access: The Docker sandbox has network access enabled by default. Commands run inside can make outbound requests.
• Auto-Update: The optional auto-updater downloads binaries from GitHub Releases. Verify releases at the repository if concerned.

License

MIT — see LICENSE

Acknowledgments

Rig · serenity · usearch · fastembed · Typst · cargo-dist