RantAIClaw

Production multi-agent runtime in 100% Rust. Powers autonomous AI employees that communicate across channels, execute tools, manage memory, and run skills — all from a single binary.

RantAIClaw is designed for the boring real-world constraints production agents actually hit: cold start under 200ms, idle memory around 15MB, deterministic concurrency, secure-by-default tool execution, and per-profile isolation for multi-tenant deployments.

Why this exists

Most agent frameworks were written in Python or TypeScript and inherit the runtime they were built on — cold starts measured in seconds, GC pauses, fragile concurrency, dependency sprawl. Those tradeoffs are fine for notebooks. They are not fine for an agent that needs to stay connected to thousands of channels for months at a time, run inside a container with a 256MB limit, and execute tools that touch real systems.

RantAIClaw rewrites the runtime from the bottom up in Rust:

Metric	RantAIClaw	Typical Python alternative
Cold start	< 200ms	2–5s
Idle memory	~15 MB	200–500 MB
Binary size	~12 MB	(interpreter + deps)
Concurrent channels	Thousands	Hundreds

No garbage collector. No runtime overhead. Just async Rust on tokio.

What it does

Multi-channel. Connect one agent to Discord, Slack, Telegram, WhatsApp (Cloud or Web), Matrix (E2EE), Mattermost, Signal, Email, IRC, DingTalk, Lark, iMessage, Linq, QQ, Nextcloud Talk, and CLI — simultaneously, each with its own independent lifecycle.
Multi-provider. Route to OpenRouter, OpenAI, Anthropic, Gemini, Copilot — plus 25+ OpenAI-compatible and regional providers (Groq, Mistral, xAI, DeepSeek, Together, Fireworks, Perplexity, Cohere, Moonshot, GLM, Qwen, MiniMax, Baidu, Doubao, ZAI). Default model is anthropic/claude-sonnet-4.6.
Resilient provider wrapper. Retries with exponential backoff, multi-provider fallback, per-model fallback chains, API-key rotation on 429, Retry-After parsing.
Per-profile workspaces. Everything an agent owns lives under ~/.rantaiclaw/profiles/<name>/. Run multiple agents on one host without cross-contamination.
ClawHub + open-skills. Install community skills from ClawHub , or auto-sync the open-skills repo (besoeasy/open-skills) every 7 days for a shared baseline.
MCP supervision. Spawn and restart Model Context Protocol servers as supervised child processes (runtime tool integration is in-progress — see MCP Servers).
Agentic tools. 20+ built-in tools (shell, file_read, file_write, glob_search, cron_*, memory_*, web_search_tool, git_operations, pdf_read, screenshot, image_info, etc.) with security boundaries.
Three-level autonomy + four onboarding presets. Runtime states read_only / supervised / full, configured via L1–L4 preset bundles that write autonomy.toml + command_allowlist.toml + forbidden_paths.toml to disk.
Five memory backends. SQLite (default), Lucid (sqlite + workspace markdown overlay — RantAIClaw-only), Markdown, PostgreSQL, none. Hybrid lexical + vector recall, soul-snapshot for cold-boot survival.
Sandbox auto-detection. Landlock (Linux) → Firejail → Bubblewrap → Docker → Noop, picked at runtime based on what the host supports.
AEAD secret store. ChaCha20-Poly1305, master key at ~/.rantaiclaw/.secret_key, ciphertext stored inline in config.toml as enc2:<hex>.

Where to start

If you want to…	Read
Install and run your first agent	Getting Started
Understand how the runtime is built	Concepts → Architecture
Understand the per-profile workspace layout	Concepts → Profiles
See how a single message flows through the system	Concepts → Agent Loop
Decide how much autonomy to grant your agent	Concepts → Autonomy Levels
Look up a CLI command, config field, or channel	Reference

Project status

RantAIClaw is alpha (current 0.6.x series). Core subsystems — runtime, channels, providers, tools, gateway pairing, audit log, sandbox detection — are stable. Some areas are partially-built and clearly flagged in their respective docs:

Live config (PATCH /config/*) — handlers exist in src/gateway/config_api.rs but are not currently mounted on the running router.
MCP runtime tool integration — supervisor manages MCP processes; agent loop does not yet consume their tools.
Approval timeouts — timeout_secs is written to policy files but no consumer reads it; CLI prompts block indefinitely.

Where the docs make a claim about behavior, the file path / function citation is in the page so you can verify against the source.

Source and license

Source: github.com/RantAI-dev/RantAIClaw
License: GNU Affero General Public License v3.0 (AGPL-3.0)
Skills marketplace: clawhub.ai
Open skills repo: github.com/besoeasy/open-skills