One Rust binary.SpeaksDiscord.
A multi-agent runtime with a built-in TUI that listens, thinks, and replies across 10+ channels from a single ~12 MB binary. Sub-200ms cold start. ~15 MB idle. Per-profile isolation.
curl -fsSL https://raw.githubusercontent.com/RantAI-dev/RantAIClaw/main/scripts/bootstrap.sh | bashA TUI you'll actually live in.
Built on ratatui. Full-screen, keyboard-driven, opinionated. Chat with your agent, watch tools execute mid-run, inspect memory, tail the audit log — from one process. Run rantaiclaw chat and you're in.
Tools mid-run
See every tool call, its inputs, its sandbox decision, and its result inline as the agent thinks.
Memory browser
Walk Core / Daily / Conversation entries. Search by query. Forget what shouldn't survive.
Live audit
Tail the JSONL audit log in real time. Approvals, denials, sandbox backend, exit codes.
The boring real-world constraints production agents hit — solved at the runtime layer.
100% Rust
Async tokio runtime, no GC, ~12 MB binary, sub-200ms cold start, ~15 MB idle memory. Survives constrained containers.
18 Channels
Discord, Slack, Telegram, WhatsApp (Cloud + Web), Matrix E2EE, Mattermost, Signal, Email, IRC, DingTalk, Lark, iMessage, and more — concurrently, each with its own lifecycle.
30+ LLM Providers
OpenRouter, Anthropic, OpenAI, Gemini, Copilot, plus 25+ OpenAI-compatible and regional providers. Per-model routing, multi-provider fallback, API-key rotation on 429.
Secure by Default
Deny-by-default allowlists, auto-detected sandbox (Landlock → Firejail → Bubblewrap → Docker), AEAD-encrypted secret store, append-only JSONL audit log.
5 Memory Backends
SQLite (default), Lucid (sqlite + markdown overlay), Markdown, PostgreSQL, none. Hybrid BM25 + cosine recall. Soul-snapshot for cold-boot survival.
Skills + MCP
Install skills from the library, auto-sync every 7 days, supervise MCP servers as child processes with exponential-backoff restart.
Tuned for the workloads agents actually run on.
From channel ingress to reply — six steps, one binary.
Channel ingress
Each channel runs in its own tokio task with its own reconnect strategy. Messages normalize into a common shape before the agent loop sees them.
Context assembly
Tools → Hardware → Skills → Workspace → Project context → Time → Channel capabilities. Memory recall (hybrid BM25 + cosine) is prepended to the user message.
Provider call
ReliableProvider wraps every call — exponential backoff, multi-provider fallback, per-model fallback chains, API-key rotation on 429.
Approval gate
Every tool call passes through the approval gate. Autonomy levels (read_only / supervised / full) and per-tool allowlists decide whether to run, prompt, or deny.
Tool execution
Sandboxed (Landlock / Firejail / Bubblewrap / Docker) with strict input validation. Sequential by default; parallel when configured.
Reply egress
Result fed back to the provider until the LLM returns a text-only response. Channel formats per-platform (markdown, mentions, drafts) and delivers.
One command. No Rust toolchain. No compiler.
Auto-detects your platform (Linux x86_64/aarch64/armv7, macOS Intel/Apple Silicon, Windows x86_64), downloads the matching prebuilt binary from the latest release, verifies its SHA-256 checksum, and installs it to ~/.cargo/bin, ~/.local/bin, or your Windows PATH.
# Installs the binary AND runs the full guided setup wizard
# (provider, approvals, channels, persona, skills, MCP).
# Add --skip-setup to install only.
curl -fsSL https://raw.githubusercontent.com/RantAI-dev/RantAIClaw/main/scripts/bootstrap.sh | bash
rantaiclaw --version
rantaiclaw chat # start chatting!Read the docs.
Ship an agent.
From the trait + factory architecture to the agent loop to the autonomy model — the docs explain how RantAIClaw is built, so you can extend it.