Rust Agent Ecosystem
AI agents are mostly Python. But a growing cluster of projects builds them in Rust, trading ecosystem size for performance, safety, and small binaries. Here’s who does it and how.
Source: big_model_radar#97 (OpenClaw ecosystem daily report, 2026-03-26).
The landscape
| Project | Stars | Binary | RAM | Channels | Sandbox | Focus |
|---|---|---|---|---|---|---|
| ZeroClaw | 30.1k | small | <5MB | 25+ (incl. iMessage, IRC) | default | Edge/IoT, runs on $10 hardware |
| IronClaw | 11.8k | medium | – | REPL, Telegram, web | WASM | Security-first, credential protection |
| Moltis | 2.6k | 44MB | – | Telegram, Discord, WhatsApp, Teams | Docker + Apple | Personal server, 46 crates |
| ZeptoClaw | 589 | ~6MB | ~6MB | 11 channels | 6 runtimes | Ultra-light, 50ms startup |
| sgr-agent (ours) | – | – | – | CLI | – | Agent framework, SGR patterns |
For comparison, Python agents: OpenClaw (~430K LoC, Node.js), Hermes Agent (76k stars, Python), NanoBot (Python).
What they have in common
Single binary deployment. No pip, no node_modules, no Docker required (though most support it). ZeroClaw runs on ESP32. ZeptoClaw is 6MB.
Multi-channel as core feature. Not just CLI. Telegram, Discord, Slack, WhatsApp are table stakes. ZeroClaw does 25+ including iMessage and Matrix.
Security beyond “trust the LLM.” WASM sandboxing (IronClaw), Apple Container sandboxing (Moltis), SSRF prevention (ZeptoClaw), prompt injection detection (all of them). Rust’s type system helps here – compile-time tenant isolation in IronClaw.
MCP support. Every Rust agent supports Model Context Protocol. It’s the standard connector.
What differs
ZeroClaw bets on edge. <5MB RAM, hardware peripherals (ESP32, STM32, Arduino via traits). React 19 dashboard. 30k stars says the market wants lightweight agents.
IronClaw bets on security. WASM sandbox for tool execution, credentials never exposed to tools, capability-based permissions. PostgreSQL-backed memory with hybrid search. NEAR AI is behind it.
Moltis bets on modularity. 46 crates, feature-gated. You assemble what you need. We contributed reqwest bump (merged) and openai-oxide provider (rejected). See moltis-rust-agent.
ZeptoClaw bets on minimalism. 6MB binary, 50ms startup, 6MB RAM. 33 tools, 16 providers, multi-tenant. The “Alpine Linux” of agents.
Where our stack fits
Our approach is different. We don’t build an agent server. We build the framework layer that agent servers use:
openai-oxide → LLM client (any Rust agent can use)
sgr-agent-core → Tool/Backend traits (any agent can implement)
sgr-agent-tools → 14 reusable tools (any agent can embed)
sgr-agent → Agent loop with SGR patterns
project-openai-oxide is already a crate on crates.io. The rejected Moltis PR (#521) was an attempt to prove this – 888 lines replacing 5300. The architecture works, the social proof needs growing.
The gap in our stack: no multi-channel (Telegram, Discord). No sandboxing. No deployment story beyond “run locally.” These are the features that get Rust agents to 10k+ stars.
Patterns worth stealing
- WASM tool sandboxing (IronClaw) – tools run in WASM, can’t access host filesystem. Our FileBackend trait could add a WasmFs backend.
- Hardware peripheral traits (ZeroClaw) – same trait-based abstraction we use for FileBackend, but for GPIO/sensors. Interesting for airq/air-signal.
- Subscription OAuth (ZeroClaw) – route through existing ChatGPT/Claude subscriptions instead of API keys. Lowers user friction.
- Multi-tenant compile-time isolation (IronClaw) – TenantCtx generic parameter on every service. Type system prevents cross-tenant data leaks.
Links
- big_model_radar#97 – source comparison
- moltis-rust-agent – our Moltis contributions
- agent-toolkit-landscape – broader agent ecosystem (Python + Rust + JS)
- hermes-agent – top Python agent for comparison (76k stars)