Hermes Agent (Rust)

English | 中文 | 日本語 | 한국어

A production-grade Rust rewrite of Hermes Agent — the self-improving AI agent by Nous Research.

84,000+ lines of Rust · 16 crates · 641 tests · 17 platform adapters · 30 tool backends · 8 memory plugins · 6 cross-platform release targets

Highlights

Single Binary, Zero Dependencies

One ~16MB binary. No Python, no pip, no virtualenv, no Docker required. Runs on Raspberry Pi, $3/month VPS, air-gapped servers, Docker scratch images.

scp hermes user@server:~/
./hermes

Self-Evolution Policy Engine

The agent learns from its own execution. A three-layer adaptive system:

• L1 — Model & Retry Tuning. Multi-armed bandit selects the best model per task based on historical success rate, latency, and cost. Retry strategy adjusts dynamically based on task complexity.
• L2 — Long-Task Planning. Automatically decides parallelism, subtask splitting, and checkpoint intervals for complex prompts.
• L3 — Prompt & Memory Shaping. System prompts and memory context are optimized and trimmed per-request based on accumulated feedback.

Policy versioning with canary rollout, hard-gate rollback, and audit logging. The engine improves over time without manual tuning.

True Concurrency

Rust's tokio runtime gives real parallel execution — not Python's cooperative asyncio. JoinSet dispatches tool calls across OS threads. A 30-second browser scrape doesn't block a 50ms file read. The gateway processes messages from 17 platforms simultaneously without a GIL.

17 Platform Adapters

Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Mattermost, DingTalk, Feishu, WeCom, Weixin, Email, SMS, BlueBubbles, Home Assistant, Webhook, API Server.

30 Tool Backends

File operations, terminal, browser, code execution, web search, vision, image generation, TTS, transcription, memory, messaging, delegation, cron jobs, skills, session search, Home Assistant, RL training, URL safety, OSV vulnerability check, and more.
The built-in memory tool follows Python parity semantics: action=add|replace|remove, target=memory|user, with old_text substring matching for replace/remove updates.
Built-in store limits also match Python defaults: memory ≈ 2200 chars and user ≈ 1375 chars.

8 Memory Plugins

Mem0, Honcho, Holographic, Hindsight, ByteRover, OpenViking, RetainDB, Supermemory.
Built-in ~/.hermes/memories/MEMORY.md and USER.md snapshots are also injected at session start for prompt-stable long memory context.

6 Terminal Backends

Local, Docker, SSH, Daytona, Modal, Singularity.

MCP (Model Context Protocol) Support

Built-in MCP client and server. Connect to external tool providers or expose Hermes tools to other MCP-compatible agents.

ACP (Agent Communication Protocol)

Inter-agent communication with session management, event streaming, and permission controls.

Architecture

16-Crate Workspace

crates/
├── hermes-core           # Shared types, traits, error hierarchy
├── hermes-agent          # Agent loop, LLM providers, context, memory plugins
├── hermes-tools          # Tool registry, dispatch, 30 tool backends
├── hermes-gateway        # Message gateway, 17 platform adapters
├── hermes-cli            # CLI/TUI binary, slash commands
├── hermes-config         # Configuration loading, merging, YAML compat
├── hermes-intelligence   # Self-evolution engine, model routing, prompt building
├── hermes-skills         # Skill management, store, security guard
├── hermes-environments   # Terminal backends (Local/Docker/SSH/Daytona/Modal/Singularity)
├── hermes-cron           # Cron scheduling and persistence
├── hermes-mcp            # Model Context Protocol client/server
├── hermes-acp            # Agent Communication Protocol
├── hermes-rl             # Reinforcement learning runs
├── hermes-http           # HTTP/WebSocket API server
├── hermes-auth           # OAuth token exchange
└── hermes-telemetry      # OpenTelemetry integration

Trait-Based Abstraction

Error Hierarchy

AgentError (top-level)
├── LlmApi(String)
├── ToolExecution(String)      ← auto-converted from ToolError
├── Gateway(String)            ← auto-converted from GatewayError
├── Config(String)             ← auto-converted from ConfigError
├── RateLimited { retry_after_secs }
├── Interrupted { message }
├── ContextTooLong
├── MaxTurnsExceeded
└── Io(String)

Every error type converts automatically via From traits. The compiler ensures every error path is handled.

Install

Download the latest release binary for your platform:

# macOS (Apple Silicon)
curl -LO https://github.com/Lumio-Research/hermes-agent-rs/releases/latest/download/hermes-macos-aarch64.tar.gz
tar xzf hermes-macos-aarch64.tar.gz && sudo mv hermes /usr/local/bin/

# macOS (Intel)
curl -LO https://github.com/Lumio-Research/hermes-agent-rs/releases/latest/download/hermes-macos-x86_64.tar.gz
tar xzf hermes-macos-x86_64.tar.gz && sudo mv hermes /usr/local/bin/

# Linux (x86_64)
curl -LO https://github.com/Lumio-Research/hermes-agent-rs/releases/latest/download/hermes-linux-x86_64.tar.gz
tar xzf hermes-linux-x86_64.tar.gz && sudo mv hermes /usr/local/bin/

# Linux (ARM64)
curl -LO https://github.com/Lumio-Research/hermes-agent-rs/releases/latest/download/hermes-linux-aarch64.tar.gz
tar xzf hermes-linux-aarch64.tar.gz && sudo mv hermes /usr/local/bin/

# Linux (musl / Alpine / Docker)
curl -LO https://github.com/Lumio-Research/hermes-agent-rs/releases/latest/download/hermes-linux-x86_64-musl.tar.gz
tar xzf hermes-linux-x86_64-musl.tar.gz && sudo mv hermes /usr/local/bin/

# Windows (x86_64)
# Download hermes-windows-x86_64.zip from the releases page

All release binaries: https://github.com/Lumio-Research/hermes-agent-rs/releases

Building from source

cargo build --release
# Binary at target/release/hermes

Running

hermes              # Interactive chat
hermes --help       # All commands
hermes gateway start  # Start multi-platform gateway
hermes doctor       # Check dependencies and config

Testing

cargo test --workspace   # 641 tests

Contributing

git config core.hooksPath scripts/git-hooks  # Enable pre-commit fmt check

License

MIT — see LICENSE.

Based on Hermes Agent by Nous Research.