Hermes-Agent-Windows-Rust 🇨🇳 中文

A high-performance AI agent desktop application for Windows, built with Rust + React. Streaming tool calls, multi-provider LLM support, 19 built-in tools, and a 9-page Web UI.

Based on NousResearch/hermes-agent (MIT License). This is a ground-up Rust rewrite targeting Windows desktop with native WebView2 GUI.

UI & feature design inspired by hermes-hud and hermes-workspace.

---

🇨🇳 中文说明

---

Features

Core Agent

• Streaming Tool Calls — Real-time streaming responses with live tool execution. Resolved a critical parsing bug in the streaming tool_call accumulator (commit d92ca75).
• Interruptible — Stop any in-progress streaming request instantly.
• Multi-turn Sessions — Persistent conversation history with SQLite + FTS5 full-text search.
• Tool Auto-Skill Generation — After complex tasks, the agent autonomously creates reusable skills.

Multi-Provider LLM Support

• MiniMax (default)
• OpenAI (GPT-4 series)
• Anthropic (Claude series)
• models.dev registry with context-length awareness
• Per-conversation model override (/model command)

Built-in Tools (19 total)

Tool	Description
terminal	Execute PowerShell/cmd commands
file_read / file_write / file_edit	Full file operations
list_dir	Directory listing with tree view
skill_create / skill_list / skill_view / skill_delete	Skill lifecycle management
memory_save / memory_search	Persistent memory store
browser_navigate / web_search	Web automation
execute_code	Sandboxed code execution
interrupt	Streaming interrupt
delegate	Subagent delegation

Frontend UI (9 pages)

• Chat — Main conversation interface with tool call cards, streaming markdown, session switcher
• Files — Monaco editor with file tree, save/load workflow
• Terminal — xterm.js terminal emulator embedded in browser
• Memory — Persistent memory editor with section-based storage
• Skills — Browse, create, and manage auto-generated skills
• Dashboard — Real-time HUD with stats (sessions, skills, messages)
• Inspector — Per-session message history and token usage
• Settings — Theme switcher (default/ares/slate/mono), provider API keys
• HUD — Agent telemetry overlay (total sessions, skills, messages)

Desktop Ready

• Tauri 2.x — Native Windows WebView2 packaging
• React + Vite — Fast HMR development, TypeScript throughout
• Playwright — E2E test suite included in devDependencies

---

Quick Start

Option A — Download Pre-built (Recommended)

1. Download the latest release: v0.1.0 → release-package.zip
2. Extract to any folder
3. Double-click start.bat
4. Open http://localhost:1420 → Settings → enter your API key → Save
5. Start chatting!

Option B — Build from Source

Prerequisites:

• Rust 1.93+
• Node.js 18+
• Windows 10/11

Build:

# Clone the repo
git clone https://github.com/tcflying/hermes-agent-windows-rust.git
cd hermes-agent-windows-rust

# Build Rust backend
cargo build --release

# Install frontend dependencies
cd crates/ui && npm install && cd ../..

Run:

start.bat    # Starts backend (port 3848) + frontend (port 1420)
stop.bat     # Stops all services

• Frontend: http://localhost:1420
• Backend API: http://localhost:3848

---

Architecture

hermes-rs/
├── crates/
│   ├── agent/          # Core agent loop, tool execution, memory, prompt building
│   │   └── src/
│   │       ├── chat.rs          # Main streaming chat handler (FIXED: tool_call parser)
│   │       ├── tools.json        # 19 tool definitions
│   │       ├── memory.rs         # MemoryStore, MemoryManager, MemorySnapshot
│   │       ├── memory_nudge.rs   # Periodic nudge injector
│   │       ├── skill_commands.rs # /skill slash command handler
│   │       ├── iteration.rs      # IterationBudget (max 30 tool calls/turn)
│   │       ├── interrupt.rs      # InterruptFlag for streaming stop
│   │       └── auxiliary_client.rs # Vision + summarization clients
│   ├── cli/            # HermesCLI — slash commands, skin engine, setup wizard
│   ├── config/         # YAML config loader, multi-provider schema
│   ├── gateway/        # Axum HTTP server, SSE streaming, session routing
│   │   └── platforms/ # Telegram, Discord, Slack, WhatsApp, Signal adapters
│   ├── session/        # SQLite + FTS5 session DB, trajectory saving
│   ├── tool-registry/  # Central tool registry (stub, pending integration)
│   ├── ui/             # React + TypeScript frontend
│   │   └── src/
│   │       ├── pages/  # 9 page components (Chat, Files, Terminal, Memory...)
│   │       ├── components/ # WorkspaceLayout, FileTree, ToolCallCard...
│   │       └── App.tsx # React Router v7 routing
│   └── utils/          # Shared utilities
├── start.bat           # One-click dev startup
├── stop.bat            # Stop all services
└── Cargo.toml          # Workspace root

Tech Stack

Layer	Technology
GUI Shell	Tauri 2.x (WebView2)
Frontend	React 18, TypeScript, Vite 6
Backend	Rust, Tokio, Axum 0.8
Database	SQLite + FTS5 (rusqlite, bundled)
Terminal	xterm.js
Code Editor	Monaco Editor
Markdown	react-markdown + shiki
HTTP	reqwest (streaming), tower-http
CLI	clap (derive)
Logging	tracing + tracing-subscriber

---

Key Fix: Streaming tool_calls Parser

A critical bug in the streaming tool_call accumulation logic caused {"error":"invalid function arguments json string"} on every tool call during streaming responses.

Root Cause: The original state machine used index field to detect tool call boundaries — but the index was unreliable across providers and multiple concurrent tool calls caused id flush timing errors.

Fix (commit d92ca75): Replaced index-based detection with id-change detection:

• Each incoming tool_call chunk is checked for a new id
• When id changes, the previous tool call is flushed to the accumulated list
• No mutexes needed for accumulation state — simple owned variables

File: crates/agent/src/chat.rs lines 183–283

---

Configuration

Config file: ~/.hermes/config.yaml

llm:
  provider: minimax
  model: MiniMax-M2.7-highspeed
  api_key: your-api-key

display:
  skin: default        # default / ares / slate / mono
  tool_progress: true
  streaming_thinking: true

gateway:
  host: 0.0.0.0
  port: 3848

memory:
  max_size: 50000
  nudge_interval: 50

---

Development

# Run tests
cargo test

# TypeScript check (frontend)
cd crates/ui && npx tsc --noEmit

# Format
cargo fmt
cargo clippy --fix

---

Project Status

Component	Status
Core chat + streaming	✅ Working
Tool calling (19 tools)	✅ Working
Skill auto-creation	✅ Working
Memory system	✅ Basic (MemoryStore, nudge)
Session DB + FTS5	✅ Working
Multi-provider LLM	✅ Working
React UI (9 pages)	✅ Working
HTTP Gateway	✅ Working
CLI + slash commands	✅ Working
Streaming interrupt	✅ Working
Platform adapters	✅ Telegram, Discord, Slack, WhatsApp, Signal
Skill self-evolution	✅ Auto-create, merge duplicates, prune stale
ACP adapter (VS Code)	🔜 Not started

---

License

MIT — Same as NousResearch/hermes-agent.