momo — AI Agent Runtime

A from-scratch AI Agent Runtime in Go, built without any agent frameworks. Supports tool registration with JSON Schema, LLM autonomous decision-making, multi-turn multi-session conversations with real-time streaming, and disk-persisted session history.

Demo

📥 下载演示视频 (90MB) — 展示 momo 的实时流式响应、工具调用和多会话管理功能。

Architecture

cmd/server/main.go           Entry point, HTTP server
internal/
  agent/
    agent.go                  Core agent loop + tool execution
    tool.go                   Tool interface + Registry + JSON Schema
    session.go                Session management + persistence + OpenAI format
    context.go                Context compression (turn-based, non-destructive)
    prompts.go                Go template system prompt (embedded templates)
    store.go                  JSON file-backed session persistence
    templates/
      momo.md.tpl             Main system prompt (rules, tools, workflow)
      title.md                Session title generation prompt
      summary.md              Context summary prompt
  tools/
    calculator.go             Arithmetic expression evaluator
    web.go                    Web fetch / search (URL fetching + HTML extraction)
    read_docs.go              Local file reader
    time.go                   Current date/time
    weather.go                Real-time weather (wttr.in)
  web/
    handler.go                Web UI with SSE streaming

Core Loop

User Input → LLM → Tool Call? → Execute → LLM → Tool Call? ...
                ↓                          ↓
          Final Answer              Streaming via SSE

Key Features

• Tool Registration: Each tool implements Tool interface with name, description, and JSON Schema parameters. LLM autonomously decides which tool to call.
• Template System Prompt: momo.md.tpl rendered with env data (date, platform, working dir) — no hardcoded prompt strings.
• SSE Streaming: Tool calls and results streamed to frontend via NDJSON over POST, no polling.
• Real-time Thinking Display: 🔧 tool calls, 📎 results, 🤔 LLM thinking shown live, latest step auto-expands and flows like water.
• Concurrent Sessions: Independent sessions with per-session mutex, concurrent-safe. Switch between sessions freely — each maintains its own context.
• Disk Persistence: Sessions saved as JSON files, survive server restart. Original messages never deleted — compression only affects LLM context.
• Turn-based Context Compression: 100-turn limit, oldest turns summarized into compact text when exceeded. Original data preserved on disk.
• Concurrency-safe: sync.RWMutex on Session, SetStatus() for atomic status updates, per-request event callbacks via context (no shared mutable state).

Tools

Tool	Description
calculator	Arithmetic expression evaluator (+, -, *, /, **, sqrt, sin, cos, etc.)
web	Fetch any URL and return readable text content. Also used for web search.
read_docs	Read any local file.
get_time	Get current date and time.
weather	Real-time weather via wttr.in (free, no API key).

Quick Start

Prerequisites

• Go 1.22+
• OpenAI API key (or any OpenAI-compatible API)

Run

# Set your API key
export OPENAI_API_KEY=sk-...

# Run the web server
go run ./cmd/server/

# With custom model and API endpoint
go run ./cmd/server/ --model gpt-4o-mini --api-url https://api.openai.com/v1

使用 DeepSeek

# 设置 DeepSeek API key（从 https://platform.deepseek.com/ 获取）
export OPENAI_API_KEY=sk-xxxxxxxxxxxxxxxx

# 使用 DeepSeek-V3 模型运行
go run ./cmd/server/ --model deepseek-chat --api-url https://api.deepseek.com/v1

# 使用 DeepSeek-R1 推理模型运行
go run ./cmd/server/ --model deepseek-reasoner --api-url https://api.deepseek.com/v1

Open http://localhost:8080 in your browser.

Usage Examples

1. Calculator: "What is 2 + 2?" or "Calculate sqrt(144) * 3"
2. Web: "What's the latest AI news?" or "Fetch https://example.com"
3. Read Docs: "Read the file README.md"
4. Weather: "What's the weather in Tokyo?"
5. Multi-session: Open sidebar, create new sessions, ask different questions in parallel

Configuration

go run ./cmd/server/ \
  --api-key sk-... \           # OpenAI API key (default: OPENAI_API_KEY env)
  --model gpt-4o-mini \         # LLM model (DeepSeek: deepseek-chat / deepseek-reasoner)
  --api-url https://api.openai.com/v1 \  # OpenAI-compatible API base (DeepSeek: https://api.deepseek.com/v1)
  --port 8080 \                 # HTTP server port
  --max-turns 999 \             # Max turns per request
  --data-dir ./data             # Session persistence directory

Testing

go test ./...

Dependencies

• github.com/sashabaranov/go-openai — OpenAI API client
• golang.org/x/net — HTML parsing for web content