724 Office AI Agent

7/24 Office AI Agent System

Skill by ara.so — Daily 2026 Skills collection.

A 24/7 production AI agent in ~3,500 lines of pure Python with no framework dependencies. Features 26 built-in tools, three-layer memory (session + compressed + vector), MCP/plugin support, runtime tool creation, self-repair diagnostics, and cron scheduling.

Installation

git clone https://github.com/wangziqi06/724-office.git
cd 724-office

# Only 3 runtime dependencies
pip install croniter lancedb websocket-client

# Optional: WeChat silk audio decoding
pip install pilk

# Set up directories
mkdir -p workspace/memory workspace/files

# Configure
cp config.example.json config.json

Configuration (config.json)

{
  "models": {
    "default": {
      "api_base": "https://api.openai.com/v1",
      "api_key": "${OPENAI_API_KEY}",
      "model": "gpt-4o",
      "max_tokens": 4096
    },
    "embedding": {
      "api_base": "https://api.openai.com/v1",
      "api_key": "${OPENAI_API_KEY}",
      "model": "text-embedding-3-small"
    }
  },
  "messaging": {
    "platform": "wxwork",
    "corp_id": "${WXWORK_CORP_ID}",
    "corp_secret": "${WXWORK_CORP_SECRET}",
    "agent_id": "${WXWORK_AGENT_ID}",
    "token": "${WXWORK_TOKEN}",
    "encoding_aes_key": "${WXWORK_AES_KEY}"
  },
  "memory": {
    "session_max_messages": 40,
    "compression_overlap": 5,
    "dedup_threshold": 0.92,
    "retrieval_top_k": 5,
    "lancedb_path": "workspace/memory"
  },
  "asr": {
    "api_base": "https://api.openai.com/v1",
    "api_key": "${OPENAI_API_KEY}",
    "model": "whisper-1"
  },
  "scheduler": {
    "jobs_file": "workspace/jobs.json",
    "timezone": "Asia/Shanghai"
  },
  "server": {
    "host": "0.0.0.0",
    "port": 8080
  },
  "workspace": "workspace",
  "mcp_servers": {}
}

Set environment variables rather than hardcoding secrets:

export OPENAI_API_KEY="sk-..."
export WXWORK_CORP_ID="..."
export WXWORK_CORP_SECRET="..."

Running the Agent

# Start the HTTP server (listens on :8080 by default)
python3 xiaowang.py

# Point your messaging platform webhook to:
# http://YOUR_SERVER_IP:8080/

File Structure

724-office/
├── xiaowang.py      # Entry point: HTTP server, debounce, ASR, media download
├── llm.py           # Tool-use loop, session management, memory injection
├── tools.py         # 26 built-in tools + @tool decorator + plugin loader
├── memory.py        # Three-layer memory pipeline
├── scheduler.py     # Cron + one-shot scheduling, jobs.json persistence
├── mcp_client.py    # JSON-RPC MCP client (stdio + HTTP)
├── router.py        # Multi-tenant Docker routing
├── config.py        # Config loading and env interpolation
└── workspace/
    ├── memory/      # LanceDB vector store
    ├── files/       # Agent file storage
    ├── SOUL.md      # Agent personality
    ├── AGENT.md     # Operational procedures
    └── USER.md      # User preferences/context

Adding a Built-in Tool

Tools are registered with the @tool decorator in tools.py:

from tools import tool

@tool(
    name="fetch_weather",
    description="Get current weather for a city.",
    parameters={
        "type": "object",
        "properties": {
            "city": {
                "type": "string",
                "description": "City name, e.g. 'Beijing'"
            },
            "units": {
                "type": "string",
                "enum": ["metric", "imperial"],
                "default": "metric"
            }
        },
        "required": ["city"]
    }
)
def fetch_weather(city: str, units: str = "metric") -> str:
    import urllib.request, json
    api_key = os.environ["OPENWEATHER_API_KEY"]
    url = f"https://api.openweathermap.org/data/2.5/weather?q={city}&units={units}&appid={api_key}"
    with urllib.request.urlopen(url) as r:
        data = json.loads(r.read())
    temp = data["main"]["temp"]
    desc = data["weather"][0]["description"]
    return f"{city}: {temp}°, {desc}"

The tool is automatically available to the LLM in the next tool-use loop iteration.

Runtime Tool Creation (Agent Creates Its Own Tools)

The agent can call create_tool during a conversation to write and load a new Python tool without restarting:

User: "Create a tool that converts Markdown to HTML."

Agent calls: create_tool({
  "name": "md_to_html",
  "description": "Convert a Markdown string to HTML.",
  "parameters": { ... },
  "code": "import markdown\ndef md_to_html(text): return markdown.markdown(text)"
})

The tool is saved to workspace/custom_tools/md_to_html.py and hot-loaded immediately.

Connecting an MCP Server

Edit config.json to add MCP servers (stdio or HTTP):

{
  "mcp_servers": {
    "filesystem": {
      "transport": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/data"]
    },
    "myapi": {
      "transport": "http",
      "url": "http://localhost:3000/mcp"
    }
  }
}

MCP tools are namespaced as servername__toolname (double underscore). Reload without restart:

User: "reload MCP servers"
# Agent calls: reload_mcp()

Scheduling Tasks

The agent uses schedule tool internally, but you can also call the scheduler API directly:

from scheduler import Scheduler
import json

sched = Scheduler(jobs_file="workspace/jobs.json", timezone="Asia/Shanghai")

# One-shot task (ISO 8601)
sched.add_job(
    job_id="morning_brief",
    trigger="2026-04-01T09:00:00",
    action={"type": "message", "content": "Good morning! Here's your daily brief."},
    user_id="user_001"
)

# Recurring cron task
sched.add_job(
    job_id="weekly_report",
    trigger="0 9 * * MON",          # Every Monday 09:00
    action={"type": "llm_task", "prompt": "Generate weekly summary"},
    user_id="user_001"
)

sched.start()

Jobs persist in workspace/jobs.json across restarts.

Three-Layer Memory System

from memory import MemoryManager

mem = MemoryManager(config["memory"])

# Layer 1 — session history (auto-managed, last 40 msgs)
mem.append_session(user_id="u1", session_id="s1", role="user", content="Hello!")

# Layer 2 — long-term compressed (triggered on session overflow)
# LLM extracts structured facts; deduped at cosine similarity 0.92
mem.compress_and_store(user_id="u1", messages=evicted_messages)

# Layer 3 — vector retrieval (injected into system prompt automatically)
results = mem.retrieve(user_id="u1", query="user's dietary preferences", top_k=5)
for r in results:
    print(r["content"], r["score"])

The LLM pipeline in llm.py injects retrieved memories automatically before each call:

# Simplified from llm.py
relevant = memory.retrieve(user_id, query=user_message, top_k=5)
memory_block = "\n".join(f"- {m['content']}" for m in relevant)
system_prompt = base_prompt + f"\n\n## Relevant Memory\n{memory_block}"

Personality Files

Create these in workspace/ to shape agent behavior:

`workspace/SOUL.md` — Personality and values:

# Agent Soul
You are Xiao Wang, a diligent 24/7 office assistant.
- Always respond in the user's language
- Be concise but thorough
- Proactively suggest next steps

`workspace/AGENT.md` — Operational procedures:

# Operational Guide
## On Error
1. Check logs in workspace/logs/
2. Run self_check() tool
3. Notify owner if critical

## Daily Routine
- 09:00 Morning brief
- 17:00 EOD summary

`workspace/USER.md` — User context:

# User Profile
- Name: Alice
- Timezone: UTC+8
- Prefers bullet-point summaries
- Primary language: English

Tool-Use Loop (Core LLM Flow)

# Simplified representation of llm.py's main loop
async def run(user_id, session_id, user_message, media=None):
    messages = memory.get_session(user_id, session_id)
    messages.append({"role": "user", "content": user_message})

    for iteration in range(20):          # max 20 tool iterations