724 Office: the tool table lists 37 and the header claims 36
7/24 Office — Self-evolving AI Agent system. 36 tools, 10,000 lines pure Python, modular architecture, MCP plugins, three-layer memory, nudge system, AI mirror, 24/7 production.
At a glance
- What is it?
- wangziqi06/724-office is a self-evolving AI agent presented as roughly 10,000 lines of pure Python with no framework dependency, running 24/7 across multiple users with a three-layer memory, a Docker multi-tenant router and tools the agent can write at runtime. Counting what the page actually lists produces 37 tools rather than 36, and the architecture diagram stops at an open box.
- Who is it for?
- This is a personal production agent that was open sourced after the fact, and the page reads better as an account of one system than as a product you can adopt. Before running anything from it, know what you are turning on: a single admin module carries shell execution, file writes, an audit tool and runtime creation of new Python tools, with a circuit breaker as the only stated limit and no approval step anywhere in this generation's feature list.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 107 days ago.
- What is it written in?
- Mainly JavaScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- wangziqi06/724-office1,020 ★ · -3
| Repository | 2026-09-12 | 2026-11-15 |
|---|---|---|
| wangziqi06/724-office | 1k | 1k |
-3 Stars in the last 64 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The tool table lists 37 and the header claims 36
The header says 36 tools, 20 files, and the repository description repeats 36. Counting the tool names in the table gives 37, spread over twelve categories: two core, four files, three scheduling, six media send, three video, four search, one visualization, two AI Mirror, six diagnostics, two memory, three plugins and one MCP. The same header says 20 files, while eighteen Python files sit at the root. The module table accounts for thirteen of them and comes to a little over 8,100 lines, against the roughly 10,000 that the description claims, so the shortfall sits in the rows that the table does not reach.
None of this is fatal, but it sets the tone for the whole page. The counts are stated twice in the summary and once in the per-repository description, while the detailed table that a reader would check them against disagrees. The number that matters in practice is not the total; it is that one module holds eighteen of them.
The repository language is JavaScript and the code is Python
The repository is tagged as a JavaScript project. What the page describes is a Python one: around 10,000 lines of pure Python, zero framework dependency, no LangChain, no LlamaIndex, no CrewAI, just the standard library plus a few small packages. The default branch is master rather than main, there are no GitHub releases, and the license is MIT with the file at the root.
The page offers an explanation for the language tag close by. A section near the top points at a successor in the subdirectory xiaowang-v2/, described as a single Node process with a single WAL SQLite and a hand-written agentic loop. A Node rewrite of this size sitting inside the repository is enough to move the language statistic off Python, which means the tag describes the successor rather than the ten thousand lines a reader came for. The page also notes the system was built solo with AI co-development tools and run in production across multiple users.
The architecture diagram ends with a box that has no bottom
The diagram the page offers for its own structure stops in the middle:
+-----------------+
| Messaging |
| Platform |
+--------+--------+
|
+--------v--------+
| router.py | Multi-tenant routing
| Auto-provision | Container reconciliation
| Group routing | Health checking
+--------+--------+
|
+--------v--------+
| xiaowang.py | Entry pointThe last box is opened and never closed, and nothing below it is drawn. Everything the rest of the page describes is missing from the picture: the seven tool modules, the three-layer memory, the scheduler, the messaging wrapper and the per-user containers the router provisions. The module table fares no better at its end, where the scheduler row is cut in the middle of its responsibility column, after the words Cron and a plus sign. Two of the page's structural sections are therefore partial, which is worth knowing before treating either as a map of the system.
One module holds exec, write_file and create_tool
The admin module is the largest in the table at about 860 lines and it is where eighteen of the thirty seven tools live: the exec and message pair, the four file operations, the six diagnostics including code_audit and daily_digest, the two memory compaction calls, the three plugin calls and the one MCP reload. The registry module that carries the decorator and the circuit breaker holds none of the tools itself.
So one file holds shell execution, arbitrary file writing and editing, an audit pass, and create_tool, which lets the agent write, save and load new Python tools while it is running, with list_custom_tools and remove_tool alongside it. The stated limits around that surface are a circuit breaker that disables a tool after three consecutive failures in a session, and a tool loop that allows up to twenty iterations per conversation with automatic retry. Approval appears nowhere in this generation's feature list or tool table. It appears in the successor, as one of five deterministic jobs assigned to the harness.
The nudge system fires when the model has tools and does not call them
The most distinctive mechanism is a correction loop aimed at the model's own behaviour. The nudge system is described as structural behaviour correction that auto-detects when the LLM has tools available but does not use them, and there is a nudge module at the root with nudge integration sitting inside the LLM module's roughly 1,260 lines. Around it sit four more controls aimed at cost and stability: dynamic tool filtering through five context profiles for voice, scheduler, group, diagnostic and default use, a budget-aware system prompt that tracks token usage as the prompt is assembled, the three-failure circuit breaker, and an inactivity guard that skips cron tasks for users dormant past a three day threshold. Messaging API calls retry three times with 2, 4 and 8 second delays, sessions are recorded daily as a black box, and the router rebuilds missing containers from its routing table on startup.
Memory keeps 40 messages, archives at 100KB, and dedupes at 0.92
The memory module is the longest in the table at about 1,100 lines and it works in three layers:
Layer 1: Session (short-term)
- Last 40 messages per session, JSON files
- Overflow triggers compression
- Auto-archive sessions >100KB
Layer 2: Compressed (long-term)
- LLM extracts structured facts from evicted messages
- Deduplication via cosine similarity (threshold: 0.92)
- Stored as vectors in LanceDB
Layer 3: Retrieval (active recall)
- User message -> embedding -> vector search
- Top-K relevant memories injected into system prompt
- Budget-aware injection (tracks token usage)Two numbers define the edges. Forty messages is the window before compression starts, and 100KB is where a session file is archived rather than kept live, so the long-term layer is fed by whatever the short-term layer evicts rather than by the full history. The 0.92 cosine threshold is what stands between a user and a memory store that keeps near-duplicates of the same fact, and it is a fixed constant rather than something a user can tune from a config file on the page.
One container per user, plus a second one for the group chat
The router provisions Docker containers, one per user, health checks them and rebuilds any that are missing from its routing table when it starts. Group conversations get their own independent container with at-mention gating, a context buffer of the last twenty messages and speaker identification, so a group thread does not share a container with a person's private session. Messaging itself is a pluggable platform layer with debounce, message splitting and streaming media upload, wrapped by a module of about 447 lines that also handles CDN upload and download.
External capability comes in two shapes. MCP servers are reached over JSON-RPC on stdio or HTTP, hot-reloaded without a restart, with a reload_mcp tool as the manual trigger. Runtime tools are the other shape, and the more unusual one: the agent writes, saves and loads new Python tools for itself. Nothing on the page shows an MCP server configuration shipping with the repository, so both routes need endpoints you supply.
Only example config files are committed, and no releases exist
Nothing runnable sits at the root. The configuration file is config.example.json, the Dockerfile is Dockerfile.example and the compose file is docker-compose.example.yml, so a working setup starts by copying and renaming three files whose contents the page never reproduces. The remaining root entries are the eighteen Python modules and directories already covered, plus archive.py, audit.py, mcp_client.py, nudge.py and router.py, with the printed module table ending at scheduler.py before those last five appear.
The rest is small and dated. No GitHub releases exist, so there is no tag to pin and no changelog to read; the change history is the news list on the page. The repository is MIT licensed, the default branch is master, the last commit landed on 2026-07-31, and the counts are 1,016 stars, 170 forks and 4 open issues. The successor sits beside all of it in xiaowang-v2/, with a single Node process, a WAL SQLite and 330 or more offline selftest assertions, documented in Chinese.
Editorial conclusion
This is a personal production agent that was open sourced after the fact, and the page reads better as an account of one system than as a product you can adopt. Before running anything from it, know what you are turning on: a single admin module carries shell execution, file writes, an audit tool and runtime creation of new Python tools, with a circuit breaker as the only stated limit and no approval step anywhere in this generation's feature list. That is a different safety posture from the successor, which names approval gates among the five jobs it assigns to the harness instead of the model. On the packaging side, nothing runnable is committed: the config, Dockerfile and compose files all ship as .example files, there are no releases, and the default branch is master. If you want a maintained version of these ideas rather than these files, the successor is in the same repository and documented in Chinese.
Frequently asked questions
What is wangziqi06/724-office?
A self-evolving AI agent system described as roughly 10,000 lines of pure Python with zero framework dependency: no LangChain, no LlamaIndex and no CrewAI, only the standard library plus a few small packages. It is presented as a production agent running 24/7 across multiple users, built solo with AI co-development tools.
How many tools does 724-office have?
The page claims 36 in both its header and its tool table title. The table itself lists 37 tool names across twelve categories, of which 18 sit in the admin module alongside exec, the file operations and runtime tool creation.
Does 724-office need LangChain or another agent framework?
No. The page states zero framework dependency and names LangChain, LlamaIndex and CrewAI as the frameworks it avoids. What it does require is an OpenAI-compatible function calling endpoint for the tool use loop, and external capability arrives through MCP servers reached over JSON-RPC on stdio or HTTP.
What memory does 724-office use?
Three layers in one module of about 1,100 lines. Sessions keep the last 40 messages in JSON files, compress on overflow and auto-archive past 100KB. Evicted messages are reduced by the LLM to structured facts, deduplicated by cosine similarity at a 0.92 threshold and stored as vectors in LanceDB, with the top matches injected into the system prompt under a token budget.
What is the xiaowang-v2 directory inside 724-office?
The successor generation, kept in a subdirectory of the same repository: a single Node process with a single WAL SQLite and a hand-written agentic loop, also with zero framework dependency, running 24/7 over WeCom as a personal agent. Its documentation is in Chinese, and it ships design blueprints plus more than 330 offline selftest assertions.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/wangziqi06-724-office)Star history chart for your README
[](https://hysenlabs.com/projects/wangziqi06-724-office)Was this useful?