wechat-intelligence-hub Is a Local CRM Hidden Inside Your WeChat History

A read-only intelligence layer that turns chat logs into follow-ups, deal tracking, and daily briefings, using deterministic heuristics, SQLite, and agent-friendly CLI tools.

8 min read • View on GitHub • More from Rion-Wu-tech

A wide editorial scene shows a desk covered in stacked chat transcripts, a local hard drive, and trays labeled with business states such as pending reply, promise, and deal. A hand sorts paper into the trays while a sealed laptop sits untouched in the background, explaining the project's read-only, local-first model.
This repo treats chat history like a business archive, not a live messaging surface.
Key Takeaways

Most WeChat tools try to sit in the chat flow. This one refuses. It reads local history, extracts business signals, and leaves the app alone.

That refusal is the point. wechat-intelligence-hub is not a bot wrapper, and it is not a cloud CRM in disguise. It is a local intelligence layer built for people who need to turn old chats into new work.

The surprising part is what it does not do

The repo starts by rejecting the obvious product shape. It does not inject into WeChat, it does not control the UI, and it does not try to keep a conversation going on your behalf. Instead, it treats local databases as a source of truth and builds analysis on top.

Reader 核心不获取密钥、不重签名、不注入、不Hook 微信;可选的实验性接入助手有独立授权和副作用边界

That boundary matters because it changes the trust model. A read-only system can be audited, reasoned about, and removed without leaving a behavioral footprint in the app itself.

Why local-first matters here

The project's reader layer sits behind a clean-room boundary. It works with local SQLite and SQLCipher-backed data, rejects unsafe paths, and keeps the acquisition layer separate from the intelligence layer. In practice, that means privacy is enforced by structure, not by promises.

The pipeline is the product: local data in, structured intelligence out, with the opportunity store as a side branch and a feedback loop.

The most important design choice is the one the repo keeps repeating: read, classify, report. Not write, not chat, not automate the live app.

A close-up shows a stream of message bubbles being scanned by mechanical rule filters. One phrase like 'I'll send it tomorrow' is caught by a selector arm and routed into a tray for promises while unrelated chat bubbles fall away, explaining how deterministic heuristics isolate actionable signals.
Before anything becomes intelligence, the repo uses rules to separate commitments from noise.

Why the opportunity store feels like a CRM

This is where the project becomes legible to a business user. The opportunity store separates a raw candidate from an active opportunity, then moves it through stages like New Lead and Received Payment. That sounds simple, but it is the bridge between message history and a working pipeline.

LayerWhat it tracksWhy it matters
CandidateA possible lead detected from chat signalsKeeps raw signals separate from confirmed business
OpportunityA triaged, active deal with a stageTurns chat into a process, not just a search result
Stage machineNew Lead to Received PaymentShows where follow-up work is stuck
Briefing outputDaily summaries and remindersMakes the pipeline visible to a human operator

The important move here is not classification for its own sake. It is state management. A message about a quote, a draft, or a callback can be promoted from noise into something a human can actually act on.

Deterministic heuristics beat vibe-based AI in this repo

The code leans on rules because the job is specific. If the system needs to spot a commitment, a commercial term, or a reply promise, regex and stateful logic are often more dependable than a general model trying to sound helpful.

COMMERCIAL_TERMS = (...)
PROMISE_TERMS = (...)

if matches_commercial_term(message) and infer_promise_action(message):
    mark_as_follow_up_candidate(message)

That approach is boring in the best way. It is repeatable, cheap to run, and easier to audit when the output affects a sales lead or a client follow-up.

WeChat Intelligence Hub does one thing: it turns your local WeChat conversations into searchable, auditable, and actionable personal intelligence. It is not a collection of prompts or a tutorial. It is a standalone project with executable entry points, clear boundaries, and test suites.

Rion-Wu-tech, Author/Maintainer · Rion-Wu-tech/wechat-intelligence-hub README

Why Codex skills change the shape of the product

The skills/ layer makes the repo feel agent-ready instead of just scriptable. An AI agent can call the CLI, pull context from the local reader, and surface a ranked slice of history without being allowed to rewrite the source of truth.

That is a subtle but important shift. The agent is not the product. It is a consumer of the product's outputs, which keeps the human in control while still making the workflow faster.

What it replaces, and what it does not

The cleanest comparison is not between features. It is between jobs. Chatbot wrappers respond in place. Local viewers help you search. SaaS CRMs formalize sales work. This repo extracts operational intelligence from a private archive and keeps the whole process local.

Project typePrimary goalRead/write behaviorData locationBest for
wechat-intelligence-hubExtract local business intelligenceRead-onlyLocal SQLite or SQLCipherFollow-ups, briefs, opportunity tracking
chatgpt-on-wechatChat in the app with an LLMRead and writeCloud or connected servicesLive conversational automation
wechatyBuild active WeChat botsRead and writeDepends on puppet providerMessage automation and integration
wechat-local-viewerInspect chat historyRead-onlyLocal dataSearching and browsing archives
SaaS WeChat CRMManage customer relationshipsRead and writeVendor cloudTeam sales workflows

So the project is not trying to win a chatbot race. It is building a different category: private, deterministic business intelligence from your own message history.

Sources and project trail