Question 1

Which platforms work end to end?

Accepted Answer

Windows is the best documented, with automatic discovery and initialisation. Linux works but depends on the WeChat build, your Python dependencies and user permissions, and needs sqlcipher development headers installed. macOS can query chats but has no automatic initialisation, so you supply credentials yourself.

Question 2

Is the MCP tool surface actually read-only?

Accepted Answer

No. The README says the tool list is derived from the assistant's own tool table and is not purely read-only. Write calls require an explicit confirm, and there is no tool at all for sending messages, publishing articles or changing configuration.

Question 3

How does it get my database key, and what is the account risk?

Accepted Answer

You close WeChat, run init, log in when prompted, and the CLI discovers the data directory and extracts the keys. The README states plainly that third-party tooling may trigger account risk controls, restrictions or bans, and that checking local law and WeChat's terms is your responsibility.

Question 4

What does Jev actually do in this tool?

Accepted Answer

Very little, and only in passing: setting typesafeApiKey switches article topic and relevance judgement for the digest from the LLM parsing path to Jev decision models. Cutoffs, model choice and any quality difference are not documented.

Question 5

What are the install prerequisites and setup steps?

Accepted Answer

Node.js 22.13+ and Python 3.10+, plus a logged-in Windows WeChat. Run npm install -g weflow-cli, then weflow-cli check, then pip install the requirements; WeChat 3.x also needs requirements-3x. Close WeChat and run weflow-cli init, which discovers the data directory and extracts the keys, and you can point it with init --path. On Linux first install libsqlcipher-dev plus sqlcipher3, cryptography, html2text and zstandard.

Question 6

What can it do beyond reading chats?

Accepted Answer

It exports chats to HTML, Excel, TXT or JSON; builds Official Account digests with AI summaries; syncs WeRead notes into an Obsidian vault with semantic search and RAG; exposes conversations, favourites, moments, todos and the knowledge base to editors through MCP; and runs a local assistant in WeChat plus a desktop panel, with a self-contained 2D or 3D concept graph that opens by double-click and needs no server.

Question 7

How does the assistant protect what it sends out?

Accepted Answer

Only the final question and reply text go to the configured LLM, with the agent loop, three-layer memory and database queries all local. A new install defaults to assistantPrivacy=strict, which redacts PII such as phone numbers, emails and links from tool output and keeps third-party chat bodies from leaving the machine, with balanced and open as looser modes and ollama as a fully offline engine. The assistant also rejects all senders until a whitelist is set.

Weflow CLI

Highlights

Questions about this entry

More Jev resources like this

From the Jev community