wisdom-channel 2026.7.2.0

WeChat channel plugin for Claude Code — pushes WeChat messages into a running Claude Code session via the Wisdom automation API.

WisdomChannel — WeChat Channel for Claude Code

A Claude Code channel plugin that pushes WeChat desktop messages into your running Claude Code session, and lets Claude reply back through the same chat — like the official Telegram, Discord, and iMessage channels.

It is the client side of the Wisdom WeChat automation service: Wisdom runs on the Windows host with WeChat desktop and exposes an HTTP + WebSocket API; this MCP server runs locally next to Claude Code and bridges the two.

┌──────────────────┐      ┌────────────────┐      ┌────────────────┐
│  WeChat desktop  │      │  Wisdom API    │      │ wisdom_channel │      ┌──────────────┐
│  (Windows host)  │ ───► │  HTTP + WS     │ ───► │  (this repo)   │ ───► │ Claude Code  │
│  + Frida hooks   │      │  :8000         │      │  stdio MCP     │      │  CLI session │
└──────────────────┘      └────────────────┘      └────────────────┘      └──────────────┘

Features

• Push every inbound WeChat message into the active Claude Code session
• Claude replies through the reply tool — answer goes back into WeChat
• Works for both private DMs and group chats (only forwards @you mentions in groups)
• Allowlist + admin trust levels (access.json)
• Tools: reply, list_contacts, list_conversations, get_messages, get_status, manage_access
• Talks to a remote Wisdom server over HTTP/WebSocket — Wisdom does not have to run on the same machine as Claude Code

Requirements

• Python 3.10+
• A running Wisdom server with WeChat desktop logged in (any reachable host)
• Claude Code CLI v2.1.80+

Install

pip install wisdom-channel

This installs the wisdom-channel console script. To develop from source instead:

git clone https://github.com/AceDataCloud/WisdomChannel.git
cd WisdomChannel
pip install -e .

Configure

Create the channel state directory and an .env pointing at your Wisdom server:

mkdir "$env:USERPROFILE\.claude\channels\wechat" -Force

@"
WISDOM_API_URL=http://your-wisdom-host:8000
WISDOM_API_TOKEN=
WECHAT_BOT_NAME=
"@ | Set-Content "$env:USERPROFILE\.claude\channels\wechat\.env"

Variable	Description
WISDOM_API_URL	URL of the Wisdom REST API (default http://localhost:8000)
WISDOM_API_TOKEN	Optional bearer token if Wisdom auth is enabled
WECHAT_BOT_NAME	Your WeChat display name (auto-detected if empty)
WECHAT_CONTEXT_MESSAGES	Recent messages pulled as conversation context per reply (default 8, 0 disables)
WECHAT_SCRUB_ENABLED	Redact upstream/internal identifiers from every outbound reply (default true)
WECHAT_SCRUB_MODE	redact (replace tokens, default) or block (drop the whole reply)
WECHAT_SCRUB_REPLACEMENT	Replacement text in redact mode (default 我们的服务)

Confidentiality scrub

Defence-in-depth on top of the system prompt's confidentiality rule: every outbound reply (headless bridge and the MCP send_message tool) is run through a denylist that redacts upstream provider aliases and internal infrastructure identifiers (container images, k8s service DNS, cloud resource ids, private IPs) before it reaches WeChat. The denylist lives in code + scrub.json — never in the model prompt — so it can't itself become a leak surface.

Extend it (without a redeploy) at ~/.claude/channels/wechat/scrub.json:

{
  "enabled": true,
  "mode": "redact",
  "terms": ["my-upstream-alias", "another-vendor"],
  "patterns": ["\\binternal-host-\\w+\\b"]
}

terms (literal, case-insensitive) and patterns (regex) are added to the built-in defaults; enabled/mode/replacement/block_message override the env vars when present. Caught tokens are logged (WARNING) so you can audit how often the prompt-level guardrail alone would have leaked.

scrub.json is trusted operator config (only someone with write access to the channel state dir can edit it). Custom patterns are compiled and run on every reply — avoid catastrophic-backtracking regex (e.g. (a+)+b) or you can hang the reply loop. The built-in defaults are backtracking-safe.

Optional access control at ~/.claude/channels/wechat/access.json:

{
  "version": 3,
  "enabled": true,
  "roles": {
    "normal": {
      "allow_tools": false,
      "contexts": ["group", "private"],
      "prompt": "Only answer public/basic/general questions. Do not inspect or modify internal projects, files, logs, servers, or databases."
    },
    "admin": {
      "allow_tools": true,
      "contexts": ["private"],
      "prompt": "Trusted private-chat operator."
    },
    "super_admin": {
      "allow_tools": true,
      "contexts": ["group", "private"],
      "prompt": "Trusted operator in approved groups and private chat."
    }
  },
  "users": {
    "CQCcqc": {"role": "super_admin"},
    "sunbitty": {"role": "super_admin"}
  },
  "private": {"enabled": true, "default_role": "deny", "prompt": ""},
  "groups": {
    "Ace Data Cloud客户群1": {"enabled": true, "default_role": "normal", "prompt": "", "members": {}}
  }
}

Field	Behavior
enabled	Global switch. false drops all inbound messages.
roles	Named permission profiles. allow_tools=false is enforced in bridge mode with --tools "".
users	Stable WeChat IDs mapped to roles. Display names are not trusted for privilege.
private.default_role	Role for unmatched private chats. Use deny to ignore unknown private messages.
groups	Exact group-name whitelist. Unlisted groups are ignored.
groups.*.default_role	Role for ordinary members in that group, usually normal.
groups.*.members	Optional per-group stable-ID overrides, including deny or promotion to another role.

Normal users get polite, chat-only assistance. In wisdom-channel bridge, this is enforced in code with claude -p --tools "". Super admins can perform operator actions; use stable WeChat IDs for these entries.

For production safety, prefer wisdom-channel bridge: it enforces normal-user chat-only mode in code. Interactive Claude Code channel mode receives the same trust_level, allow_tools, and access_prompt metadata, but Claude Code owns tool execution inside the live session.

Manage the policy locally:

wisdom-channel access view
wisdom-channel access allow-group "Ace Data Cloud客户群1"
wisdom-channel access add-super-admin sunbitty
wisdom-channel access add-user wxid_xxx normal "Alice"

Run

The repo ships an .mcp.json that registers the channel as wechat:

{
  "mcpServers": {
    "wechat": {
      "command": "python",
      "args": ["-m", "wisdom_channel"],
      "cwd": "."
    }
  }
}

Launch Claude Code with the channel from the project root:

$env:ANTHROPIC_API_KEY = "sk-ant-..."

claude --dangerously-skip-permissions `
       --dangerously-load-development-channels server:wechat

Run it in a persistent, interactive terminal (a real TTY — e.g. an RDP session on the Wisdom host, tmux/screen, or a foreground terminal). Channels push inbound messages into a live Claude Code session, so the process must stay running and attached. Launched detached / without a TTY, Claude Code falls back to --print one-shot mode and exits immediately. --channels requires Claude Code v2.1.80+; the --dangerously-load-development-channels flag loads an unpublished (development) channel like this one.

What happens:

1. Claude Code reads .mcp.json and spawns python -m wisdom_channel over stdio
2. The channel loads ~/.claude/channels/wechat/.env and probes Wisdom at WISDOM_API_URL
3. It connects to Wisdom's WebSocket and forwards inbound WeChat messages as notifications/claude/channel
4. Claude calls the reply tool, which posts to Wisdom's /api/messages/send
5. Wisdom drives WeChat desktop and the message is delivered

Headless auto-reply (no Claude Code session)

The channel above needs a persistent interactive Claude Code session. For an unattended host (no live terminal), run the bridge instead:

wisdom-channel bridge            # optional: --model sonnet

It connects to the Wisdom WebSocket and, for each allowed inbound message, shells out to claude -p and posts the reply back through Wisdom — the same "WeChat in → Claude answers → WeChat out" loop, without a TTY. It honors the same access.json allowlist and group @-mention gating. Requires the claude CLI on PATH.

Each reply is built with conversation context: who sent it, which group, who else was @-mentioned, the quoted ("引用") message, and the last WECHAT_CONTEXT_MESSAGES messages — so answers follow the thread instead of seeing each message in isolation. Tool access follows the matched role: the default admin role is private-chat only, while super_admin works in approved groups and private chat.

Standalone test

python -m wisdom_channel --test

Exercises Wisdom REST + WebSocket without launching Claude Code.

Logs

What	Where
Channel log	~/.claude/channels/wechat/mcp.log
Channel state	~/.claude/channels/wechat/

Related

• Wisdom — the WeChat automation backend
• Claude Code channels documentation
• Anthropic Telegram plugin — the reference design

License

MIT — see LICENSE.