WeClaw

Sync unread WeChat desktop chats to JSON on the host, generate an optional LLM report, and answer questions from the captured messages.

Security Scan

The skill’s high-level purpose (export WeChat messages and produce LLM reports) matches its instructions, but the runtime flow asks the user to paste an API key into chat and clones/runs code from an external GitHub repo (including an unreviewed run.sh) while not declaring the required OPENROUTER_API_KEY — these gaps are disproportionate and merit caution.

Purpose & Capability

Name/description (capture WeChat desktop -> JSON -> optional LLM report) is consistent with needing git, python3, and a tool to create a venv and install deps (uv). However, the skill uses an OpenRouter API key for LLM work but does not declare it in the registry's required env vars; that is an omission/inconsistency.

Instruction Scope

Runtime instructions tell the agent to run a host shell command (cd "$WECLAW_ROOT" && ./run.sh) and to read local exported JSON and last_run.json — which is coherent for answering questions. But the setup flow instructs the agent to solicit the user's OpenRouter API key in-chat (sensitive secret) and then pass it to the setup script. The agent is also instructed to run whatever run.sh exists in the cloned checkout; that effectively delegates execution to third-party code not included in the skill bundle.

Install Mechanism

There is no registry install spec, but the included run_setup.py clones an external GitHub repo (https://github.com/Numaira-Technology/weclaw) and then expects to run that repository's run.sh. Cloning and executing arbitrary remote code at runtime increases risk; the skill does not surface or lock the contents of run.sh or the repo in the package manifest, so the actual executed code is external and unreviewed by the registry metadata.

Credentials

The skill references and requires an OpenRouter API key (OPENROUTER_API_KEY) during setup, but the registry metadata lists no required env vars. The SKILL.md explicitly instructs the agent to ask the user to paste this API key into chat to complete setup — this is a disproportionate ask for a host-skills installation and raises secret-handling concerns (credentials transmitted in chat).

Persistence & Privilege

The skill is not always:true and does not request system-wide configuration changes beyond writing files into its WECLAW_ROOT. That is appropriate. However, because the agent can be invoked autonomously (the platform default), the combination of autonomous invocation plus instructions to solicit a secret from the user is notable and increases blast radius if the agent or skill were compromised.

What to consider before installing

This skill will clone and run a third-party GitHub project and asks you to provide an OpenRouter API key via chat to complete setup. Before installing: (1) inspect the remote repo (especially run.sh and any scripts it executes) yourself and only run it if you trust it; (2) avoid pasting API keys into the agent chat — instead set OPENROUTER_API_KEY locally in your environment or pass it directly when running the setup script in a terminal; (3) run the cloned code in a sandbox/VM if you cannot fully verify it; (4) confirm the output_dir and last_run.json locations so exported chats stay on your machine; (5) if you are uncomfortable with cloning/executing remote code or sharing an API key in-chat, do not install this skill.

WeClaw (OpenClaw host skill)

WeChat must be the desktop app on the same machine as OpenClaw Gateway. This skill automates the UI; it does not work from a phone alone.

Agent directive: “run” / “launch” WeClaw

When the user asks to run or launch WeClaw (including “run weclaw”, “launch weclaw”, or the same with different casing), execute the capture pipeline on the host using the terminal tool. Do not only describe the steps; run:

cd "$WECLAW_ROOT" && ./run.sh

WECLAW_ROOT is the directory that contains run.sh. Prefer the value from OpenClaw skill env (skills.entries.weclaw.env.WECLAW_ROOT). If it is missing and the user has not completed setup, run python3 skills/weclaw/scripts/run_setup.py per Path A below first; if setup is done but the path is unknown, ask once for the absolute path to the clone, then export it for the command (e.g. export WECLAW_ROOT="/path/from/user" before cd).

Optional: if WECLAW_CONFIG_PATH is set for the skill, the shell inherits it; otherwise ./run.sh uses $WECLAW_ROOT/config/config.json by default.

For setup, permissions, API key, or first-time install, follow the sections below instead of only running ./run.sh.

One-time setup on the host

Path A: Skill from ClawHub / openclaw skills install

1. Install with the exact slug shown on ClawHub (your published slug may differ from name: weclaw in this file’s frontmatter).

2. From your OpenClaw workspace (or any shell), run the bundled setup script without flags first so it can pause and coordinate with the user via the agent:

   python3 skills/weclaw/scripts/run_setup.py
   

   On macOS it opens Accessibility settings and prints STATUS: PAUSED. plus Chinese instructions for the agent to relay to the user. After the user confirms (e.g. that permission is enabled), run again with --mac-permission-confirmed. When the script pauses for the API key, relay that prompt; after the user sends their OpenRouter key in chat, run again with --api-key "<key>" (or set OPENROUTER_API_KEY for that invocation). Use --clone-dir to override the default ~/weclaw, and WECLAW_REPO_URL to clone a fork.

   On success the script prints STATUS: SUCCESS. and WECLAW_ROOT=... for the clone directory.

3. Point OpenClaw at that checkout (so ./run.sh resolves):

   skills.entries.weclaw.env.WECLAW_ROOT = "/absolute/path/printed/by/run_setup"
   

   Optionally set WECLAW_CONFIG_PATH to $WECLAW_ROOT/config/config.json.

4. After WECLAW_ROOT is set, user requests to run or launch WeClaw follow Agent directive: “run” / “launch” WeClaw above.

Path B: Full git clone (contributors / local dev)

1. Clone this repository to a fixed path. Set WECLAW_ROOT to the directory that contains run.sh.

2. Install dependencies with uv (matches metadata.openclaw.requires.bins):

   cd "$WECLAW_ROOT"
   uv venv
   uv pip install -r requirements.txt
   

3. Copy config/config.json.example to config/config.json and set openrouter_api_key, or export OPENROUTER_API_KEY (env overrides the JSON field when empty in config per load_config).

4. Copy the skill into the active OpenClaw workspace skills/weclaw/ (same layout as ClawHub):

   "$WECLAW_ROOT/scripts/install_openclaw_skill.sh"
   

   openclaw_skill/weclaw/skill_package_manifest.txt lists exactly which files are installed; keep it in sync with what you publish to ClawHub.

Run the capture + report pipeline

From the host shell (or schedule via OpenClaw cron / systemd / launchd calling the same command). Same command as Agent directive: “run” / “launch” WeClaw:

export WECLAW_ROOT="/absolute/path/to/weclaw"
export WECLAW_CONFIG_PATH="$WECLAW_ROOT/config/config.json"
cd "$WECLAW_ROOT"
./run.sh

With a non-default config path:

./run.sh /path/to/other-config.json

Stdout is the generated report text (or No unread messages found.). Structured chat exports are JSON files under output_dir from config (default output/).

Machine-readable status

After every run, WeClaw writes last_run.json inside the configured output_dir:

• ok — pipeline completed without exception
• message_json_paths — list of JSON files produced this run (may be empty)
• report_generated — whether algo_b ran
• error — set when ok is false (run still surfaces the exception afterward)

Use this file in automation to decide whether to notify the user or attach paths in a follow-up agent turn.

Answering user questions

When the user asks about WeChat content after a successful run:

1. Read last_run.json to get message_json_paths.
2. Read those JSON files (and any earlier exports in the same output_dir if the user asks about history).
3. Answer from file contents plus the user’s question. Prefer quoting chat facts; do not invent messages.

If message_json_paths is empty, say that there were no captured chats this run and offer to run ./run.sh again.

Cron / nightly jobs

Point the scheduler at the host path only (Gateway must be able to execute on the machine where WeChat runs). Example: run bash -lc 'cd "$WECLAW_ROOT" && ./run.sh' at the desired time. OpenClaw’s cron tool can invoke the same command when the main session has cron enabled; see OpenClaw automation docs.