EverClaw
ReviewAudited by ClawScan on May 10, 2026.
Overview
EverClaw is a coherent local research and knowledge-base skill, but it is always-on and persistent, so users should expect automatic local KB checks and long-running learning when invoked.
This appears purpose-aligned rather than malicious. Before installing, make sure you want an always-on local knowledge-base skill that can run long research tasks, spawn subagents, store downloaded material under the configured KB path, and reuse that material in future answers. Keep auto/proactive modes disabled unless needed, protect any provider API keys, and periodically review or clean the local KB.
Findings (6)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
The agent may inspect the skill’s local state and knowledge base at session start even when the user did not explicitly call the skill.
The skill asks the agent to run local startup checks without displaying them in chat. This is disclosed in the skill file and appears intended to reduce noise, but it is automatic behavior users should understand.
**At the beginning of every session**, perform the following check silently (do not inform the user):
Install only if you want always-on KB behavior; use the documented master switch or disable the skill if you do not want automatic session-start checks.
Future answers may be shaped by whatever has been stored in the local knowledge base, including downloaded web material.
The skill intentionally creates persistent local memory and prioritizes it for future answers. This is purpose-aligned, but stored or retrieved content can become stale, incorrect, or influence later sessions.
learned topics are stored locally and should be used to answer related questions with source citations instead of relying on parametric memory
Review citations, keep the KB curated, avoid storing secrets or private material as research content, and clear or disable the KB if it becomes untrusted.
A learning run may continue for a long time and produce persistent local files.
The skill explicitly supports long-running autonomous activity. The artifacts show this as a disclosed core function, not hidden persistence, but it can consume time, tokens, API quota, or local storage.
You are a domain expert system with **long-running autonomous learning capability**. You can research any topic for hours
Use explicit duration limits, keep proactive/auto modes off unless needed, and monitor long learning jobs and storage growth.
Parallel research may increase API usage, resource consumption, and the amount of material downloaded into the local KB.
The skill can coordinate multiple subagents for research. This is consistent with its learning purpose and has a configurable limit, but users should notice the parallel automation.
Spawns parallel subagents (up to `maxChildrenPerAgent`); pipeline-batches if subtopics exceed the limit
Set a reasonable `maxChildrenPerAgent`, use shorter `--hours` limits for unfamiliar topics, and review generated results before relying on them.
If configured, the skill’s research activity may use the user’s provider quota or incur provider-side usage.
The README recommends configuring a Google AI Studio key for web search. This is expected for the research workflow, and the visible artifacts do not show credential logging or exfiltration.
"apiKey": "{YOUR_GOOGLE_AI_STUDIO_KEY}"Store provider keys only in trusted configuration, monitor usage, and avoid pasting secrets into normal chat messages.
Users may not have full assurance that the registry package exactly matches the homepage repository.
The registry source is not identified, although a homepage is listed. The provided artifacts do not show remote install scripts or hidden dependencies, so this is a provenance notice rather than a behavioral concern.
Source: unknown; Homepage: https://github.com/EdgePro001/ClawExpert
Install from trusted registries, compare with the public repository if provenance matters, and review updates before enabling always-on behavior.