Erupt AI Claw
INFO
Erupt AI Claw lets you drive the server directly through natural language — operating annotation-driven data and business, executing shell commands, reading and writing files, and extending custom Skills, as simply as chatting with a colleague.
A self-evolving agent that works and learns simultaneously, supports dozens of models interchangeably, auto-expands its skill library, and intelligently searches conversation memory. The agent learns while working and can run on an ordinary computer.
Usage
<!-- Supported from version 1.14.1 and above -->
<dependency>
<groupId>xyz.erupt</groupId>
<artifactId>erupt-ai-claw</artifactId>
<version>${LATEST}</version>
</dependency>erupt:
ai:
claw:
# Whether to enable Claw
enabled: true
# Whether to allow executing shell commands (default false; must be enabled explicitly)
enable-exec-shell: false
# Allowlist of directories outside the sandbox where shell commands may run (absolute paths)
shell-allowed-paths: []
# Daily background archiving of stale skills (moved to .archive, never deleted)
skill-curator-enabled: true
# A skill unused for this many days is considered stale
skill-stale-days: 30WARNING
Claw has full tool permissions by default. In production environments, restrict the available Tool scope via role configuration (supported from v1.14.3+, see Authorize Tools by Role). Security first!
Token Preparation
The effectiveness of Claw varies by model. Mainstream public models are recommended.
Feature Guide
| Page | Contents |
|---|---|
| Install & Use Skills | 700,000+ Agent Skills; create, patch and run Skills from the conversation |
| Operate Erupt Data by Chat | CRUD any Erupt model in natural language |
| Shell & File Operations | Sandboxed shell and file access, with layered guardrails |
| Control a Browser | Drive a browser through the chrome-devtools MCP |
| Long-Term Memory | Cross-session memory persisted per user |
| Inspect App Runtime (Spring Boot) | Threads, configuration, GC and scheduler info via chat |
| Tool Reference & Role Authorization | Every built-in @Tool, authorized per role |
Contributors
YuePeng
Claude Opus 5 (1M context)