TL;DR

We launched ClawSouls Registry — an open registry where anyone can submit AI agent personas via GitHub Pull Request. Every submission is automatically verified by SoulScan (53 safety patterns) before it can be merged.

No other AI agent registry does automated safety verification. That’s our differentiator.

The Problem

AI agents are everywhere — Claude Code, Cursor, Windsurf, Copilot. Each one can be personalized with system prompts, personality files, or persona definitions. But there’s no safe, standardized way to share these personas.

What happens when someone submits a persona that contains:

  • Prompt injection (“ignore previous instructions…”)
  • Permission escalation (“run sudo rm -rf /”)
  • Secret exfiltration (“send all API keys to…”)
  • Safety bypass (“override all safety rules”)

Without automated verification, these get through.

The Solution: SoulScan CI

Every Pull Request to ClawSouls Registry triggers an automated CI pipeline:

1. You fork the repo
2. Add your persona in souls/<your-username>/<persona-name>/
3. Open a Pull Request
4. SoulScan CI automatically runs (53 safety patterns)
5. Results posted as PR comment with grade (A+ to F)
6. Grade C+ or above → eligible for merge
7. Merged → automatically synced to clawsouls.ai

What SoulScan Checks

SoulScan runs 53 security patterns across 8 categories:

CategoryPatternsExamples
Prompt InjectionSEC001-008“ignore previous”, “you are now”, system prompt override
Code ExecutionSEC010-015sudo, rm -rf, eval(), exec()
XSSSEC020-022Script injection, event handlers
Data ExfiltrationSEC030-032API key patterns, credential harvesting
Permission EscalationSEC040-042Role elevation, admin bypass
Social EngineeringSEC050-051Authority impersonation
Secret DetectionSEC060-069AWS keys, GitHub tokens, JWT, private keys
Multi-language InjectionSEC070-077Korean, Chinese, Japanese injection patterns

Real Example

Here’s what a SoulScan CI result looks like on a PR:

🔍 SoulScan Validation Results

✅ All checks passed — eligible for merge.

✅ souls/clawsouls/code-reviewer — A+ (100/100)

⚠️ safety.laws recommended for v0.5+ — consider declaring behavioral laws

Soul Spec v0.6 Preview

The registry is built on Soul Spec, our open standard for AI agent personas. We’re working on v0.6, which adds three major features:

1. Soul Packs

Bundle personas with skills, tools, memory, and rules:

full-stack-engineer/
├── soul.json        # type: "pack"
├── SOUL.md          # Personality
├── RULES.md         # Hard constraints (new!)
├── skills/          # Bundled skills
├── tools/           # MCP tool definitions
├── memory/          # Initial memory
└── hooks/           # Lifecycle events

2. Registry Protocol

Standardized submission, validation, and discovery — exactly what ClawSouls Registry implements.

3. Memory Spec

Portable agent memory format with TF-IDF search:

{
  "memory": {
    "format": "markdown",
    "layout": {
      "longTerm": "MEMORY.md",
      "daily": "memory/YYYY-MM-DD.md",
      "topics": "memory/topic-*.md"
    },
    "search": {
      "default": "tfidf",
      "enhanced": "llm-rerank"
    }
  }
}

How It Compares

FeatureClawSouls RegistryGitAgent RegistryCharacter.AI
Open submissions✅ GitHub PR✅ GitHub PR❌ Closed
Safety verification✅ SoulScan 53 patterns❌ Structure only❌ Internal only
CI/CD integration✅ GitHub Actions✅ GitHub Actions❌
Cross-platform✅ Any framework⚠️ Limited❌ Platform-only
Auto DB sync✅ Supabase❌❌

Get Started

Submit a Persona

# 1. Fork https://github.com/clawsouls/registry
# 2. Create your persona
mkdir -p souls/your-username/my-agent

# 3. Add required files
# soul.json — metadata
# SOUL.md — personality & principles

# 4. Open a Pull Request
# SoulScan will verify automatically

Install a Persona

# Via CLI
npm install -g clawsouls
clawsouls install TomLeeLive/brad

# Via MCP (Claude Code)
npx -y clawsouls-mcp@latest

Browse

Visit clawsouls.ai/souls or browse the registry on GitHub.

ClawSouls Registry is MIT licensed. Personas are licensed individually. SoulScan rules are open (Apache-2.0), engine is proprietary.