Diagnose and fix Openclaw AI agent workspaces. One command tells you what's working, what's broken, and fixes it for you.
npm i -g souls# Auto-discover workspaces and pick one
souls
# Diagnose a specific workspace
souls doctor ~/.openclaw/workspace-garyAuto-discovery checks ~/.openclaw, ~/.clawdbot, and ~/.moltbook.
The CLI reads your SOUL.md, AGENTS.md, TOOLS.md, USER.md, and MEMORY.md, loads quality rules from souls.zip, and runs an interactive diagnostic with per-category scores, specific findings, and actionable fixes.
The soul that was sharp on day one becomes bloated by week two. Instructions contradict each other. Delegation boundaries blur. Anti-patterns creep in as generic filler. Nobody notices until the agent starts producing mediocre output and no one can explain why.
souls catches drift before it becomes a problem. The rules are distilled from 160+ arXiv papers across agent persona design, multi-agent coordination, autonomous systems, and production reliability, plus 500+ hours of production OpenClaw usage.
When the research says detailed experiential identities outperform generic role labels (Xu 2023, EMNLP 2024), that becomes a rule that flags "You are a helpful assistant" as a failure. When "Lost in the Middle" (Liu et al., TACL 2024) shows models ignore mid-context instructions, that becomes a rule about where to place hard rules in your soul. When scaling law research shows coordination overhead grows quadratically with team size, that becomes a rule about agent span of control.
Every rule traces back to either a published finding or a production failure we hit firsthand and wrote a lesson about.
Rules are constantly updated to keep up with new research, findings, and best practices.
Identity quality. Is your SOUL.md a lived-in partner or a generic job description? Are beliefs encoded as experience or compliance? Can someone predict the agent's behavior just from reading the soul?
Delegation boundaries. Does every agent know what's not its domain? Are owners named? Are fallback rules explicit? Agents without boundaries drift into out-of-domain work the moment something seems urgent.
Security. Credentials in workspace files. Secrets in markdown. Exposed connection strings. The stuff that gets committed accidentally and lives there forever.
Structural integrity. Missing sections, oversized files, contradicting rules, dead weight that wastes context tokens without affecting behavior.
Found issues come with proposed fixes. Select a finding and apply the fix directly. The fix engine carries soul engineering judgment: it preserves voice, protects identity content, and refuses destructive changes.
Rules are maintained at souls.zip and fetched at runtime. Your workspace content is sent to whichever model provider your OpenClaw config selects for analysis. souls.zip and souls cli does not use telemetry.
- Node.js 18+
- OpenClaw setup under
~/.openclaw,~/.clawdbot, or~/.moltbook - At least one configured model provider
| Key | Action |
|---|---|
↑ ↓ |
Navigate |
Enter |
Select |
r |
Run another workspace |
q |
Quit |
Esc |
Back |
If npm i -g souls fails with EACCES on macOS/Linux:
sudo npm i -g soulsOnly needed when npm globals live in a protected directory (/usr/local/lib/node_modules). Not needed with nvm, fnm, or volta.