openclaw-manager

A skill primarily intended for Claude Code — it teaches Claude Code how to safely perform administrative operations on an OpenClaw installation: editing agent workspaces, authoring skills, configuring sub-agents, managing openclaw.json, and troubleshooting the gateway service.

The typical use case: you ssh into a server that has OpenClaw installed, start Claude Code in the terminal, and this skill activates automatically on any mention of OpenClaw, ~/.openclaw, SOUL.md, or equivalent Hebrew phrases. From that point on, Claude Code follows the OpenClaw conventions encoded here instead of guessing file paths, command names, or reload semantics.

The skill is also compatible with any other AgentSkills-spec runtime (OpenClaw agents themselves, etc.), but Claude Code is the target.

OpenClaw is a file-driven AI agent platform that runs as a local Node.js gateway on port 18789. Every agent is a directory of Markdown files (SOUL.md, IDENTITY.md, USER.md, …), every skill is a directory with a SKILL.md, and every tool is a documented capability surfaced by the gateway. Because state lives in files, a careless edit can silently corrupt an agent — this skill encodes the conventions an administrator must follow.

What this skill does

When loaded, it trains the host agent (Claude Code, an OpenClaw agent itself, or any compatible runtime) to:

• Use the canonical workspace layout and never invent, rename, or merge core files.
• Respect skill loading precedence (<workspace>/skills > <workspace>/.agents/skills > ~/.agents/skills > ~/.openclaw/skills > bundled).
• Prefer hot reload over restarts — most Markdown and config edits are picked up automatically.
• Understand that session tools are depth-gated — only orchestrator agents can spawn sub-agents.
• Keep secrets out of Markdown files — auth-profiles.json and models.json are the only right places for keys.

The skill ships with six reference documents covering every area of OpenClaw administration:

File	Covers
SKILL.md	Entry point — pre-flight discovery, decision tree, non-negotiable rules.
references/walkthrough.md	Happy-path setup for a new customer bot, Telegram + WhatsApp, interactive (agent asks the operator).
references/workspace-structure.md	The seven core Markdown files and what belongs in each.
references/skills-authoring.md	Writing, installing, and debugging skills; ClawHub workflow.
references/cli-and-reload.md	openclaw CLI, systemd service, hot-reload semantics, log inspection.
references/agents-and-subagents.md	Agent config, sub-agent spawning, depth limits, trust boundaries.
references/config-files.md	openclaw.json, models.json, auth-profiles.json — structure and safe edits.
references/safety-rules.md	Destructive-op guards, secret handling, tenancy isolation.
references/cost-and-models.md	Where cost is visible, model picker by task type, cache-hit rate, cost-investigation checklist.

Installation

Option 1 — Claude Code (primary intended use)

On your workstation — or, more commonly, over SSH on the OpenClaw host itself:

mkdir -p ~/.claude/skills
curl -L https://github.com/ariweber/openclaw-manager/raw/main/openclaw-manager.zip \
  -o /tmp/openclaw-manager.zip
unzip /tmp/openclaw-manager.zip -d ~/.claude/skills/

Restart Claude Code. The skill auto-triggers on any mention of OpenClaw, ~/.openclaw, SOUL.md, openclaw serve, gateway port 18789, or their Hebrew equivalents (אופןקלאו, להוסיף סוכן, תקן את הבוט, etc.).

This is the intended workflow: ssh into the server, run claude in the terminal, and Claude Code handles the OpenClaw administration with this skill loaded.

Option 2 — Install into an OpenClaw workspace (for OpenClaw agents to use)

If you want an OpenClaw agent itself (not Claude Code) to have access to this knowledge:

# From the machine where OpenClaw is installed
curl -L https://github.com/ariweber/openclaw-manager/raw/main/openclaw-manager.zip \
  -o /tmp/openclaw-manager.zip
unzip /tmp/openclaw-manager.zip -d <workspace>/skills/

# Confirm OpenClaw picked it up
openclaw skills list | grep openclaw-manager

Replace <workspace> with the workspace path from openclaw.json (e.g. ~/.openclaw/workspace for a single-agent install, or /home/<user>/workspaces/<name> for a per-tenant setup).

No restart needed — OpenClaw hot-reloads skill directories. If the skill doesn't appear, check openclaw skills list for rejection reasons (usually a requires mismatch).

Option 3 — Install globally for all OpenClaw workspaces on the host

mkdir -p ~/.openclaw/skills
unzip /tmp/openclaw-manager.zip -d ~/.openclaw/skills/

Any OpenClaw workspace on this host that does not have a higher-precedence copy will load this one.

Option 4 — Install from source (this repo directly)

git clone https://github.com/ariweber/openclaw-manager.git
# For Claude Code:
cp -r openclaw-manager ~/.claude/skills/
# Or for an OpenClaw workspace:
cp -r openclaw-manager <workspace>/skills/

The repo layout is already a valid skill directory — SKILL.md at the root plus a references/ subdirectory. The loose .md files at the repo root are the same reference files; the references/ layout inside the zip is what Claude Code / OpenClaw expects.

Verifying it loaded

openclaw skills list

You should see an openclaw-manager entry with its source path and an eligible: true flag. If it shows eligible: false, the output will print the reason (most common: the skill's requires clause didn't match your environment).

To see the skill's entry point:

openclaw skills show openclaw-manager

Using it

Once installed, the skill activates automatically whenever the loaded agent is asked to do OpenClaw administrative work — you do not invoke it by name. Trigger phrases include anything about:

• Editing an agent's personality, identity, or behavior (SOUL.md, IDENTITY.md, USER.md)
• Adding, installing, or debugging skills
• Creating a new agent or sub-agent
• Changing models or rotating auth tokens
• Restarting or diagnosing the gateway
• Hebrew equivalents: להוסיף סוכן, לערוך את הסוכן, לבנות סקיל, תקן את הבוט

The skill's first act in any session is a pre-flight discovery sequence — it finds the config path, checks which user owns the running gateway, and lists the workspaces on the host — before making any change.

Updating

# From a workspace install
rm -rf <workspace>/skills/openclaw-manager
curl -L https://github.com/ariweber/openclaw-manager/raw/main/openclaw-manager.zip \
  | bsdtar -xvf - -C <workspace>/skills/

OpenClaw will hot-reload the new version within a couple of seconds. Confirm with openclaw skills list.

Repo contents

.
├── SKILL.md                        # Skill entry point
├── agents-and-subagents.md         # Reference (flat copy)
├── cli-and-reload.md               # Reference (flat copy)
├── config-files.md                 # Reference (flat copy)
├── safety-rules.md                 # Reference (flat copy)
├── skills-authoring.md             # Reference (flat copy)
├── workspace-structure.md          # Reference (flat copy)
└── openclaw-manager.zip            # Packaged skill (SKILL.md + references/)

The flat .md files at the repo root are for easy reading on GitHub. The .zip is the installable bundle — it contains the same files but organised in the layout OpenClaw expects (SKILL.md at the top, references under references/).

License

MIT.