> ## Documentation Index
> Fetch the complete documentation index at: https://docs.clawkitchen.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Agents and skills

> This doc explains the practical mental model behind agents, teams, tools, and skills in ClawRecipes.

***

## What is an agent?

In OpenClaw, an agent is a configured assistant with:

* a workspace
* an identity/persona
* tool permissions
* a model/runtime configuration

In ClawRecipes, an **agent recipe** scaffolds that into a folder like:

```text theme={null}
~/.openclaw/workspace-<agentId>/
```

Typical files include:

* `SOUL.md`
* `AGENTS.md`
* `TOOLS.md`

***

## What is a team?

A team is a set of role agents sharing one workspace root.

Example:

```text theme={null}
~/.openclaw/workspace-development-team/
```

Inside that workspace you usually get:

* `work/` ticket lanes
* `roles/` role folders
* `shared/` artifacts
* `notes/` planning/status docs

Role agent ids usually look like:

```text theme={null}
development-team-lead
development-team-dev
development-team-test
```

***

## What is a skill?

A skill is a packaged capability.

Examples:

* Gmail / Calendar tools
* web / places integrations
* social/X tooling
* helper CLIs or packaged instructions

Skills can expose tools, configuration, scripts, or helper behavior.

***

## Tool policy: what an agent is allowed to do

Recipes can define tool policy and apply it into OpenClaw config when you scaffold with `--apply-config`.

Common patterns:

### Safe-by-default

```yaml theme={null}
tools:
  profile: coding
  allow: ["group:fs", "group:web"]
  deny: ["exec"]
```

### Dev / devops style

```yaml theme={null}
tools:
  profile: coding
  allow: ["group:fs", "group:web", "group:runtime"]
  deny: []
```

### About `exec`

`exec` is the shell-command tool.

If an agent needs to run commands like:

* `npm test`
* `git status`
* `pnpm build`
* `pytest`

then its runtime/tool policy must allow that capability.

***

## How to change an agent after scaffolding

There are two layers.

### 1) Workspace files

Edit the agent/team files directly:

* `SOUL.md`
* `AGENTS.md`
* `TOOLS.md`
* notes/status docs

### 2) OpenClaw config

If you need to change permissions, identity, or config-level behavior, update the OpenClaw agent config.

A common way is to re-run scaffold with `--apply-config`.

Examples:

```bash theme={null}
openclaw recipes scaffold project-manager --agent-id pm --overwrite --apply-config
openclaw recipes scaffold-team development-team --team-id development-team --overwrite --apply-config
openclaw gateway restart
```

***

## Installing skills

### Install a single skill globally

```bash theme={null}
openclaw recipes install-skill agentchat --yes
```

### Install a skill for one agent

```bash theme={null}
openclaw recipes install-skill agentchat --yes --agent-id dev
```

### Install a skill for one team

```bash theme={null}
openclaw recipes install-skill agentchat --yes --team-id development-team
```

### Install the skills declared by a recipe

```bash theme={null}
openclaw recipes install-skill development-team --yes
```

***

## Where skills get installed

By default:

* global skills → `~/.openclaw/skills/`
* agent-scoped skills → `~/.openclaw/workspace-<agentId>/skills/`
* team-scoped skills → `~/.openclaw/workspace-<teamId>/skills/`

***

## Removing skills

There is not currently a first-class ClawRecipes remove-skill command.

For now, remove the skill folder manually and restart the gateway.

***

## Removing a team

```bash theme={null}
openclaw recipes remove-team --team-id development-team --plan --json
openclaw recipes remove-team --team-id development-team --yes
openclaw gateway restart
```

This is safer than manually deleting everything by hand.

***

## Good practical advice

* keep high-risk tools limited to the roles that actually need them
* encode durable tool-policy changes in the recipe, not just ad hoc config edits
* use team workspaces when you want collaboration and durable ticket lanes
* use single agents when you want one specialized role without the extra team machinery

***

## Useful next reads

* [TEAM\_WORKFLOW.md](team-workflow)
* [RECIPE\_FORMAT.md](recipe-format)
* [COMMANDS.md](commands)