| name | capabilities |
| description | Show what this NanoClaw agent can do — installed skills, MCP tools, group config. Read-only. Use when the user asks "what can you do", "qué sabes hacer", "what's installed", or runs /capabilities. |
/capabilities — Agent Capabilities Report
Generate a clean, read-only summary of what this agent group is wired to do. No state changes — pure introspection.
Run when the user asks any of:
- "qué puedes hacer" / "what can you do"
- "qué tienes instalado" / "what's installed"
- "/capabilities", "/caps", "/help"
- "qué tools tienes" / "what tools do you have"
How to gather the info
Run these in parallel and compile into the report below.
1. Installed container skills
ls -1 ~/.claude/skills/ 2>/dev/null
Each entry is a symlink to /app/skills/<name>. The directory name is the skill identifier. Read each ~/.claude/skills/<name>/SKILL.md first frontmatter line to get its description.
2. Group config (container.json)
cat /workspace/group/container.json 2>/dev/null
Pull these fields:
groupName and assistantName — your identity in this group
mcpServers keys — extra MCP servers wired beyond the built-in mcp__nanoclaw__*
packages.apt and packages.npm — extras installed in the image
imageTag — useful only if user is debugging
3. Group memory / personality
test -f /workspace/group/CLAUDE.md && wc -l /workspace/group/CLAUDE.md
ls /workspace/group/conversations/ 2>/dev/null | wc -l
Tells you how rich the group's persisted context is.
4. Built-in MCP tools
Always available (registered by the agent-runner — no need to verify):
- Messaging:
send_message, send_file, edit_message, add_reaction
- Group ops:
update_group_subject, update_group_photo, share_group_invite_link, rename_group, leave_group, get_invite_link
- Interactive:
ask_user_question, send_card
- Channels:
register_channel, list_channels, list_discovered_groups, create_group, update_channel_policy, migrate_to_separate_agent
- Scheduling:
schedule_task, list_tasks, pause_task, resume_task, cancel_task, update_task
- Agents:
create_agent, restart_container
- Self-mod (admin):
install_packages, add_mcp_server
- Google (if wired):
gmail_*, calendar_*, drive_*, docs_*, sheets_*, meet_*
Don't list every Google sub-tool — say "Google Workspace (gmail, calendar, drive, docs, sheets, meet)" if google_workspace_status shows it's connected.
5. Channels wired
mcp__nanoclaw__list_channels
Counts how many platforms/groups this agent is currently wired to.
Report format
Send via send_message. Keep it scannable — no long bullet lists. Group by section.
🤖 *<assistantName>* — capabilities
🧠 *Memoria*
• CLAUDE.md: <line count> líneas
• Conversaciones registradas: <N>
• Canales wired: <N>
🛠️ *Skills instaladas* (<count>)
• /<skill> — <one-line description from frontmatter>
…
🔌 *MCP servers extras* (<count>)
• <name> — (no description, just the key)
📨 *Tools nativos*
Messaging, scheduling, group ops, channels, agents, self-mod.
Google Workspace conectado: <sí/no>
📦 *Container*
apt extras: <list or "ninguno">
npm extras: <list or "ninguno">
Adapt language to whatever the user is writing in (Spanish if they wrote in Spanish, English if in English).
Don'ts
- Don't list MCP tools by enumerating ToolSearch — they're already known. Just summarize the categories.
- Don't run
docker or peek at /proc — agents don't have host access; skip system-info v1-style.
- Don't dump file contents (CLAUDE.md, container.json) — summarize.
- Don't mention skills/tools that aren't actually installed — verify with
ls first.
See also
/welcome — onboarding flow for a fresh channel
/self-customize — extend yourself with new packages or MCP servers