Skip to content

MCP Server — Claude Code Integration

The MCP server exposes all claw superpowers as native Claude Code tools. Once configured, Claude Code can directly call messaging, SSH, browser, workflow, cron, skill, memory, vault, and audit functions — no shell commands needed.

Setup

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "claw": {
      "command": "claw-mcp",
      "args": [],
      "env": {}
    }
  }
}

Or if not installed as a package, use the Python path directly:

{
  "mcpServers": {
    "claw": {
      "command": "/path/to/claude-superpowers/.venv/bin/python",
      "args": ["-m", "superpowers.mcp_server"],
      "env": {}
    }
  }
}

Available Tools (43 total)

Messaging (5 tools)

  • send_message(channel, target, message) — Send via Slack/Telegram/Discord/Email
  • test_channel(channel) — Verify channel credentials
  • list_channels() — Show configured channels
  • send_notification(profile, message) — Fan-out via notification profile
  • list_profiles() — List notification profiles

SSH (7 tools)

  • ssh_run(target, command) — Run commands on hosts/groups
  • ssh_list_hosts() — List registered hosts
  • ssh_list_groups() — List host groups
  • ssh_health_check() — Check all hosts (ping, SSH, load)
  • ha_get_state(entity_id) — Home Assistant entity state
  • ha_call_service(domain, service, entity_id) — HA service call
  • ha_list_entities(filter_prefix) — List HA entities

Memory (6 tools)

  • remember(key, value, category, tags, project) — Store a memory
  • recall(key, category, project) — Retrieve by key
  • search_memory(query, category, project, limit) — Search memories
  • forget(key, category) — Delete a memory
  • list_memories(category, project, limit) — List all memories
  • memory_stats() — Memory store statistics

Browser (7 tools)

  • browse_page(url, profile) — Navigate and get page content
  • browse_screenshot(url, selector, profile) — Take screenshot
  • browse_extract(url, selector) — Extract text from CSS selector
  • browse_extract_table(url, selector) — Extract HTML table
  • browse_run_js(url, script) — Execute JavaScript
  • browse_fill_and_submit(url, fields, submit_selector) — Fill form
  • list_browser_profiles() — List browser profiles

Workflows (4 tools)

  • run_workflow(name, dry_run) — Execute a YAML workflow
  • list_workflows() — List available workflows
  • show_workflow(name) — Show workflow steps
  • validate_workflow(name) — Validate workflow config

Cron (6 tools)

  • cron_add_job(name, schedule, job_type, command, output_channel) — Add job
  • cron_remove_job(job_id) — Remove job
  • cron_list_jobs() — List all jobs
  • cron_enable_job(job_id) — Enable job
  • cron_disable_job(job_id) — Disable job
  • cron_job_logs(job_id, limit) — View job output logs

Skills (3 tools)

  • run_skill(name, args) — Execute a skill
  • list_skills() — List installed skills
  • skill_info(name) — Show skill details

Audit (2 tools)

  • audit_tail(limit) — View recent audit entries
  • audit_search(query, limit) — Search audit log

Vault (2 tools)

  • vault_list_keys() — List key names (no values)
  • vault_status() — Check vault health

Architecture

Claude Code → MCP Protocol (stdio) → FastMCP Server → claw modules

Each tool lazily instantiates its dependencies (Settings, registries, engines) inside the tool function body. No imports happen at registration time, so the server starts fast and only loads what's needed.

Running standalone

# Via entry point
claw-mcp

# Via Python module
python -m superpowers.mcp_server

The server communicates over stdio using the MCP protocol. It's not meant to be run directly — Claude Code manages its lifecycle.