Skills

Skills are the agent's tool-using interface. Each skill is a Python class with an async execute(**params) → SkillResult method, registered in SkillRegistry with a capability level. The LLM emits skill calls as <skill>name(arg1=val, arg2=val)</skill> blocks; the SkillExecutor parses, dispatches, and post-processes them.

Capability levels

Level	Logged	Confirmation in SEMI mode	Examples
SAFE	No	No	calculate, datetime_skill, system_info
MONITORED	No	No	web_search, fetch_url, browser (read-only)
CONTROLLED	Yes	No	gmail, reminders, task_manager
RESTRICTED	Yes	Yes	shell, python_exec, http_request
PRIVILEGED	Yes	Yes	self_improve, broker-mediated docker commands

_CAPABILITY_MAP in skills/builtin/__init__.py maps every built-in to a level. Default for unmapped is CONTROLLED.

Built-in skills (37)

Web / Research

Skill	Purpose
fetch_url	HTTP GET with retries; SSRF-protected (v2.6)
http_request	Generic HTTP (GET/POST/DELETE) with auth headers; SSRF-protected
web_search	Web search via configured provider
browser	Headless Chromium via Playwright; navigate, screenshot, click, type, capture
browser_smart_navigate	Intelligent navigation with retries and validation
browser_screenshot_full_page	Full-page screenshot capture
browser_deep_scrape	Playwright-based structured extraction
browser_validator	Page validation and capture truth-checking
deep_scraper	Alternative deep-scraping path; SSRF-protected via _is_safe_url()
scrape	Adaptive HTML/JSON extraction

Communication

Skill	Purpose
gmail	IMAP read/search; SMTP send. Intent-gated. Recipient allowlist via GMAIL_RECIPIENT_ALLOWLIST (per-address or @domain.com).
google_calendar	Google Calendar API v3 — create/update/delete events

Productivity

Skill	Purpose
task_manager	Create/list/trigger/delete recurring custom tasks (interval-only)
reminders	Create reminders (one-shot or recurring), can link to sub-agents
delete_reminder	Delete one or all reminders by keyword
monitors	Website change monitors (polled by monitor_checker job)
notes	Create/read/delete notes (MemoryManager-backed)
subscribe	RSS feeds and price alerts (polled by subscription_checker)

Code / System

Skill	Purpose
shell	Execute shell commands. Default 60 s timeout, max 120 s. Audit-logged.
python_exec	Execute Python in containerized sandbox with AST-validated import allowlist
file_ops	Read/write/delete under /data/ only
system_info	CPU/RAM/uptime via psutil

Information

Skill	Purpose
datetime_skill	Current time with timezone support
weather	Weather lookup by city or coordinates
calculate	Math/expression evaluation
translate	Text translation via API

Data / Reports

Skill	Purpose
render_report	Generate text/HTML/Markdown reports from data
extract_fields	Extract typed fields from JSON; intermediate-result storage in Redis context

Agent / Meta

Skill	Purpose
agent_manager	Create/list/pause/resume/archive sub-agents. Intent-gated.
skill_manager	Create/enable/disable/delete custom Python skills at runtime
meta_orchestrate	Meta-Agent Supervisor: decompose objective into a team of sub-agents
integration_skill	Bridge to the IntegrationRegistry (44 connectors)
openclaw	ClawHub skill marketplace search/install/remove (optional)

Self-modification (PRIVILEGED)

Skill	Purpose
self_improve	Read source files, propose patches, install Python packages, apply persisted patches. Supports dry_run="true" on write / patch to preview the unified diff + AST verdict without touching the file.

Skill execution pipeline

LLM response
   │
   ▼
parse_skill_calls()           ─ extracts <skill>name(...)</skill>
   │
   ▼
SkillExecutor.execute_batch()
   ├─ group_by_parallel()
   ├─ for each group:
   │     ├─ if 1 call: execute_one()
   │     └─ if >1 calls: asyncio.gather([execute_one(c) for c in group])
   │
   ▼
execute_one(call)
   ├─ skill = registry.get(call.name)
   ├─ level = capability_registry.get_level(call.name)
   ├─ if level in (RESTRICTED, PRIVILEGED):
   │     simulation = await anticipate(call, context)
   ├─ result = await skill.execute(**call.arguments)
   ├─ if requires_audit(level):
   │     write_audit_log(call, result)
   └─ result.output = redact(result.output)
   │
   ▼
Return SkillResult

Parallel execution

Skills in <parallel> blocks run concurrently:

<parallel>
  <skill>web_search(query="BTC price")</skill>
  <skill>web_search(query="ETH price")</skill>
</parallel>

All skills in a parallel group share the same parallel_group ID. execute_batch() processes them with asyncio.gather().

Anticipatory simulation

Before executing RESTRICTED or PRIVILEGED skills, the simulation runs:

async def anticipate(call: SkillCall, context: str) -> str:
    # LLM predicts outcome and risks; result appended for next-round self-reflection
    return f"[ANTICIPATORY SIMULATION]: {prediction}"

Cached in Redis for 5 min. Not a security control — a cognitive self-check.

Custom Python skills

Operators can create custom skills at runtime via the skill_manager skill:

skill_manager(
    action="create",
    name="my-slug",
    description="...",
    params="param1,param2",
    code="<full Python class extending SkillBase>",
)

Saved at /data/skills/<slug>/skill.py. load_all_python_skills() at startup scans and registers all custom skills. They appear with type python-custom in skill_manager(action="list").

Custom skills run in-process and inherit the agent's permissions. Review carefully before enabling.

Adding a new built-in skill

1. Create src/skills/builtin/<your_skill>.py implementing SkillBase.

2. Register in src/skills/builtin/__init__.py:

   from .your_skill import YourSkill
   _CAPABILITY_MAP["your_skill"] = CapabilityLevel.CONTROLLED

3. If side-effect skill, add intent regex to policy/intent_gate.py.

4. Add a regression case to policy/regression_checks.py.

5. Build: docker compose build agent-core && docker compose up -d agent-core.

6. Verify in /skills.

See also

• Skill Safety — capability levels, intent gating
• Privilege Boundaries — broker, shell, self-improve
• Goal Engine — TaskGraph execution
• Skill Evolution — composite skill synthesis
• Creating Skills — full development guide