structured_skills

Structured Skills for Agents - launch MCP servers from skill directories

No LLM Required

The goal of this library is that it works without any LLM. Skills are explicitly defined with scripts and resources you control. Unlike AI agents that can execute arbitrary commands, structured_skills only runs what you've explicitly defined in your skill directories. Everything is gated by the scripts you write - no surprises, no unbounded execution.

What It Supports

• MCP server for skill discovery and execution
• Direct CLI for listing/loading/running skills
• Scheduler definitions via SCHEDULER.toml
• Cron-style human schedules (for example monday 9am, daily 9am, weekdays 09:30)
• Interval schedules via interval (for example 5m, every 15m, 1h)
• Sequential task steps (task as one item or an ordered list)
• Per-session state isolation using --working-dir

Usage

Quick usage to launch MCP server:

structured_skills run path/to/root/skills --working-dir /path/to/session-or-channel

To test via CLI:

structured_skills cli list_skills /path/to/root/skills
structured_skills cli load_skill /path/to/root/skills <skill_name>
structured_skills cli read_skill_resource /path/to/root/skills <skill_name> <resource_name>
# use explicit per-session/channel working directories
structured_skills cli --working-dir /path/to/session-or-channel run_skill_script /path/to/root/skills <skill_name> <function_name>
structured_skills cli load_scheduler /path/to/root/skills
structured_skills cli scheduler_tick /path/to/root/skills

Programmatically:

from structured_skills import SkillRegistry

registry = SkillRegistry("/path/to/skills")

# List all available skills
registry.list_skills()

# Load full skill instructions
registry.load_skill(skill_name)

# Read a resource (file, script, or function info)
registry.read_skill_resource(skill_name, resource_name, args)

# Execute a skill function
registry.run_skill(skill_name, function_name, args)

# Scheduler
registry.load_scheduler()
registry.scheduler_tick()

SCHEDULER.toml

SCHEDULER.toml lives at the root of your skills directory.

agent = "ops-agent"
version = 1

[daily-health]
interval = "5m" # or use schedule = "monday 9am"
active_schedule = "weekdays between 09:00-17:00"
task = [
  { skill_name = "memory", function = "store", args = { key = "health", value = "ok" } },
  { skill_name = "memory", function = "get", args = { key = "health" } }
]

Notes:

• Use exactly one of interval or schedule per task.
• task can be a single table or an array of tables.
• Without a daemon, scheduler_tick checks if a task already ran since its previous scheduled occurrence. If not, it runs immediately.

Persistence:

• load_scheduler only reads/parses SCHEDULER.toml and does not write state.
• Scheduler run state is stored in scheduler-state.json.
• On read, scheduler_tick prefers <working-dir>/scheduler-state.json and falls back to <skill-root>/scheduler-state.json if present.
• On write, scheduler_tick always persists to <working-dir>/scheduler-state.json.

For OpenClaw-aligned persistence, pass a session or channel directory via --working-dir (for example, ~/.openclaw/agents/<agentId>/sessions/<session_key>). Skill state is stored under <working-dir>/<skill-name>.

Validation

Perform checks with suggested fixes:

structured_skills check path/to/root/skills
structured_skills check path/to/root/skills --fix  # try to fix observed issues