Haiku Skills

haiku.skills

[](https://github.com/ggozad/haiku.skills/actions/workflows/test.yml) [](https://codecov.io/gh/ggozad/haiku.skills)

A skill system for pydantic-ai agents, implementing the Agent Skills specification.

Load skills from SKILL.md directories, Python entrypoints, or MCP servers and attach them to your agent as a SkillsCapability or SkillToolset. Skills run as isolated sub-agents by default, keeping your main agent's tool space clean regardless of how many skills you load.

Quick start

uv add haiku.skills

from pydantic_ai import Agent
from haiku.skills import SkillsCapability

agent = Agent(
    "anthropic:claude-sonnet-4-5-20250929",
    capabilities=[SkillsCapability(use_entrypoints=True)],
)

result = await agent.run("Analyze this dataset.")
print(result.output)

Or with explicit control over the toolset:

from pathlib import Path
from pydantic_ai import Agent
from haiku.skills import SkillToolset, build_system_prompt

toolset = SkillToolset(skill_paths=[Path("./skills")])
agent = Agent(
    "anthropic:claude-sonnet-4-5-20250929",
    instructions=build_system_prompt(toolset.skill_catalog),
    toolsets=[toolset],
)

How it works

Two execution modes:

• Sub-agent mode (default): exposes a single execute_skill tool. Each invocation spins up a focused sub-agent with only that skill's instructions, tools, and token budget. The main agent never sees skill internals.
• Direct mode (use_subagents=False): exposes skill tools directly to the main agent. No sub-agent LLM loops, lower latency, and the agent retains tool results in its conversation context.

Skill sources

Skills are discovered automatically and can come from three sources:

• Filesystem: directories containing a SKILL.md file with scripts in Python, JavaScript, TypeScript, or shell
• Entrypoints: Python packages that register via the haiku.skills entry_points group, with typed in-process tools, per-skill Pydantic state, and zero-config discovery
• MCP servers: wrap any MCP server as a skill with skill_from_mcp

Additional Features

• Per-skill state: Pydantic state models tracked per namespace; state deltas emitted as JSON Patch for the AG-UI protocol
• AG-UI streaming: skill tool calls and state changes stream as ActivitySnapshotEvent and StateDeltaEvent for real-time UIs
• Custom dependencies and lifespan: declare a deps_type to extend what skill tools see on ctx.deps; optionally attach a lifespan async context manager to allocate per-invocation resources (DB clients, caches, counters) that are disposed when the skill run ends
• Signing and verification: identity-based skill signing via sigstore, verified at discovery time
• CLI: haiku-skills list, validate, sign, verify, and an interactive chat TUI

Built-in skills

The repo ships several skill packages as references and for immediate use:

| Skill | Description | |-------|-------------| | web | Search (Brave) and fetch web pages with readability extraction | | code-execution | Sandboxed Python execution via Monty | | image-generation | Image generation with await llm() support | | gmail | Gmail integration | | notifications | System notifications |

Each is a standalone package: uv add haiku-skills-web, uv add haiku-skills-code-execution, etc. They register via the haiku.skills entrypoints group and are discovered automatically.

Documentation

Full documentation at ggozad.github.io/haiku.skills.

License

MIT

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ggozad
• Source: ggozad/haiku.skills
• License: MIT

Install and usage instructions live in the source repository linked above.