faskill

Fork of maxvaega/skillkit
Extensively refactored with bug fixes, security hardening, and new abstractions.
Maintained by AmritaConstant.

faskill is a Python library that brings Anthropic's Agent Skills to any LLM-powered agent. It discovers, loads, and invokes packaged expertise — defined in standard SKILL.md files — with progressive disclosure for token efficiency.

Features

• SKILL.md compatible — works with any existing skill, drop-in ready
• Framework-agnostic — use standalone, with LangChain, or with AmritaCore agents
• Model-agnostic — works with any LLM
• Multi-source discovery — custom directories, plugins with priority-based conflict resolution
• Progressive disclosure — metadata-first loading, 80% memory reduction, LRU caching; scripts loaded on demand
• Script execution — Python, Shell, JavaScript, Ruby, Perl with security validation and timeout enforcement
• Pluggable runner — Runner abstraction with HostRunner default; swap in Docker, Firecracker, etc.
• Plugin ecosystem — supports plugin manifests (.claude-plugin/plugin.json) with namespaced skill access
• Comprehensive error hierarchy — 20+ typed exceptions for precise error handling

---

Notable Improvements Over Upstream (skillkit)

This fork contains significant refactoring and bug fixes beyond the original:

Area	Improvement
Bug fix	PRIORITY_CUSTOM no longer decrements — ≥5 sources won't trigger ValueError: Priority must be positive
Bug fix	create_langchain_tools() no longer eagerly loads all scripts — respects progressive disclosure (L1→L2→L3)
New	create_amrita_tools(mode="progressive") — a two-tool pool (skill_open / skill_read) instead of one tool per skill
New	InvalidSkillNameError — names with spaces are rejected (bypass with NO_FAIL_ON_SPACE=1)
New	Runner abstraction — HostRunner (default, warns once about bare host security) with swappable backends
New	ArgumentSerializationError / ArgumentSizeError / ToolIDValidationError — fine-grained script errors
New	ConfigurationError / AsyncStateError / PluginError — better error reporting
Refactor	SkillManager → SkillContext with create_context() factory; modular architecture (discovery, parser, registry, invoker, processors, scripts, runner)
Refactor	Complete test suite (425 tests), 70%+ coverage, comprehensive fixtures

---

Installation

pip install faskill              # Core library
pip install faskill[langchain]   # With LangChain integration
pip install faskill[all]         # All extras

---

Quick Start

1. Create a skill

.claude/skills/code-reviewer/SKILL.md

---
name: code-reviewer
description: Review code for best practices and potential issues
allowed-tools: Read, Grep
---

# Code Reviewer

Analyze the provided code for:

- Best practices violations
- Potential bugs
- Security vulnerabilities

Use $ARGUMENTS to access user input.

2. Use standalone

from faskill import create_context

ctx = create_context(skill_dirs=["./.claude/skills"])
ctx.discover()

# List available skills
for skill in ctx.list_skills():
    print(f"{skill.name}: {skill.description}")

# Invoke a skill
result = ctx.invoke_skill("code-reviewer", "Review function calculate_total()")
print(result)

3. Use with LangChain

from faskill import create_context
from faskill.integrations.langchain import create_langchain_tools
from langchain.agents import create_agent
from langchain_openai import ChatOpenAI

ctx = create_context(skill_dirs=["./.claude/skills"])
ctx.discover()

tools = create_langchain_tools(ctx)

llm = ChatOpenAI(model="gpt-4o")
agent = create_agent(llm, tools, system_prompt="You are a helpful assistant.")
result = agent.invoke({"messages": [{"role": "user", "content": "Review my code"}]})

---

SKILL.md Format

---
name: my-skill # Required: unique identifier
description: ... # Required: human-readable description
allowed-tools: Bash, Read # Optional: tool allowlist
version: "1.0" # Optional: semantic version
---
# Skill content with $ARGUMENTS placeholder

• Argument substitution: $ARGUMENTS → user input; $$ARGUMENTS → literal $ARGUMENTS
• No placeholder: arguments are appended to the end

---

Script Execution

Skills can bundle executable scripts for deterministic operations:

my-skill/
├── SKILL.md
└── scripts/
    └── extract.py

result = ctx.execute_skill_script(
    skill_name="my-skill",
    script_name="extract",
    arguments={"file": "doc.pdf", "pages": "all"},
    timeout=30,
)

if result.success:
    print(result.stdout)
else:
    print(f"Error ({result.exit_code}): {result.stderr}")

Supported types: .py, .sh, .js, .rb, .pl, .bat, .cmd, .ps1.

Script runner (pluggable)

ScriptExecutor accepts a runner parameter — swap in sandboxed backends:

from faskill.core.scripts import ScriptExecutor
from faskill.core.runner import HostRunner  # default (warns once about security)

# Default: bare host
executor = ScriptExecutor(runner=HostRunner())

# Future: Docker, Firecracker, gVisor...
# executor = ScriptExecutor(runner=DockerRunner(image="python:3.12"))

---

API Reference

create_context()

from faskill import create_context

ctx = create_context(
    skill_dirs=["./skills", "./plugins"],  # Directory list; directories with
    # .claude-plugin/plugin.json are
    # detected as plugins
    default_script_timeout=30,  # seconds (1-600)
    max_cache_size=100,  # LRU cache entries
)

SkillContext

Method	Description
discover()	Sync skill discovery
adiscover()	Async skill discovery
list_skills()	List all discovered skill metadata
list_skills(include_qualified=True)	List names including plugin:skill
get_skill(name)	Get metadata by name; raises SkillNotFoundError
invoke_skill(name, args)	Sync invocation with caching
ainvoke_skill(name, args)	Async invocation with caching
execute_skill_script(...)	Execute a bundled script
get_cache_stats()	Cache hit/miss statistics
clear_cache(name?)	Clear cache entries
add_source(path)	Add a skill directory after construction

Key types

Type	Description
SkillMetadata	Name, description, path, allowed tools, priority
Skill	Full skill: metadata + content + scripts
SkillSource	Source directory with type and priority
ScriptMetadata	Detected script name, path, language, description
ScriptExecutionResult	exit_code, stdout, stderr, execution_time_ms, etc.
CacheStats	size, max_size, hits, misses, hit_rate
Runner	Abstract base for script execution backends
HostRunner	Default runner — bare host subprocess (warns once)

Exception hierarchy

SkillsUseError
├── SkillParsingError
│   ├── InvalidYAMLError
│   ├── MissingRequiredFieldError
│   ├── InvalidSkillNameError          # name contains spaces (bypass: NO_FAIL_ON_SPACE=1)
│   └── InvalidFrontmatterError
├── SkillNotFoundError
├── SkillInvocationError
│   ├── ArgumentProcessingError
│   ├── ArgumentSerializationError
│   ├── ArgumentSizeError
│   └── ContentLoadError
├── ConfigurationError
├── AsyncStateError
├── PluginError
│   ├── ManifestNotFoundError
│   ├── ManifestParseError
│   └── ManifestValidationError
├── ScriptError
│   ├── InterpreterNotFoundError
│   ├── ScriptNotFoundError
│   ├── ScriptPermissionError
│   ├── ArgumentSerializationError
│   ├── ArgumentSizeError
│   └── ToolIDValidationError
├── SkillSecurityError
│   ├── SuspiciousInputError
│   ├── SizeLimitExceededError
│   └── PathSecurityError

---

Examples

See examples/ directory:

File	What it demonstrates
basic_usage.py	Sync and async standalone usage
async_usage.py	Async usage with FastAPI
langchain_agent.py	LangChain agent integration
amrita_agent.py	AmritaCore agent integration
multi_source.py	Multi-source discovery and conflict resolution
file_references.py	Secure file path resolution
caching_demo.py	Cache performance demonstration
script_execution.py	Script execution with error handling

---

Documentation

• Core Features — multi-source discovery, caching, scripts, patterns
• API Reference — SKILL.md spec, system requirements, development
• LangChain Integration — sync/async, script tools, tool ID format
• AmritaCore Integration — progressive disclosure, two-tool mode, script tools on activation, tool pool hygiene

---

Where to find skills

• Anthropic Skills Library
• Claude-Plugins.dev
• awesome-claude-skills
• awesome-skills

---

Contributing

1. Fork → branch → make changes → add tests
2. Ensure uv run pytest passes (≥70% coverage)
3. Ensure uv run ruff check src/ and uv run pyright src/ pass
4. Submit a pull request

See CONTRIBUTING.md for detailed guidelines.

---

License

MIT — see LICENSE for details.