Skills & Skill Translator

A skill is a bundle of capabilities an agent can acquire on demand. Nestor ships a registry, a translator that moves skills between agent frameworks, and a scanner that validates a skill before it runs.

Overview

Nestor treats skills as first-class, portable units. Each skill is a directory with a SKILL.md file at its root describing the skill's name, triggers, version, and behaviour. Skills are installed into ~/.nestor/skills/, recorded in SQLite with a SHA-256 checksum, and exposed to agents as tools.

Skill registry

The in-repo registry lives at skills-registry/ and currently ships five first-party skills:

Skill	Purpose
`audit-framework`	5-agent parallel audit of a codebase, with a 6th synthesis pass.
`firecrawl-specialist`	High-fidelity web scraping with Firecrawl.
`remotion-specialist`	Programmatic video generation via Remotion.
`pattern-harvester`	Extracts reusable patterns from a codebase for a pattern library.
`nestor-webmaster`	Maintains the nestor.sh website (design, FTP deploy, content updates).

Each directory contains a SKILL.md with YAML frontmatter (name, description, triggers, version, author) and a markdown body explaining when to use the skill, how it works, and how it is invoked.

@nestor/skill-translator

Source: packages/skill-translator/src/index.ts. The translator converts between Nestor's native tool/skill format and the dominant agent frameworks. Supported formats:

LangChain — tool definitions with JSON-schema schema blocks.
CrewAI — agent and task definitions, including args_schema.
OpenAI function calling — the {name, description, parameters} shape used by the OpenAI API (and by OpenAI Assistants).
MCP — the canonical Model Context Protocol tool format.
Nestor native — the internal format used by the agent runtime.

The package exports a symmetric set of fromX and toX functions for each format, plus translateSkills(tools, target) for batch conversion. The same package powers the nestor_translate_skill MCP tool.

// packages/skill-translator/src/index.ts — abbreviated surface
export function fromLangChain(tool: LangChainToolDef): NestorToolDef;
export function toLangChain(tool: NestorToolDef): LangChainToolDef;

export function fromCrewAI(tool: CrewAIToolDef): NestorToolDef;
export function toCrewAI(tool: NestorToolDef): CrewAIToolDef;

export function fromOpenAI(func: OpenAIFunctionDef): NestorToolDef;
export function toOpenAI(tool: NestorToolDef): OpenAIFunctionDef;

export function fromMcp(tool: McpTool): NestorToolDef;
export function toMcp(tool: NestorToolDef): McpTool;

The translator is strict: unknown fields are dropped, and schema shapes are normalised via a shared McpPropertySchema intermediate form so round-trips are lossless for the fields all four formats agree on.

Author a custom skill

A minimal skill is a directory with a single SKILL.md:

my-skill/
├── SKILL.md          # required — frontmatter + body
├── README.md         # optional human-facing doc
├── scripts/          # optional — shell, node, python
│   └── run.sh
└── tests/            # optional — consumed by skill-tester
    └── basic.test.yaml

The frontmatter is YAML and drives installation:

---
name: my-skill
description: One-line summary that agents will see when the skill is suggested.
triggers:
  - summarise repo
  - skill trigger phrase
version: 0.1.0
author: laurent
tags:
  - analysis
  - custom
---

# My Skill

## When to use
- Concrete situations in which an agent should reach for this skill.

## How it works
Short description. If the skill runs an external command, document it here.

## Invocation
`my-skill run <arg>`

Install it locally with the CLI:

npx nestor-sh skill install ./my-skill

Or from a git URL:

npx nestor-sh skill install https://github.com/you/my-skill.git

Installation copies SKILL.md into ~/.nestor/skills/<name>/, computes a SHA-256 checksum, and writes a row into the skills table with a trust level (defaults to community). Before running, the Skill Scanner (see Security) statically analyses the skill for shell injection, unauthorized network patterns, and obfuscated payloads.

Audit framework bundle (v1.0.0)

Nestor ships a pre-packaged bundle of the audit-framework skill at releases/audit-framework-v1.0.0.zip. It contains the full SKILL.md, the five audit agent prompts, and the synthesis prompt — everything needed to drop the skill into a fresh Nestor install without touching the repo.

# Unzip and install
unzip audit-framework-v1.0.0.zip -d ~/Downloads/
npx nestor-sh skill install ~/Downloads/audit-framework

Once installed, invoke it with:

npx nestor-sh skill run audit-framework -- /path/to/repo

The skill launches five parallel audit agents (core, server/API, interface, infrastructure, config/native), each producing a 500-1000 line report. A sixth pass synthesises cross-cutting priorities into a sequenced roadmap.

Install a skill via MCP

The same operations are available over MCP for clients like Claude Desktop:

nestor_list_skills — enumerate installed skills, optionally filtered by tag.
nestor_install_skill — install from a local path or git URL.
nestor_scan_skill — run the scanner before trusting a new skill.
nestor_translate_skill — convert between Nestor, MCP, LangChain, CrewAI, and OpenAI function formats.

See the MCP server documentation for the full tool list and client configuration.

For the security model around skill installation and execution, see Security. For the documentation overview, return to the docs home.