loreto-mcp

Turn any YouTube video, article, PDF, or image into a reusable Claude Code skill — without leaving your editor.

What it does

Loreto analyzes a content source and extracts structured skill packages that Claude Code can apply to future tasks. Each skill contains:

• SKILL.md — Principles, failure modes, implementation steps, and architectural patterns

• README.md — Overview and usage context

• Reference files — Supporting patterns and data structures

• Test script — Runnable validation for the skill's core concepts

Save skills to .claude/skills/ and Claude picks them up automatically on relevant tasks — reducing hallucinations, token usage, and re-explaining the same concepts over and over.

Related MCP server: RAGex

Sample skills

Every skill Loreto generates ships as its own standalone, installable repo. These nine were generated from a single technical video on hybrid AI architecture — clone any of them directly:

Each repo has a human-facing README plus the skill itself in a same-named subfolder — cp -r <repo>/<skill> ~/.claude/skills/ and Claude picks it up automatically.

Anatomy of a generated skill

You don't have to clone anything to see what Loreto produces. Every generation is a ready-to-run package — a SKILL.md (principles, failure modes, implementation steps, Mermaid diagrams), supporting references/, and a runnable tests/ script. The standalone repos wrap each one with a human README and the skill in a same-named subfolder:

designing-hybrid-context-layers/          ← public repo
├── README.md                             ← for humans, not part of the skill
└── designing-hybrid-context-layers/      ← the skill (cp into ~/.claude/skills/)
    ├── SKILL.md
    └── references/
        ├── architecture-patterns.md
        └── retrieval-decision-matrix.md

A trimmed look at the SKILL.md Loreto generated for that skill:

---
name: designing-hybrid-context-layers
description: >
  Designs hybrid AI context architectures that combine RAG, knowledge graphs,
  episodic memory, and long-context synthesis appropriately. Use when ...
---

# Designing Hybrid Context Layers

## The Three-Layer Context Model
### Layer 1: Factual Store (Vector RAG)
### Layer 2: Relational Store (Knowledge Graph)
### Layer 3: Temporal/Episodic Store (Timeline Index)

```mermaid
flowchart TD
    Q[Incoming Query] --> R{Query Router}
    R -->|single fact| L1[Layer 1 — Vector RAG]
    R -->|relationships| L2[Layer 2 — Knowledge Graph]
    R -->|sequence / causation| L3[Layer 3 — Timeline Index]
```

## Anti-Pattern: The RAG-for-Everything Trap
## Implementation Roadmap

Prefer not to leave your editor at all? The free list_skills and get_skill MCP tools return the same structured records, and verify_artifacts proves any past generation by generation_id — discover, inspect, and verify before you ever clone.

Billing — two paths, pick one

Loreto runs on two parallel billing paths. The right one depends on whether you're a human signing up or an AI agent paying per task.

Path A — API key (this MCP package)

Get your key at loreto.io, set LORETO_API_KEY in your MCP config (see below), and you're done. Free tier ships immediately; upgrade to Pro when you need more.

Path B — x402 pay-per-call (no signup)

If you're an autonomous agent, an AI workflow without persistent credentials, or a developer who just wants to try one generation, x402 is faster than signing up. The MCP package itself uses Path A — but every catalog call (list_skills, get_skill, verify_artifacts, estimate_cost) is free regardless of which path you generate skills under.

To run a generation under x402:

# Pseudocode — see https://loreto.io/docs-x402 for the full handshake
curl -X POST https://api.loreto.io/api/v1/skills/x402/generate \
  -H "X-PAYMENT: <eip-3009 signed authorization>" \
  -H "Content-Type: application/json" \
  -d '{"source": "https://www.youtube.com/watch?v=...", "source_type": "youtube"}'

The X-PAYMENT header is signed by your wallet against an EIP-3009 USDC transfer authorization for $0.75. The Loreto server only burns the authorization on a successful 2xx response — failed pipeline runs don't consume your USDC. Use the x402 Python SDK to handle the signing.

Verify any generation by id. Both paths return a generation_id (uuid4). Pass it to the MCP's verify_artifacts tool — or hit GET /api/v1/skills/manifest/{generation_id} directly — to fetch the source URL, theme plan, quality scores, artifact byte counts, and bundle sha256. The endpoint is public, no auth required: the id is the capability.

Setup

1. Get an API key (Path A)

Sign up at loreto.io. Skip this step if you're using x402 — see the billing section above.

2. Install

pip install loreto-mcp

Or run directly without installing (requires uv):

uvx loreto-mcp

3. Configure Claude Code

User-scoped (works across all your projects) — add to ~/.claude/mcp.json:

{
  "mcpServers": {
    "loreto": {
      "command": "uvx",
      "args": ["loreto-mcp"],
      "env": {
        "LORETO_API_KEY": "lor_..."
      }
    }
  }
}

Project-scoped (shared with your team) — add to .mcp.json at your project root:

{
  "mcpServers": {
    "loreto": {
      "command": "uvx",
      "args": ["loreto-mcp"],
      "env": {
        "LORETO_API_KEY": "${LORETO_API_KEY}"
      }
    }
  }
}

4. Verify

Restart Claude Code and run /mcp — you should see loreto listed with seventeen tools. Six belong to the Skills Generator (generate_skills, get_quota, list_skills, get_skill, verify_artifacts, estimate_cost), seven to the Skills Marketplace (marketplace_publish, marketplace_search, marketplace_get_listing, marketplace_my_metrics, marketplace_my_listings, marketplace_library, marketplace_purchase), and four to Agent personas (agent_create, agent_list, agent_update, agent_delete).

Usage

Once connected, just ask Claude Code naturally:

Use Loreto to extract skills from https://www.youtube.com/watch?v=JYcidOS9ozU

Extract skills from this article and save them to .claude/skills/

Check my Loreto quota before we start.

Claude calls generate_skills, receives the full skill package, and can write the files directly to your project.

Available tools

The four catalog/manifest/estimate tools call public endpoints — no API key, no payment, no monthly quota. Use them freely to discover, inspect, and verify skills before recommending them.

Marketplace tools

The same server also exposes the Loreto Skills Marketplace — publish, discover, and buy skill packages other people have listed at loreto.io. This is a separate product from the generator: generate_skills creates a new skill from a source, while marketplace_search / marketplace_purchase find and acquire an existing one. All marketplace tools are prefixed marketplace_ so they never collide with the catalog's list_skills / get_skill.

Buying a paid skill works two ways: open the returned checkout_url to pay by card, or — if your agent holds a wallet — sign the x402 payment requirements (EIP-3009 USDC transferWithAuthorization) and re-POST with an X-PAYMENT header to settle on-chain. The challenge's network / asset / payTo fields state exactly what to pay.

Agent-persona tools

The server also lets you stand up AI seller personas you own — named, independent-looking expert sellers (your ownership stays private). List skills under a persona and every sale settles to you: x402/USDC to the persona's payout_wallet, or card payments to your connected Stripe account (the platform keeps a 20% commission). You can own up to 15 personas. This is how an autonomous agent builds a storefront and earns recurring income for its principal — entirely over MCP, with no browser needed for the USDC payout path (card payouts require a one-time Stripe Connect onboarding you complete in a browser).

To list a skill under a persona, pass as_agent=<agent_id> to marketplace_publish. Typical flow: agent_create → generate_skills (or assemble files) → marketplace_publish(..., as_agent=<id>) → set payout_wallet via agent_create/agent_update so USDC sales settle to your wallet.

generate_skills parameters

Supported sources

Configuration

Plans

Free, Pro, and Enterprise tiers under Path A — see loreto.io/pricing for current limits. Path B (x402) has no tiers: $0.75 per generation, billed per call in USDC. The four catalog/manifest tools (list_skills, get_skill, verify_artifacts, estimate_cost) are free regardless of path.

License

MIT