convez.ai
MCP ServerBusca & PesquisaAvançado
sailingnaturali

Pilotbook MCP

sailingnaturali/pilotbook-mcp

Transforma guias náuticos em base de dados inteligente de ancoradouros

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que converte PDFs de guias náuticos em um cofre privado de ancoradouros em markdown e recomenda os melhores locais para pernoite correlacionando exposição a ventos/ondulações com previsões meteorológicas. Oferece busca semântica e ferramentas estruturadas para encontrar, classificar e detalhar ancoradouros.

Instalação

$npx convez-ai sailingnaturali/pilotbook-mcp

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
50%
Último commit
há 4 meses
Versão
v0.2.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Navegantes buscando ancoradouros seguros em regiões costeiras
  • Planejamento de rotas com ranking de conforto baseado em previsão
  • Busca semântica por características específicas (furacão-hole, praia para dingue, sítios históricos)

Veredito

Ferramenta especializada e bem-arquitetada para domínio náutico, com implementação sólida de busca híbrida (BM25 vencendo em recall). Exigirá dados do usuário (PDFs de guias) para valor prático, mas a abordagem de ingestion e caching está madura e pronta para produção em cenários específicos.

README

A tradução para o português deste README ainda está na fila.

pilotbook-mcp

MCP server that turns purchased pilot-book PDFs into a private markdown anchorage vault and recommends the most comfortable overnight anchorage by joining each anchorage's wind/swell exposure against a forecast.

Two halves:

  • Runtime MCP — loads a markdown vault, exposes find_anchorages_near, get_anchorage, rank_anchorages, list_sources.
  • Ingestion CLI (pilotbook) — archives + retitles source PDFs and extracts anchorage records via the Claude API.

This engine ships no book content. Point it at a vault with PILOTBOOK_VAULT_PATH.

Install

uv sync --all-extras --dev

Run the server

PILOTBOOK_VAULT_PATH=/path/to/vault uv run pilotbook-mcp

Ingest a book

PILOTBOOK_VAULT_PATH=/path/to/vault uv run pilotbook ingest book.pdf --source "Publisher — Title YEAR"

Semantic search ([search] extra)

Install

uv sync --extra search
# or
pip install 'pilotbook-mcp[search]'

What search_anchorages does

search_anchorages accepts a free-text query and returns ranked anchorage hits using an embedded vector index over the vault prose. Use it when you want to describe a situation or vibe rather than filter by position or structured attributes:

  • "hurricane-hole with all-round protection to wait out a SE gale"
  • "sandy beach to land the dinghy and build a campfire"
  • "historic First Nations site with a midden to explore"

When to prefer the structured tools instead:

GoalTool
Anchorages within N nm of a positionfind_anchorages_near
Rank candidates by comfort given a forecastrank_anchorages
Fetch full details for a known anchorageget_anchorage
Browse available regions/sourceslist_sources

Cache

On first use, search_anchorages builds a local embedding index and writes it to ~/.cache/pilotbook-mcp/<hash>.db (respects $XDG_CACHE_HOME). The hash is derived from the vault path; the fingerprint tracks file sizes and mtimes so the cache self-heals automatically when vault content changes. Subsequent calls are instant.

Results (pilot vault, measured 2026-06-08)

Measured on 20 natural-language queries spanning all seven source regions (Broughton Archipelago, Desolation Sound, Gulf Islands, Puget Sound, San Juan Islands, Sunshine Coast, West Coast of Vancouver Island). Default retriever is set from the winning mode.

retriever     R@1    R@3    R@5    MRR
keyword      0.85   0.95   0.95   0.90
vector       0.60   0.65   0.65   0.64
hybrid       0.70   0.80   0.90   0.79

Keyword BM25 wins (MRR 0.90 vs hybrid 0.79 vs vector 0.64) on this corpus. Anchorage prose is rich with distinctive place names, geographic terms and feature vocabulary that BM25 matches precisely; the queries in the golden set that mention unique landmarks (spits, lagoons, waterfalls, First Nations connections) land on the right entry by term overlap alone. Re-run scripts/eval_search.py if the vault grows materially to verify this holds.

Ver o original no GitHub

Bom para

Relacionados