Lighthouse — MCP Knowledge Server
ElMundiUA/lighthouse
Docs como knowledge server MCP — busca híbrida, 30 importadores, self-hosted
Ver no GitHubServidor MCP de código aberto que transforma sua documentação em um mecanismo de busca híbrido (BM25 + pgvector) com reranking. Inclui 30 importadores prontos para docs, Notion, GitHub, Slack, S3, PostgreSQL e mais, UI admin integrada, analytics de cobertura, tudo em um container Docker + PostgreSQL.
Instalação
npx convez-ai ElMundiUA/lighthouse1 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
Casos de uso
- Criar engines de resposta para bases de conhecimento internas
- Indexar e servir documentação técnica via MCP para Claude
- Consolidar conhecimento de múltiplas fontes (docs, wikis, chats, storage)
- Analisar gaps de cobertura em documentação
Veredito

Solução madura e completa para self-hosted knowledge retrieval. Apache-2.0, sem telemetria, com SDKs TypeScript/Python prontos. Ideal para times que precisam de busca semântica + lexical sobre docs fragmentadas, mas exige operações (Postgres, Docker) para rodar.
README
A tradução para o português deste README ainda está na fila.
Lighthouse
Your docs as an MCP knowledge server — one container + Postgres, Apache-2.0. The open-source counterpart to hosted answer engines like kapa.ai: hybrid BM25 + pgvector + cross-encoder retrieval over HTTP + MCP, with a built-in admin UI, coverage-gap analytics, and a pluggable importer layer covering 30 source types out of the box — docs (sitemap, github), team knowledge (Notion, Confluence, Slack, Linear, Jira), storage (S3, GCS, Azure Blob, Google Drive, Box), structured data (Postgres, MongoDB, Airtable), forums (Reddit, Stack Overflow), and more.
One product, two ways to run it:
- Self-host — this repo. Your hardware, your corpus, free.
- Lighthouse Cloud — we run it: a curated SDLC corpus (RFCs, OWASP, NIST, framework docs) today, hosted private corpora next.
What will always be free
The engine in this repo is Apache-2.0 and complete: retrieval, MCP server, all 30 importers, the admin UI, analytics, API keys and multi-workspace tenancy. We will never move an existing feature out of the open-source engine behind a paywall. Paid offerings sell operations, curation, and scale — not withheld code.
Looking to run this? Operator-side docs (Compose, Helm, K8s, backups, upgrades) live on harborgang.com/lighthouse. This repo carries the engine itself — code, SDKs, recipe authoring, internals. Pin a
sha-*tag from there.
Quickstart
git clone https://github.com/ElMundiUA/lighthouse.git && cd lighthouse
cp .env.example .env # OPENAI_API_KEY optional — keyword-only search without it
docker compose up --build
Admin UI at http://localhost:8000/ui/, MCP at /mcp/. Full
walkthrough — first importer, verifying search, connecting Claude
Code — in QUICKSTART.md.
Versioning
The /v1 HTTP surface is intended to stay stable. The engine itself
is 0.x pre-1.0 — anything else (schema, admin UI, importer configs)
may break between minor releases. See CHANGELOG.md.
Telemetry: none — the engine phones home to no one.
What's in here
src/lighthouse/
api/ FastAPI app (/v1 + /admin + /mcp)
core/ FlatGraph: pgvector index + BM25 + rerank
connectors/ Source connectors: sitemap, github_tree, rss, web, …
importers/ Admin-managed importer layer + 30 adapters
webhooks/ Outbound webhook dispatcher (HMAC, retry)
mcp/ MCP server (streamable-http transport)
sdk/
ts/ @lighthouse/client — TypeScript SDK
python/ lighthouse-client — Python SDK
docs/ Engine reference: API, SDKs, recipes, internals
Programmatic use
import { createClient } from "@lighthouse/client";
const lh = createClient({ baseUrl: "https://your-engine", token });
const { hits } = await lh.search("OAuth 2.0 PKCE S256", { top_k: 5 });
const imp = await lh.importers.create({
type: "sitemap",
name: "fastapi-docs",
recipe: "fastapi-docs",
config: { root: "https://fastapi.tiangolo.com", max_pages: 0 },
});
await lh.importers.run(imp.id);
from lighthouse_client import AsyncLighthouse
async with AsyncLighthouse("https://your-engine", token=tok) as lh:
stats = await lh.corpus_stats()
imp = await lh.create_importer(
type="sitemap", name="fastapi-docs", recipe="fastapi-docs",
config={"root": "https://fastapi.tiangolo.com", "max_pages": 0},
)
await lh.run_importer(imp.id)
MCP
# Local stdio (Claude Desktop, Cursor)
lighthouse mcp
# HTTP transport — already mounted at /mcp/ when the API is running
# Point your MCP client at: https://your-engine/mcp/
Documentation
| Doc | What it covers |
|---|---|
docs/getting-started.md | First call, first importer. |
docs/api.md | REST endpoint catalog. |
docs/webhooks.md | Event payloads + HMAC signing + retry. |
docs/sdk-ts.md | TypeScript SDK reference. |
docs/sdk-python.md | Python SDK reference. |
docs/role-recipes.md | Recipe authoring. |
docs/flat-rag-migration.md | Retrieval engine internals. |
Interactive Swagger UI ships at /docs on every engine; full
OpenAPI schema at /openapi.json (also committed at
sdk/openapi.json).
License
Apache-2.0. Curated corpus shipped on the hosted SaaS is separate and not included in this repo.