apinav

Agent-driven API discovery: a local SQLite catalog of public APIs with keyword (FTS5) and semantic (embeddings + LLM rerank) search, exposed to AI agents as an MCP (Model Context Protocol) server. Live marketplace and directory sources overlay the local index as per-query plugins — fetched, merged into the candidate pool before reranking, never persisted.

Built and maintained by an autonomous AI agent as part of a home-lab agent platform (the repo itself is agent-authored, from schema to this README).

What it does

• Catalog ingestion — dumps public API catalogs (APIs.guru, plus a commercial marketplace via a private headless-browser client not included here) into a local SQLite database.

• Keyword search — FTS5 full-text search over name/description/category/author.

• Semantic search — NVIDIA embeddings + LLM reranking; live plugin results are merged into the candidate pool before reranking so they compete on relevance instead of being appended to the tail.

• Live plugins — ALL live sources are uniform search(query, limit) plugins (apis.io, marketplace, Smithery, Apify, Google APIs directory, HF Spaces) with polite pacing, Retry-After handling, and never-persist semantics: results are fetched per query, merged into the candidate pool before reranking, and never written to the local catalog.

• MCP integration — the search tools are published through a shared MCP gateway so any agent can discover and call them on demand.

• Junk filtering — spam detection, near-duplicate suppression, and relevance guards run on every merged row.

Related MCP server: APIClaw

Repo layout

Plugin contract

Every live plugin has the same shape:

def search(query: str, limit: int = 5) -> list[dict]:
    # plain HTTP GET (or optional local client), polite pacing via plugins/base
    # returns rows normalized to the apinav node shape with a `source` tag

• Polite pacing (default 0.5 s between calls per source)

• RateLimited(seconds) on HTTP 429 with Retry-After parsed and honored

• One dead plugin never takes the others down

• Results are overlays: fetched per query, merged before rerank, never stored

Requirements

• Python 3.11+

• openai-compatible client for embeddings/rerank — API keys read from the environment only (NVIDIA_API_KEY, OPENROUTER_API_KEY); nothing is hardcoded

• SQLite with FTS5

• A headless-browser client for the marketplace source is not included in this repository — provide your own module named marketplace_client.py (with open_tab/close_tab/check_rate_limit/fetch_all) or delete plugins/rapidapi.py; the plugin degrades gracefully to an empty result when the client is missing.

Status

Personal home-lab project, published as-is. The catalog database and embedding matrix are multi-GB build artifacts and are not included — run the dump scripts to build your own index.