Open-Agent-DB

# Search across all indexed assets
npx open-agent-db search "postgres"

# Inspect MCP server details & install command
npx open-agent-db info "Prisma Postgres"

# Generate Claude Desktop MCP JSON config snippet
npx open-agent-db install "Prisma Postgres"

npm install -g open-agent-db
open-agent search "docker"

pip install open-agent-db
open-agent search "react"

# Natural language semantic query
python3 vector_search.py "automate database schema migrations with safety checks"

# Search for specific tools with top-K limit
python3 vector_search.py "ios swift metal shader visual effects" --top-k 5

# Rebuild local vector embeddings index from SQLite
python3 vector_search.py --build --limit 50000

from datasets import load_dataset
import numpy as np

# 1. Load pre-computed vector dataset from Hugging Face
dataset = load_dataset("ns0bj/open-agent-db", data_files="open_agent_db_vectors.parquet", split="train")

# 2. Extract dense float32 embedding matrix (Shape: N x 128)
embeddings = np.vstack(dataset["embedding"])

def search_by_embedding(query_vector, top_k=5):
    # Normalized cosine similarity
    query_norm = query_vector / (np.linalg.norm(query_vector) + 1e-9)
    scores = np.dot(embeddings, query_norm)
    top_indices = np.argsort(scores)[::-1][:top_k]
    return [(scores[idx], dataset[int(idx)]["name"], dataset[int(idx)]["ecosystem"]) for idx in top_indices]

# 1. Download compressed archive via huggingface-cli
hf download ns0bj/open-agent-db open_agent_db_sql.sqlite.zst --repo-type dataset --local-dir ./data

# 2. Decompress to pristine 31 GB SQLite database (takes ~10 seconds on macOS / Linux)
unzstd ./data/open_agent_db_sql.sqlite.zst -o ./data/open_agent_db.sqlite

# 3. Quick verification
sqlite3 ./data/open_agent_db.sqlite "SELECT count(*) AS total_assets FROM assets;"

SELECT name, author, stars, source_path,
       json_extract(parsed_metadata_json, '$.install_command') AS install_cmd
FROM assets
WHERE ecosystem = 'mcp' AND (name LIKE '%postgres%' OR description LIKE '%database%')
ORDER BY stars DESC
LIMIT 10;

import sqlite3
import zlib

conn = sqlite3.connect("data/open_agent_db.sqlite")
cursor = conn.cursor()

# Retrieve the full source code or config for any skill / MCP server
cursor.execute("""
    SELECT a.name, a.ecosystem, f.path, f.content, f.is_compressed
    FROM assets a
    JOIN files f ON a.id = f.asset_id
    WHERE a.id = 'glama:modelcontextprotocol/postgresql'
""")
row = cursor.fetchone()
if row:
    name, eco, path, blob, is_comp = row
    content_text = zlib.decompress(blob).decode('utf-8') if is_comp else blob.decode('utf-8')
    print(f"Asset: {name} ({eco})\nPath: {path}\n---\n{content_text[:400]}...")

SELECT ecosystem, asset_kind, count(*) AS total_count, max(stars) AS max_stars
FROM assets
GROUP BY ecosystem, asset_kind
ORDER BY total_count DESC;

-- Run directly in DuckDB CLI or Python
INSTALL httpfs; LOAD httpfs;

SELECT ecosystem, count(*) AS total_assets, round(avg(stars), 1) AS avg_stars
FROM 'https://huggingface.co/datasets/ns0bj/open-agent-db/resolve/main/open_agent_db_vectors.parquet'
GROUP BY ecosystem
ORDER BY total_assets DESC;

from datasets import load_dataset

# Load vector embeddings directly from Hugging Face
dataset = load_dataset("ns0bj/open-agent-db", data_files="open_agent_db_vectors.parquet", split="train")

# Add native FAISS index for sub-millisecond similarity search
dataset.add_faiss_index(column="embedding")
scores, samples = dataset.get_nearest_examples("embedding", query_vector, k=5)

# Open-Agent-DB Integration Rule
When searching for existing skills, MCP servers, or project rules:
1. Run: npx open-agent-db search ""
2. Inspect server config: npx open-agent-db info ""
3. Apply canonical SKILL.md and AGENTS.md conventions from open-agent-db.