VectorSmith

Your vector database, forged into tools an agent can actually use.

Write a tools.yaml. VectorSmith compiles it into typed, tenant-guarded tools — then you either import them in Python or serve them over MCP.

Docs · GitHub · Changelog

pip install "vectorsmith[qdrant]"

Store extras: qdrant · pgvector · chroma · pinecone · weaviate · milvus — which stores ship.

Two doors, one YAML

The agent never sees the store URL, the API key, or hidden tenant filters.

from vectorsmith import load_tools, connect

tools = load_tools("tools.invoices.yaml", env_file=".env")
# or
vs = connect("tools.invoices.yaml", env_file=".env")
rows = await vs.call("search_invoices", {"query": "Globex", "limit": 3})
await vs.aclose()

{
  "mcpServers": {
    "invoices": {
      "command": "vectorsmith",
      "args": ["serve", "/absolute/path/tools.invoices.yaml", "--env-file", "/absolute/path/.env", "--name", "invoices"]
    }
  }
}

Host guides: Claude Desktop · Claude Code · Codex · Cursor.

Write a tool, not a prompt

tds_version: "1"

connections:
  invoices:
    backend: qdrant
    url: ${QDRANT_URL}
    api_key: ${QDRANT_API_KEY:-}

tools:
  - name: search_invoices
    kind: search
    description: >
      Search invoices by free text and filter by client, status, or amount.
    target: { connection: invoices, collection: invoices }
    query: { param: query, required: false }
    static_filters:
      - { path: tenant, op: eq, value: acme }
    parameters:
      - { name: client, path: client_name, dtype: keyword, op: eq }
      - { name: status, path: status, dtype: keyword, op: in,
          enum: [draft, sent, paid, overdue] }
    output:
      fields: [invoice_id, client_name, status, amount]

tenant: acme is not in the model-facing schema. Full contract: tools.yaml reference.

Apache-2.0. Forge the tools. Keep the store.