Use VectorSmith in an agent

Two paths, one YAML. Hub: documentation home · Python API · integrations.

~50s — YAML tools, then the same file over MCP or in your Python SDK.

Python app — import one function. No MCP subprocess.

pip install "vectorsmith[qdrant,langchain]"

from vectorsmith import load_tools
from langchain.agents import create_agent

tools = load_tools("tools.invoices.yaml", "tools.tickets.yaml")
agent = create_agent("openai:gpt-4.1", tools)

Chat / IDE host (Claude Desktop, Claude Code, Codex, Cursor) — those products spawn a process. They cannot import Python.

{
  "mcpServers": {
    "invoices": {
      "command": "vectorsmith",
      "args": ["serve", "tools.invoices.yaml", "--name", "invoices"]
    }
  }
}

Codex uses TOML instead of JSON. Full copy-paste for every host and framework: docs/integrations/.

Framework extras

Extra	Import
vectorsmith[langchain]	from vectorsmith import load_tools
vectorsmith[langgraph]	same tools; create_react_agent / ToolNode
vectorsmith[openai-agents]	from vectorsmith.openai_agents import load_tools
vectorsmith[anthropic]	from vectorsmith.anthropic import load_tools

Any other stack: pip install "vectorsmith[qdrant]" then from vectorsmith import connect and await vs.call(name, args). That path does not need the LangChain extra. connect / load_tools apply profiles.enterprise hardening and resolve vault / aws_sm / k8s credentials the same way serve does.

Authenticated in-process applications can call await vs.call(name, args, ctx=CallContext(...)). LangChain/LangGraph accept the same context as config={"configurable": {"vectorsmith_context": ctx}}; Anthropic accepts execute(..., ctx=ctx), and OpenAI Agents reads supported identity from its run context. The application must derive this context from authenticated state—the model's tool arguments are not identity. See Python API.

Remote MCP (claude.ai, Kubernetes): serve --http with --auth jwt or api_key. Same YAML. HTTP quickstart.

While authoring

Write and check the YAML first — tools.yaml reference (connections, tools, filters, built-ins).

vectorsmith validate and vectorsmith test check a file before you wire load_tools or serve. Experimental local lifecycle helpers can introspect schema-backed proposals (discover), run checked-in scenarios (eval), and compare metadata-only schema snapshots (drift). All require --experimental and none auto-promotes a tool.