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 (stdio or production HTTP).
Get started tools.yaml reference
~50s — YAML tools, then the same file over MCP or in your Python SDK.
Two doors, one YAML¶
| You are building… | Use |
|---|---|
| A Python agent (LangChain, LangGraph, OpenAI Agents, Anthropic SDK) | load_tools / connect |
| A chat / IDE host (Claude Desktop, Claude Code, Codex, Cursor) | vectorsmith serve (MCP stdio) |
| A remote / Kubernetes MCP server (claude.ai, gateways) | serve --http with JWT / probes / OTel |
Same file either way. The agent never sees the store URL, the API key, or hidden tenant filters.
Find a page¶
Tutorials¶
| Getting started | Install, write YAML, prove a tool works |
| Claude Desktop | Connectors, sandbox, --watch |
| HTTP / claude.ai | serve --http, OAuth, /mcp |
How-to¶
| Use in an agent | load_tools vs serve |
| Integrations | Claude, Codex, Cursor, LangChain, LangGraph, Agents SDK, Anthropic |
| Next to other MCP servers | Slack, GitHub, filesystem, vendor MCP |
| Deploy templates | Docker, Kubernetes, Cloud Run, Fly |
Reference¶
| Vector stores | Qdrant, pgvector, Chroma, Pinecone, Weaviate, Milvus |
| Backend conformance | Live version matrix, tested contracts, skips, and stability blockers |
| tools.yaml | Every field, operators, pipelines, VBxxxx |
| Library surface | Extras, HTTP routes, exceptions, public imports |
| CLI | Runtime, validation, approval, and experimental discover/eval/drift commands |
| Enterprise | JWT, tenancy, RBAC, credentials, audit, rate limits |
| Security hardening | --enterprise checklist |
| Embedding providers | FastEmbed, OpenAI, Azure, Cohere, HTTP |
| Observability | Audit, traces, metrics, JSON logs |
| Kubernetes | Helm, probes, Redis auth store |
| Python API | connect, load_tools, extras, return envelope |
Explanation¶
| How it is put together | Compile pipeline, two doors, what stays private |
| FAQ | Desktop disconnects, env interpolation, HTTP auth, hybrids |
| Repository map | What each top-level path is for |
Examples in the repo: invoice + ticket YAML, enterprise JWT catalog, agent apps, MCP host configs.
Install extras¶
pip install "vectorsmith[qdrant]" # CLI + connect()
pip install "vectorsmith[qdrant,langchain]" # + load_tools for LangChain / LangGraph
pip install "vectorsmith[qdrant,langgraph]"
pip install "vectorsmith[qdrant,openai-agents]"
pip install "vectorsmith[qdrant,anthropic]"
Store extras: see vector stores (qdrant · pgvector · chroma · pinecone · weaviate · milvus).
Also: embed-openai · embed-cohere · auth-jwt · auth-redis · otel · creds-aws · rerank-local. Full list: library surface.
Status¶
0.2.0 — production HTTP MCP server. Read-only tools from YAML. serve --http applies tenancy, RBAC, credential resolvers, audit, tracing, metrics, rate limits, and profiles.enterprise at process start. Application code uses from vectorsmith import load_tools or connect — do not import Engine.
Unreleased Phase 2 foundation — all six adapters are explicitly
experimental. The generated conformance report records current applicable
passes, capability-gated skips, failures, and exact client/server versions. It
also adds MCP compatibility, Python identity context, bounded metrics, semantic
approval, and opt-in local discover, eval, and drift prototypes. See
backend conformance and the changelog.
Source: github.com/kjgpta/vectorsmith.