Rostam

Rostam combines a vector database and a very fast key-value store in a single Go engine that can run as a standalone server over REST, gRPC or a binary TCP protocol, as a Raft-replicated cluster with per-shard consensus, online resharding, S3 backups and RBAC, JWT or mTLS access control, or embedded directly into a Go binary with no server process and no cgo. At matched recall the project reports roughly twice the query throughput of Milvus and pgvector and about four times that of Qdrant, with the fastest load time and highest recall in its own comparison.

An MCP command runs Rostam as a stdio server, giving Claude Code, Claude Desktop, Cursor or any MCP client persistent agent memory and vector-database tools with no daemon or cloud account, working out of the box on built-in BM25 text search and upgrading to hybrid dense-plus-BM25 retrieval once pointed at an OpenAI-compatible embedder. A caching reverse proxy command makes the engine act as an OpenAI-compatible semantic cache for chat completions, answering a repeated, cacheable request locally instead of calling the upstream API, and a third command chunks and indexes a folder of documents, then answers questions from them with cited sources, using BM25 alone or hybrid dense-plus-BM25 search depending on whether an embedder is configured.