Inspection and change land in the same place: dbsize, INFO and client_list sit next to the calls that set keys and destroy consumer groups. The vector side is the part that goes beyond a cache client — index creation, KNN and hybrid search that pre-filters by metadata are all present, which makes an instance usable as a retrieval backend rather than only as storage. Nothing is read-only here, so delete, json_del and xgroup_destroy are one call away from whatever session is connected.
The Redis project's own server over a live connection to an instance: 53 tools covering the core data types, pub/sub, consumer groups, and the search and vector index commands.
- Every core type through its own calls: strings, hashes, lists, sets, sorted sets and JSON documents addressed by path, most of them taking an optional expiry
- Key housekeeping without a shell: type checks, renames, expiry, deletes, and two scans — scan_keys for one cursor page, scan_all_keys to iterate a pattern to the end
- Streams end to end: append entries, read ranges, create and destroy consumer groups, read as a named consumer, and acknowledge what was processed
- Pub/sub with a handle you can return to — subscribe by channel or by pattern, read pending messages with a timeout and a message cap, then unsubscribe
- Vector search on hashes: store and retrieve float vectors, create an HNSW index with the dimension and distance metric you choose, then run KNN, or hybrid_search to filter by metadata first and rank by similarity after
- Operational reads alongside the data ones: index schemas via FT.INFO, indexed-key counts, dbsize, INFO by section, and the list of connected clients
- search_redis_documents, a query against Redis's curated documentation, in the same tool set as the data calls
A Redis instance the server can reach, and uv with a Python available (UV_PYTHON). The vector index and KNN tools target Redis 8 or later. MIT licensed.
One command — uvx --from redis-mcp-server@latest redis-mcp-server --url redis://localhost:6379/0