Redis MCP Server

What Redis MCP Server does

Redis MCP Server exposes Redis and Valkey inspection tools through the Model Context Protocol. It is designed for investigating what is stored in an instance, checking operational health, and identifying common database risks. The default posture is read-only, making it suitable for asking an agent about a production Redis instance without granting unrestricted command access.

Key enumeration uses cursor-based SCAN rather than KEYS. Results are bounded, can be filtered by glob pattern or Redis value type, and can be resumed with a returned cursor. The advisor also samples keys with SCAN when looking for large values and missing expirations.

How it works

The server connects using a Redis URL supplied through its environment. redis_get determines a key's type before returning a suitable representation: strings, hashes, lists, sets, sorted sets, and streams are handled differently. Collection reads are limited, and large string values are byte-windowed. Responses include size and truncation information so a client can tell when it received only part of a value.

redis_key_info provides type, expiration, encoding, serialized memory usage, and idle time without loading the value. This makes it a useful first check for unfamiliar or potentially large keys. Health reporting combines Redis INFO data, database sizes, and recent slow-log entries into one snapshot. The advisor reports findings for sampled big keys, missing TTLs, eviction pressure, and fork-latency risk, with severity and a suggested fix.

Commands sent through redis_command pass through a curated safety gate. Read commands are allowed by default. Mutating commands require ALLOW_WRITES=1; commands for arbitrary execution, transactions, monitoring, shutdown, cluster administration, and similar high-risk operations remain unavailable.

Setup and configuration

A standard stdio configuration can run the package through npx:

{
  "mcpServers": {
    "redis": {
      "command": "npx",
      "args": ["-y", "@yawlabs/redis-mcp@latest"],
      "env": {
        "REDIS_URL": "redis://:password@host:6379/0"
      }
    }
  }
}

REDIS_URL is required. It can use a regular Redis connection string or rediss:// for TLS. On Windows, the configuration wraps npx with cmd /c.

The server starts in read-only mode. Set ALLOW_WRITES=1 or true to permit the curated mutating commands accepted by redis_command; this does not enable blocked arbitrary-execution commands. The package also supports limits for key and value results, including REDIS_MAX_KEYS and REDIS_MAX_VALUE_BYTES, which bound returned data.

For stronger access control, use a least-privileged Redis ACL user in the connection URL. After changing a stored connection value in Yaw MCP, deactivate the connection or start a new session so the new value is loaded.

Tools and capabilities

• redis_scan: Enumerate keys with resumable, bounded SCAN operations, optional glob matching, and type filtering.
• redis_key_info: Inspect TTL, type, encoding, memory footprint, and idle time without reading the value.
• redis_get: Read supported Redis value types with collection limits and string byte windows.
• redis_command: Run commands through read and write allowlists.
• redis_health: Review memory, eviction, clients, throughput, hit/miss ratio, persistence, replication, key counts, TTL coverage, and recent slow commands.
• redis_slowlog: Retrieve recent slow commands with execution time, timestamp, and client details.
• redis_advisor: Review sampled findings for big keys, missing TTLs, eviction pressure, and fork latency.

Limitations and notes

This is an explorer and diagnostician, not a general Redis administration console. Pub/sub, MONITOR, cluster management, replication control, scripting, and other high-risk operations are intentionally excluded. Use redis-cli for those tasks.

Advisor findings about large keys and missing TTLs apply to the sampled portion of the keyspace, not necessarily every key. SCAN can return a key again across resumed scans, although keys present for the entire scan are not missed. Some health fields become null when the connected Redis server does not expose them through INFO.