@yawlabs/redis-mcp

One click adds this to your local Yaw MCP config so it's available in every Yaw Terminal session. Or install manually below.

Explore and diagnose a Redis instance from Claude Code, Cursor, and any MCP client. Read-only by default - writes opt in via a single env var - and key enumeration always uses SCAN, never the O(N) KEYS, so it is safe to point at a production instance with millions of keys.

Built and maintained by Yaw Labs.

Why this one?

• SCAN, never KEYS. Every key-enumeration path uses cursor-based SCAN with a bounded COUNT and an iteration cap. KEYS * is O(N) over the entire keyspace and blocks Redis's single-threaded event loop for the full scan - a self-inflicted outage on a large instance. SCAN yields between batches. See Security.
• Read-first command gate. Tools run a curated read-only command allowlist by default. Mutating commands (SET, DEL, EXPIRE, HSET, ...) require ALLOW_WRITES=1. Arbitrary-execution commands (EVAL, FUNCTION, SCRIPT, MULTI, MONITOR, SHUTDOWN, CLUSTER, ...) are never exposed, even with writes on - the gate is a curated allowlist, not "anything when writes are enabled".
• Type-aware reads without surprises. redis_get dispatches by value type (string / hash / list / set / zset / stream) and windows collection reads to a cap, so a million-element list can't blow out the model context. redis_key_info reads type / TTL / encoding / memory footprint without pulling the value at all.
• Health in one call. redis_health rolls up INFO + DBSIZE + recent SLOWLOG into memory pressure, eviction policy, hit rate, ops/sec, persistence status, replication role, per-database key counts (and how many lack a TTL), and the most recent slow commands.
• A real advisor. redis_advisor is the "what should I be looking at?" lint pass: big keys, missing TTLs, eviction pressure (including the dangerous noeviction + no-TTL combination), and fork-latency risk - each with a severity and an actionable fix. Keys are SCAN-sampled, so it is safe on a large instance.
• Instant startup. Ships as a single bundled file with zero runtime dependencies. No multi-minute node_modules install on every npx cold start.

Scope

This server is a read-first explorer and diagnostician, not a general Redis admin console. It deliberately does not expose EVAL/FUNCTION/SCRIPT, pub/sub, MONITOR, cluster management, or replication control. For those, use redis-cli directly. The goal here is the safe, common 90%: "what's in this instance, is it healthy, and what should I worry about?" - the questions an agent should be able to answer against a production Redis without risk.

Works against Redis 6+ and Valkey. A few redis_health fields (latest_fork_usec, aof_enabled) depend on the running server exposing them in INFO; missing fields surface as null rather than erroring.

Quick start

1. Create .mcp.json in your project root

macOS / Linux / WSL:

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

Windows:

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

Why the extra step on Windows? Since Node 20, child_process.spawn cannot directly execute .cmd files (that's what npx is on Windows). Wrapping with cmd /c is the standard workaround.

2. Restart and approve

Restart Claude Code (or your MCP client) and approve the redis MCP server when prompted.

3. (Optional) Enable writes

Read-only is the default. To let the agent run mutating commands (SET, DEL, EXPIRE, HSET, ...) via redis_command, add ALLOW_WRITES=1:

"env": {
  "REDIS_URL": "redis://...",
  "ALLOW_WRITES": "1"
}

Prefer scoping this to dev/test instances. Even with writes on, arbitrary-execution commands stay blocked.

Security

SCAN, not KEYS - this is the load-bearing choice. Redis is single-threaded. KEYS pattern walks the entire keyspace in one uninterruptible operation; on an instance with millions of keys it blocks every other client for the duration - effectively a denial of service you triggered yourself. Every key-enumeration path in this server (redis_scan, the advisor's key sampling, redis_get's set reads) uses cursor-based SCAN/SSCAN with a bounded COUNT and a hard iteration cap, which yields the event loop between batches. KEYS is explicitly rejected by the command gate with a nudge to redis_scan.

Read-only by default. Without ALLOW_WRITES=1, only commands on the read-only allowlist run; everything else is rejected before it reaches Redis. With ALLOW_WRITES=1, a curated set of mutating commands is additionally permitted - but arbitrary-execution commands (EVAL, FUNCTION, SCRIPT, MULTI/EXEC, MONITOR, SHUTDOWN, REPLICAOF, CLUSTER, MIGRATE, ...) remain blocked in all modes. The gate is fail-closed: a command on neither allowlist is rejected, so a command we never anticipated can't slip through.

Use Redis ACLs as the primary control. As with a database role, the cleanest posture is a least-privileged Redis user (ACL SETUSER mcp on >pass ~* +@read) in REDIS_URL. Redis then enforces the boundary server-side, independent of this server's gate. ALLOW_WRITES is defense-in-depth on top of that.

See SECURITY.md for vulnerability reporting.

Tools

Configuration

All env vars are read from the MCP server's environment. The last three are read by the redis-mcp launcher rather than the server, so they have no effect when a host runs dist/index.js directly - and neither does the launcher's oam version check (see Runtime):

Connecting to managed Redis (Upstash, ElastiCache, Redis Cloud, etc.)

Use a rediss:// URL for TLS. It works the same on Node and on oam, sandbox included.

If the provider serves a cert signed by a private CA that the runtime's trust store doesn't recognize (symptoms on stderr: self-signed certificate in certificate chain, unable to verify the first certificate, or self-signed certificate, the same on Node and oam), the best fix is to trust that CA: point NODE_EXTRA_CA_CERTS at its PEM file. Node and oam 0.15.3+ both read it, and the connection stays fully verified:

"env": {
  "REDIS_URL": "rediss://default:pass@host:6379",
  "NODE_EXTRA_CA_CERTS": "/path/to/provider-ca.pem"
}

If you cannot get the CA certificate, add REDIS_TLS_REJECT_UNAUTHORIZED=false instead:

"env": {
  "REDIS_URL": "rediss://default:pass@host:6379",
  "REDIS_TLS_REJECT_UNAUTHORIZED": "false"
}

This turns certificate verification off entirely - the chain and the hostname check - so the connection is still TLS-encrypted end-to-end, but no longer authenticated. Prefer NODE_EXTRA_CA_CERTS wherever you can.

Runtime

The published redis-mcp command is a small launcher that prefers the newest oam JavaScript runtime it can find and falls back to Node. It only uses the latest oam release, currently 0.15.3, and never serves on an older one. If you do not have oam, nothing changes: the fallback is automatic, and because npm has already started Node to run the launcher, it is an in-process import() of the server with no extra spawn.

Why 0.15.3. Before it, oam's TLS socket lacked some of the net.Socket methods ioredis calls and did not close itself when the server hung up, so rediss:// crashed the server on its first tool call (YawLabs/oam#132), and oam ignored NODE_EXTRA_CA_CERTS (YawLabs/oam#136). 0.15.3 fixes both, so on it TLS behaves as it does on Node. With only an older oam available, the launcher runs the server on Node instead - in its own process under npx, or handed off to Node on PATH when a host launches it with an old oam run - and says to update oam from oamjs.org. It exits instead when that old oam run host has no Node on PATH, or under REDIS_MCP_RUNTIME=oam. A host that bundles its own oam needs the host's update. A host that runs dist/index.js directly under oam skips all of this; on an oam below 0.15.3 the server then answers every rediss:// tool call with an error naming the version to update to, rather than crashing.

How the runtime is chosen:

• Already on oam. If the launcher itself is running under oam 0.15.3+ (a host that runs oam run on the bin), the server is imported into that process: nothing is discovered or spawned, OAM_BIN is ignored, and REDIS_MCP_RUNTIME=oam is satisfied. REDIS_MCP_SANDBOX=1 skips this step, because only a freshly spawned oam can apply --permission: the launcher relaunches the oam it is running on with the sandbox, even when that oam is in no install location and not on PATH (a usable OAM_BIN still takes precedence).
• Otherwise, discovery. The launcher uses OAM_BIN if it is 0.15.3 or newer. Otherwise it asks every oam it can find - the default install locations (%LOCALAPPDATA%\oam\bin then ~/.oam/bin on Windows, ~/.oam/bin elsewhere), then PATH - for its version, and spawns the newest one at 0.15.3 or newer; on a tie the installed copy wins. On Windows only oam.exe counts. An OAM_BIN that does not exist, is older, or will not run is named on stderr and discovery carries on.
• No usable oam. Under auto the server runs in the launcher's own process instead (an old oam host hands off, see below). The launcher says so on stderr when it passed over an OAM_BIN, a found oam that was too old or would not run, or (on Windows) an oam.cmd / oam.bat on PATH, when the chosen oam failed to launch, and whenever REDIS_MCP_SANDBOX=1 was set. A note that passed over a working oam that is only too old says to update it. With no oam found at all and no sandbox asked for, it falls back silently; the handoff from an old oam host is always noted, in one message written once the launcher knows whether a Node was found. Once oam has started, a failure while running the server is a startup failure, not a fallback.
• An old oam host never serves. If the launcher is running under an oam older than 0.15.3, it hands the server off to the newest usable oam, else to Node on PATH, and exits with an error when there is neither.
• Overrides. REDIS_MCP_RUNTIME=oam turns every fallback into a startup failure. REDIS_MCP_RUNTIME=node skips discovery and always runs Node: in the launcher's own process under Node, handed off to Node on PATH under oam.

No fallback carries --permission - not in-process, and not a handoff to Node - so with REDIS_MCP_SANDBOX=1 under auto the server can end up unsandboxed. Every such fallback says so on stderr (without --permission), but a host may not show stderr. Set REDIS_MCP_RUNTIME=oam alongside it when an unsandboxed server is not acceptable.

Troubleshooting

REDIS_URL is not set - Your MCP client is launching the server without the env var. On Windows especially, env vars set in bash / PowerShell profiles are not inherited by MCP servers launched via cmd. Put REDIS_URL directly in the env block of .mcp.json.

NOAUTH Authentication required - The instance requires a password and the URL has none. Add it: redis://:yourpassword@host:6379 (note the leading colon - the username is empty for the default user).

<COMMAND> mutates state and is blocked: ALLOW_WRITES is not set - You asked for a write through redis_command in read-only mode. Add ALLOW_WRITES=1 to the env block (dev/test), or - cleaner - use a Redis ACL user scoped to the access you want.

KEYS is blocked - Intentional. Use redis_scan (cursor-based) to enumerate keys; it is safe on a large keyspace where KEYS is not.

First command is slow, subsequent commands are fast - Expected. The client connects lazily on the first command; later commands reuse the connection.

Development

npm install
npm test          # build + unit tests (no live Redis needed)

The unit suite covers the pure logic - command allowlist enforcement, SCAN cursor paging, INFO/SLOWLOG parsing, and the advisor heuristics - and runs without a Redis instance. Integration tests that exercise live paths (npm run test:integration) require a disposable Redis at REDIS_URL.

License

MIT © 2026 YawLabs