redis-mcp

A safe, connection-efficient Redis MCP server.

Design goals

• No accidental data loss. Destructive tools (redis_delete, redis_expire, redis_flushdb) refuse to run unless called with confirm=true — the tool description instructs the AI to ask the user first. redis_flushdb additionally requires the literal phrase confirm_phrase="FLUSH ALL DATA". redis_set refuses to overwrite an existing key unless overwrite=true.

• Gentle on the Redis server.

  • One lazy, reused connection per MCP process (lazyConnect) — no connection churn, no pool needed since stdio MCP calls are sequential.

  • SCAN with cursor + result limits instead of KEYS (never blocks the server).

  • UNLINK instead of DEL and FLUSHDB ASYNC — deletes happen in a background thread on the Redis side.

  • Collection reads are truncated to 200 elements.

Related MCP server: mcp-database

Tools

Requirements

• Node.js ≥ 18

• A reachable Redis server

Install

git clone <this-repo> redis-mcp
cd redis-mcp
npm install

Configuration is a single env var:

• REDIS_URL — default redis://127.0.0.1:6379. Examples: redis://user:pass@host:6379/2, rediss://host:6380 (TLS).

Run manually to verify: REDIS_URL=redis://127.0.0.1:6379 node src/index.js (it prints [redis-mcp] ready on stderr).

In all snippets below, replace /absolute/path/to/redis-mcp with where you cloned it.

Claude Code

claude mcp add redis --env REDIS_URL=redis://127.0.0.1:6379 -- node /absolute/path/to/redis-mcp/src/index.js

Or add to .mcp.json in your project (or ~/.claude.json for global):

{
  "mcpServers": {
    "redis": {
      "command": "node",
      "args": ["/absolute/path/to/redis-mcp/src/index.js"],
      "env": { "REDIS_URL": "redis://127.0.0.1:6379" }
    }
  }
}

Claude Desktop

Add the same mcpServers block to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json).

Codex (OpenAI Codex CLI)

Add to ~/.codex/config.toml:

[mcp_servers.redis]
command = "node"
args = ["/absolute/path/to/redis-mcp/src/index.js"]
env = { REDIS_URL = "redis://127.0.0.1:6379" }

OpenCode

Add to opencode.json (project) or ~/.config/opencode/opencode.json (global):

{
  "mcp": {
    "redis": {
      "type": "local",
      "command": ["node", "/absolute/path/to/redis-mcp/src/index.js"],
      "environment": { "REDIS_URL": "redis://127.0.0.1:6379" },
      "enabled": true
    }
  }
}

Kilo Code

Settings → MCP Servers → Edit MCP Settings (mcp_settings.json), or .kilocode/mcp.json in your project:

{
  "mcpServers": {
    "redis": {
      "command": "node",
      "args": ["/absolute/path/to/redis-mcp/src/index.js"],
      "env": { "REDIS_URL": "redis://127.0.0.1:6379" }
    }
  }
}

Antigravity (Google)

Agent panel → MCP servers (⚙) → Manage MCP Servers → View raw config (mcp_config.json):

{
  "mcpServers": {
    "redis": {
      "command": "node",
      "args": ["/absolute/path/to/redis-mcp/src/index.js"],
      "env": { "REDIS_URL": "redis://127.0.0.1:6379" }
    }
  }
}

Safety model

1. The AI calls a destructive tool without confirm → the server refuses and returns a message telling the AI to ask the user first, listing exactly what would be deleted.

2. The user explicitly approves.

3. The AI retries with confirm=true (and the confirm phrase for flush).

The gate lives in the server, so it holds even if the client/AI forgets to ask.

License

MIT