Upstash MCP Server

The Upstash MCP Server is the official Model Context Protocol connection to your own Upstash account, letting AI agents manage and debug Redis databases, QStash message queues, Workflow runs, and Upstash Box containers directly. Instead of switching between the Upstash console, terminal, and AI assistant, a coding agent can create databases, search logs, retry failed workflows, or spin up sandbox containers — all through natural language chat in Claude Code, Cursor, Codex, or other MCP-capable clients.

What the server does

Per the provider, for Redis you can create new databases in a specific region, list existing databases sorted by memory usage, have the schema of stored data explained, find keys matching certain patterns along with their contents, create backups and clear databases, and analyze throughput spikes over recent days. For QStash and Workflow, an agent can search logs for failure causes, find and retry failed workflow runs from the last 24 hours, summarize the dead-letter queue by error type, or pause scheduled tasks. For Upstash Box, you can spin up sandbox containers, clone repositories and run tests directly inside them, create and duplicate snapshots, or identify failure causes in container logs.

Free test database without sign-up

One tool is aimed directly at AI agents: via redis_database_start_free or a direct POST request to https://upstash.com/start-redis, a free, temporary Redis database can be started without registration or an API key. The response contains credentials and a quickstart guide in markdown; the database expires after 72 hours but can be permanently claimed via an Upstash account. Per the provider, this is especially useful for agents that need scratch storage on the fly.

Credentials and permissions

Before setup, you need your account email and an API key from the Upstash console. The server supports read-only API keys: when started with one, it automatically disables every state-changing tool, such as creating databases, deleting backups, or retrying workflows — the agent can then only read and query the account, not modify it. Accessing Upstash Box additionally requires a separate Box API key, which can be wired into the server configuration to avoid repeated manual input.

Supported clients

Per the provider, the server works with any MCP-compatible client over stdio, with officially documented setup instructions for Claude Code, Cursor, Devin Desktop, OpenCode, OpenAI Codex, Visual Studio Code, Google Antigravity, Claude Desktop, Zed, DeepSeek Harness, and Gemini CLI. For clients without dedicated instructions, a generic stdio configuration using the command npx -y @upstash/mcp-server@latest is enough.

Provider's note: consider the skill instead of the MCP server

Upstash itself recommends in the repository that, for most workflows, installing the Upstash skill and letting the agent drive the upstash command line through it is preferable to running the MCP server continuously. The MCP server is best suited when the agent needs direct, interactive tool access to live resources — for example, searching logs in real time or retrying workflows mid-conversation.

Telemetry and privacy

Per the provider, the server sends anonymous diagnostic data to Upstash: SDK version, runtime version, and basic platform information such as OS and architecture. Account data, tool arguments, or results are not transmitted. Anyone who doesn't want even that can disable telemetry with the --disable-telemetry flag.

License and cost

The repository is licensed under MIT. The MCP server itself doesn't add separate costs; billing runs entirely through your existing Upstash plan for the managed resources.

Who benefits most

The server is especially useful for development teams already building on Upstash services like Redis or QStash: troubleshooting without switching consoles, automatically creating test databases for new projects, or quickly debugging failed workflows right from the editor. Per the provider, anyone working with just a single Redis database is often better served by the leaner, dedicated Redis MCP server.

An Upstash account (account email + API key from Console → Account → API Keys) plus an MCP-capable client with stdio support (Node.js/npx available). Upstash Box tools additionally require a separate Box API key.

For Claude Code:

claude mcp add upstash -- npx -y @upstash/mcp-server@latest --email YOUR_EMAIL --api-key YOUR_API_KEY

For Codex:

codex mcp add upstash -- npx -y @upstash/mcp-server@latest --email YOUR_EMAIL --api-key YOUR_API_KEY

For any other stdio-capable MCP client, the base command is enough:

npx -y @upstash/mcp-server@latest --email YOUR_EMAIL --api-key YOUR_API_KEY

For read-only access, use a read-only API key from the Upstash console — the server then automatically disables all state-changing tools.

npx -y @upstash/mcp-server@latest --email YOUR_EMAIL --api-key YOUR_API_KEY

Account email plus an API key from the Upstash console, passed as CLI flags on startup. Read-only API keys are supported and automatically disable all state-changing tools. Upstash Box requires an additional Box API key.

The server acts with the permissions of the API key used. With a read-only key, the agent can only read and query, not modify; with full access it can also create databases, delete backups, and retry workflows.

Requests run locally through the npx process and communicate directly with the Upstash API. Per the provider, the server sends anonymous telemetry (SDK version, runtime version, OS/architecture) to Upstash, but no account data, tool arguments, or results. Telemetry can be disabled via --disable-telemetry.

With a full API key, an agent can delete databases, remove backups, or modify production data. Upstash recommends using read-only API keys where possible and granting full access only once a clear need is demonstrated.