Skip to main content
The Upstash MCP server lets your agent manage and debug your Upstash resources directly, across Redis, QStash, Workflow, Vector, Search, and Upstash Box. Run it two ways:
  • Remote (hosted) — nothing to install. Connect over HTTP to https://mcp.upstash.com/mcp with OAuth or a developer API key. Covers Redis, QStash, Workflow, Vector, and Search.
  • Local (stdio) — run @upstash/mcp-server with npx, authenticated with your email and API key. Covers Redis, QStash, Workflow, and Upstash Box.
Only running commands against a single Redis database? Use the dedicated Redis MCP with only that database’s REST/TCP credentials.

Remote server (recommended)

Connect any MCP client over streamable HTTP to:
Tokens never reach the agent: data-plane tools resolve each resource’s REST credentials server-side and call the endpoint directly.

Connect

Snippets for popular clients. For every client — and the plugin-based setup — see Install by agent.
Then run /mcpAuthenticate.
The Upstash plugin for Claude Code, Claude Desktop, and Codex bundles this remote MCP server (over OAuth), so installing the plugin sets up the skills and the MCP in one step — no separate command. See Install by agent.

Authentication

Two ways in — the remote server verifies neither locally; the Upstash API is the authority.

OAuth (default)

The first tool call opens a browser consent page where you pick the account scope (personal or a team) and whether the connection is read-only. The grant is per client and revocable from the console under Account → OAuth Clients. Nothing to configure — this is what the Connect snippets above use.

Developer API key (headless / CI)

No browser. Pass an API key as an Authorization header, in the form email:API_KEY:
Create a key from the console — see Developer API → Create an API key.
Read-only API keys (and read-only OAuth grants) are supported — the server then disables every tool that would change state, and credential fields are omitted from tool results.

Feature groups

Append ?features= to the URL to scope which product tools the agent sees — comma-separated values from redis, qstash_workflow, vector_search; omitted means all:
A group covers every product it names, because those products share their tools: qstash_workflow carries the QStash and Workflow tools, and vector_search carries the Vector and Search tools. Each product’s own name works as an alias for its group — qstash, workflow, vector and search are all accepted — so ?features=vector and ?features=vector_search return the same tools.
A name the server does not recognize is dropped, so ?features=redis,typo still returns the Redis tools. But if every name is dropped, the server starts with no tools at all — and a client then reports tools/list as an unknown method rather than showing an empty list. If a fresh connection looks like it has no tools, check the spelling here first.

Tools

38 tools, grouped the way ?features= groups them. The access column is the tool’s own annotation, which clients use to decide what to confirm with you: read never changes anything, write creates or updates, destructive can delete or overwrite data. On a read-only connection every write and destructive tool is refused.
redis_run_command is annotated destructive because it takes arbitrary commands. On a read-only connection it still works, but only for commands on an allowlist of read-only ones — anything else, including a command the server does not recognize, is refused.
Every tool here takes a region (eu or us), which picks the QStash user. The three shared ones take a service (qstash or workflow) to say which side you mean.
A Search database is a Vector index whose named indexes are namespaces, so most tools serve both and take a service (vector or search). Only querying and writing documents differ, which is why those four keep product names.

Local server (@upstash/mcp-server)

A stdio server you run with npx, authenticated with your account email and a Developer API key. Adds Upstash Box tools on top of Redis, QStash, and Workflow. Repository here. Create an API key from the console — see Developer API → Create an API key. The base command every client uses is:
Add it to your client with the stdio transport, e.g. Claude Code:
See Install by agent for a config snippet for every client.
Readonly API keys are supported. When the server starts with one, it disables every tool that would modify state, such as creating databases, deleting backups, or retrying workflows. Your agent can still read and query your account, but it cannot make changes.

Upstash Box API key (optional)

For the local server to interact with Upstash Box, the agent needs your Box API key. To avoid pasting it into the chat each time, wire it into the setup — as a CLI flag or an environment variable.

Example prompts

Redis

  • “Start a free temporary Redis I can experiment with”
  • “Create a new Redis database in us-east-1”
  • “List my databases sorted by memory usage”
  • “Give me the schema of how users are stored in this Redis”
  • “Find all session keys expiring in the next hour and show me their payloads”
  • “Create a backup of this db, then clear it”
  • “Show me throughput spikes during the last 7 days”

QStash & Workflow

  • “Check the QStash logs and figure out why my webhook keeps failing”
  • “Find failed workflow runs for user @ysfk_0x in the last 24 hours”
  • “Retry the failed workflow run that started 2 hours ago”
  • “Summarize what’s in the DLQ right now, grouped by error type”
  • “Pause the daily-report schedule until Monday”
  • “List my Vector indexes and show the dimensions of each”
  • “Query my Vector index for the 5 nearest neighbours of this text”
  • “Upsert these documents into my Search index”

Upstash Box

  • “Spin up a Box, clone this repo, and run the tests”
  • “Snapshot this Box and create 5 copies from it, assign each one a GitHub issue”
  • “My Box keeps failing to start, check the logs and tell me what’s wrong”

Telemetry

The local server sends anonymous diagnostic info to Upstash with each request: the MCP server SDK version, your runtime version (Node, Bun, etc.), and basic platform info (OS and architecture). No account data, tool arguments, or results are collected. To opt out, add --disable-telemetry to the args.