Aiven MCP Server
A Model Context Protocol (MCP) server for the Aiven cloud data platform.
Manage PostgreSQL, Apache Kafka, applications, and other Aiven services directly from AI assistants like Claude, Cursor, and VS Code Copilot.
[!WARNING] Use with care. This MCP server can create, modify, and delete Aiven services and data on your behalf. AI agents may execute destructive actions (dropping databases, deleting services, producing messages) based on their interpretation of your prompts. You are fully responsible for the actions taken through this tool.
Permissions: Access is governed by the Aiven user permissions associated with the authenticated account. The MCP server can only perform actions that your Aiven user is allowed to do.
AI Agent Security: AI agents may need access credentials (database connection strings, streaming tokens) to act on your behalf. Review what your agent is doing, especially in production environments. Follow your organization's security policies and do a risk assessment before giving AI agents access to sensitive resources.
Quick Start
Option 1: Remote (hosted by Aiven)
The MCP server is hosted at https://mcp.aiven.live/mcp. Your MCP client will prompt you to authorize on Aiven.
Claude Code
claude mcp add --scope user --transport http aiven-mcp "https://mcp.aiven.live/mcp"
Cursor
Or manually add to Cursor MCP settings:
{
"mcpServers": {
"aiven-mcp": {
"url": "https://mcp.aiven.live/mcp"
}
}
}
VS Code / Copilot
Add to .vscode/mcp.json in your workspace:
{
"servers": {
"aiven-mcp": {
"type": "http",
"url": "https://mcp.aiven.live/mcp"
}
}
}
Read-Only Mode (Remote)
Enable read-only mode by adding ?read_only=true to the URL. All write operations will be excluded from the MCP:
{
"mcpServers": {
"aiven-mcp": {
"url": "https://mcp.aiven.live/mcp?read_only=true"
}
}
}
Scoped Tools (Remote)
Reduce the tool surface exposed to your AI agent by adding ?services_scope= to the URL. Useful when you only work with a subset of Aiven services and want to keep the agent's context focused. Combine values with commas. core (project/service discovery) is always included implicitly.
Valid scopes: all, core, pg, kafka, application, integrations, opensearch. Use all to explicitly load every tool (same as omitting the param). all cannot be combined with other scopes. opensearch follows the same rules as kafka or pg (core plus OpenSearch-category tools); there are no OpenSearch-specific tools yet, so it currently exposes core only.
{
"mcpServers": {
"aiven-mcp": {
"url": "https://mcp.aiven.live/mcp?services_scope=kafka"
}
}
}
You can also combine with read_only:
https://mcp.aiven.live/mcp?services_scope=pg&read_only=true
Write Exceptions in Read-Only Mode (Remote)
When read_only=true, add ?write_allowlist= to re-enable specific write tools while keeping
everything else read-only. Useful when you want mostly-read access but still need to allow one
write action, for example creating Kafka topics. Combine multiple tool names with commas. Ignored
when read_only is not enabled.
https://mcp.aiven.live/mcp?read_only=true&write_allowlist=aiven_kafka_topic_create
Marketplace Customers (Remote)
If you subscribed to Aiven through a cloud marketplace, add your marketplace as a path segment so sign-in uses the correct console:
{
"mcpServers": {
"aiven-mcp": {
"url": "https://mcp.aiven.live/mcp/<marketplace>"
}
}
}
The path segment combines with the query parameters above, for example https://mcp.aiven.live/mcp/gcp?services_scope=pg&read_only=true.
Option 2: stdio (local)
Run the server locally as a child process of your MCP client. Requires Node.js 18+.
You must provide your Aiven API token via the AIVEN_TOKEN environment variable. Create a token here.
Claude Code
claude mcp add --scope user aiven-mcp -e AIVEN_TOKEN=your-token-here -- npx -y mcp-aiven
Cursor, VS Code -- add to your MCP client config:
{
"mcpServers": {
"aiven-mcp": {
"command": "npx",
"args": ["-y", "mcp-aiven"],
"env": {
"AIVEN_TOKEN": "your-token-here"
}
}
}
}
Config file locations:
- Cursor: Cursor Settings > MCP Servers
- VS Code:
.vscode/mcp.jsonin your workspace
Option 3: Local development
Run a local build of the server (useful for development and testing):
pnpm install && pnpm generate:api-types && pnpm generate && pnpm build && AIVEN_TOKEN="<YOUR_TOKEN>" MCP_TRANSPORT="http" PORT=3000 node dist/index.js
The server listens on port 3000 by default. Connect your MCP client to http://localhost:3000/mcp.
To point a remote deployment at a custom host (e.g. your local build), set MCP_HOST:
MCP_HOST=http://localhost:3000 node dist/index.js
Environment Variables
In remote (HTTP) mode, AIVEN_TOKEN is not needed. Your MCP client sends your token as a Bearer token with each request.
Production HTTP traffic is rate-limited in two layers: Cloudflare enforces a per-client-IP limit (configured in the Cloudflare dashboard), and this server enforces MCP_HTTP_RATE_LIMIT_* per bearer token on POST /mcp.
Edge protection rollout
When EXTRA_PROTECTION=true, any mismatch between MCP_EDGE_AUTH_SECRET and the value Cloudflare injects as X-Edge-Auth causes every request to return 403 (except GET /health). Both values are environment/config on opposite sides of the wire, so the only recovery path is to fix the secret and redeploy or update Cloudflare.
Enable in this order:
- Cloudflare Transform Rule — Add a rule that sets
X-Edge-Auth(and, if used for PG tools,X-Client-IP) on traffic to the MCP origin. Note the secret value you configure. MCP_EDGE_AUTH_SECRET— Deploy the server with this env var set to the same secret as the Transform Rule. LeaveEXTRA_PROTECTIONunset orfalsefor now; verify the origin still accepts traffic.EXTRA_PROTECTION=true— Enable only after steps 1–2 are live and matched. Confirm a normal MCP request succeeds and direct origin access withoutX-Edge-Authis rejected.- Secret rotation — Update Cloudflare and
MCP_EDGE_AUTH_SECRETtogether (or briefly setEXTRA_PROTECTION=false), redeploy, then re-enable. Never rotate one side alone while the flag is on.
If EXTRA_PROTECTION=true at startup and MCP_EDGE_AUTH_SECRET is missing, the process exits immediately with an error.
While rejections continue, the server logs a misconfig warning at most once every 15 minutes (reset after a request with valid X-Edge-Auth), so a secret mismatch is visible in logs without one line per rejected request.
Tools
Core
Kafka
PostgreSQL
Applications
Documentation
Contributing
See CONTRIBUTING.md for development setup, running locally, and adding new tools.