MCP Server
The Selfhost.dev MCP Server:
Your AI DevOps Engineer.The Selfhost.dev MCP server gives Claude, Cursor, Devin Desktop and VS Code direct control of your backend: managed PostgreSQL, MySQL, Redis, ClickHouse, OpenSearch and Kafka on AWS, MongoDB in your projects, GitHub deploys and all 48 one-click configurations across 42 apps in the catalog. Provision, scale and back up, restore PostgreSQL to the second and MySQL to about the minute, then ship it all without leaving your editor. 239 production-ready tools across 37 modules.
And because every action draws on prepaid credits with no subscription, no agent can run up a surprise bill.
+$5.00 for the infrastructure your AI manages · no card needed
What we believe
One server.
Your entire backend.
One protocol. Total control.
MCP is the standard that lets your AI tools work with the system through a controlled interface, enabling actions, not direct access.
Operate through AI
Selfhost.dev exposes managed databases and Projects through MCP: projects, GitHub deployments, project databases, domains, environment variables, SSH keys and billing. Its one-click deployment scope is the whole catalog, all 48 configurations. The deepest first-party infrastructure MCP server in the category - see how competitors compare.
How it connects
239 MCP tools, 37 modules
Everything your
infrastructure needs.
Create, scale, reboot and delete managed PostgreSQL, MySQL, Redis, ClickHouse, OpenSearch and Kafka instances on AWS. Fork the first five with their data; Owners and Admins can read the master credentials again. Platform-managed or BYOC.
$ Spin up a ClickHouse cluster for analytics with 2 replicas
⚡ calling create_instance ...
✓ Instance created: analytics-ch
ap-south-1 · 3 data nodes · ClickHouse 26.8
Keeper quorum embedded · provisioning
Engine support
What each database
can do from MCP.
| Capability | PostgreSQL | MySQL | Redis | ClickHouse | OpenSearch |
|---|---|---|---|---|---|
| Versions, default in bold | 18, 17, 16 | 8.4, 8.0 | 7.2, 7.0 | 26.8 LTS | 2.19 3.8 from the console |
| Default port | 5432 | 3306 | 6379 | 8123 HTTP, 9000 native | 9200 HTTPS |
| Minimum RAM per node | 1 GB | 1 GB | 0.5 GB | 4 GB | 4 GB |
| High availability | Up to 2 replicas automatic failover | Up to 2 replicas automatic failover | Up to 2 replicas automatic failover | Up to 5 replicas on a Keeper quorum | 3 managers plus data nodes chosen at create |
| Snapshots, fork, resize, metrics and alerts | Yes | Yes | Yes | Yes | Yes |
| Data backups to S3 | - | - | - | Yes restored in place | - |
| Point-in-time recovery | Yes to the second | Yes to about the minute | - | - | - |
| Connection pooling | PgBouncer port 6432 | ProxySQL port 6033 | - | CHProxy port 9090 | - |
| Parameter tuning | Yes | Yes | - | - | - |
| Database users | Yes | Yes | - | Yes | Yes as security roles |
| Extensions and plugins | pgvector, TimescaleDB, PostGIS and 8 more | - | - | - | k-NN built in Dashboards on or off |
| Client TLS | Require or relax | Switch in the console | - | Switch in the console | HTTPS always |
| Logs and query statistics | Yes | Yes | - | Yes | Yes cluster snapshot |
As of @selfhost.dev/mcp-server 0.12.0. Kafka, the sixth managed engine, is created from MCP too: one broker or three across availability zones, with TLS on by default and an optional Schema Registry.
ClickHouse is the one engine whose topology changes the bill. Replication runs on a Keeper/Raft quorum rather than streaming replication, so at exactly one replica the platform adds a third keeper-only node to hold the quorum and that node is billed. At two or more replicas the quorum runs embedded on the data nodes and the extra line item disappears. Pass the engine to estimate_instance_cost and it prices and itemizes the difference before anything is provisioned.
One prompt, end to end
One prompt.
The whole deploy.
Ask once. Your agent provisions the database, wires the connection string, deploys from GitHub and brings the domain live, calling the right tools in order and confirming each step.
> Spin up a Postgres DB for acme/web, wire it into the app and ship it to production.
⚡ create_instance
✓ prod-db ready · us-east-1 · t4g.small · PostgreSQL 18
⚡ merge_env_vars DATABASE_URL → acme-web
✓ wired · encrypted at rest
⚡ trigger_deploy acme/web @ main
✓ build passed (nixpacks) · 38s
⚡ add_deployment_domain app.acme.com
✓ DNS verified · SSL issued
✓ Live at https://app.acme.com (42s)
One prompt. Database provisioned, config wired, app deployed, domain live. Destructive steps still ask first and every action bills against prepaid credits, so nothing runs up a surprise.
Quickstart
Three steps,
that's it.
$ npm install -g @selfhost.dev/mcp-server
✓ Installed @selfhost.dev/mcp-server
# Add to your MCP config:
# "selfhost": { "command": "npx", "args": ["@selfhost.dev/mcp-server"] }
$ npx @selfhost.dev/mcp-server
Opening browser for authentication...
✓ Signed in as you@acme.com
✓ Organization: Acme Inc
✓ 239 tools registered. Ready.
Security
Your data stays
where you put it.
The MCP server runs on your machine. It talks to the Selfhost.dev API to manage infrastructure - it never connects to your databases directly. Credentials are stored locally with owner-only permissions.
Runs locally on your machine
Service process stays in your terminal, not our cloud
No database connection
Manages infrastructure via API - never touches your data
Destructive ops require confirmation
delete, stop, reboot - Claude will always ask before acting
Credentials stored with chmod 600
~/.selfhost/credentials.json - owner-only access