Claude
Connect Claude to CapyDB with the remote MCP server at mcp.capydb.dev - sign in through the browser, no key to paste - or install the CapyDB plugin for the connector plus workflow skills.
What it does
CapyDB runs a remote MCP server at https://mcp.capydb.dev/mcp. Claude on the web, desktop and mobile, Cowork, and Claude Code connect to it directly: there is nothing to install and no API key to copy. The same tools as the local MCP server are available - projects, connection strings, SQL, schema, preview databases, backups, restore points, imports, observability and jobs.
Two ways to add it:
- The connector - the MCP server on its own.
- The CapyDB plugin - the connector plus skills that teach Claude the workflows: setting up a database and wiring its connection string into your app, the rehearse-checkpoint-change-verify loop for risky SQL, and throwaway databases with no account.
Connect in Claude
- In Claude, open Customize → Connectors and add CapyDB - from the directory, or with Add custom connector and the URL
https://mcp.capydb.dev/mcp. - Start using it. Ephemeral databases work without signing in.
- The first time a tool needs your account, Claude shows a Connect button. It opens CapyDB in your browser: sign in, pick the workspace Claude may use (switch organizations on that page if needed), and select Connect. Claude then retries the tool.
Connecting creates an API key in the workspace. A workspace member without the admin role can connect too: the consent page says so, and the key it creates is a member key. A member key can use the scopes below, but none of the organization admin actions - creating or deleting projects, changing project settings, imports, credential rotation, overwrite restores, and API key, webhook and K/V management - just as the member's own session cannot. It is marked Member key under Settings → API keys. There is no need to switch to a workspace you administer, unless Claude should do those things.
On Team and Enterprise plans, an Owner adds the connector for the organization and each member connects with their own CapyDB account.
Connect in Claude Code
claude mcp add --transport http capydb https://mcp.capydb.dev/mcpRun /mcp in a session to sign in. To install the plugin instead, which brings the connector and the skills together:
/plugin marketplace add capydatabase/claude-plugin
/plugin install capydb@capydbClaude Code can also run the local stdio server with an API key, which suits CI and headless use better than an interactive sign-in.
What the connection can do
Connecting mints an agent API key for the workspace you picked. It appears under Settings → API keys named after the app (for example "Claude (MCP)") and carries these scopes:
| Scope | Lets Claude |
|---|---|
projects:read, projects:write | List, create and manage projects and previews; run SQL |
credentials:read | Read connection strings |
jobs:read | Follow provisioning and restore jobs |
backups:read, backups:write | List and take backups, exports and restore points; restore into a preview |
organizations:read | Read usage and plan limits |
It cannot create or revoke API keys, change the organization, or restore over a production database. The key expires after 90 days; Claude then asks you to connect again. Revoke it under Settings → API keys to disconnect at once - the next tool call that needs your account asks you to sign in again.
Creating lasting databases needs an active plan; if the workspace has none, the sign-in page offers one (new accounts start with a free month).
Safety
Every tool declares whether it only reads or can change data, so Claude asks before anything destructive. Reading SQL (query_sql) runs in a read-only transaction; statements that change data go through a separate tool (execute_sql), and CapyDB refuses an UPDATE or DELETE without WHERE and any TRUNCATE. For anything else risky, follow the agent safety loop: rehearse on a preview, create a restore point, change, verify.
Other MCP clients
Any client that supports remote MCP servers with OAuth (Cursor, VS Code and others) can use https://mcp.capydb.dev/mcp. The server's authorization server is the CapyDB API (https://api.capydb.dev), discovered from the server's protected resource metadata; it supports dynamic client registration and requires PKCE.
Data
Tool calls go from your client to mcp.capydb.dev, which forwards them to the CapyDB API with your connection's key and returns the results to your conversation. The MCP server stores nothing. CapyDB records API activity and SQL statements in your project's audit trail and SQL history, as it does for the dashboard and CLI. See the privacy policy.