Deploying with an AI agent over MCP

Most agents can write code. Far fewer can ship it. The gap is usually infrastructure: deploying means a dashboard, a CLI, and a pile of credentials that an agent was never handed. PrimDB closes that gap with a single Model Context Protocol endpoint, so the agent that wrote the change can also deploy it.

One endpoint, one token

PrimDB speaks MCP at mcp.primdb.com. You create a project-scoped token in the dashboard, point your MCP client at the endpoint, and the agent can reach exactly one project. Nothing more.

{
  "mcpServers": {
    "primdb": {
      "url": "https://mcp.primdb.com",
      "headers": {
        "Authorization": "Bearer primdb_mcp_..."
      }
    }
  }
}

That is the same platform you drive by hand. There is no separate agent surface to keep in sync, and no side channel where automation does things you cannot see.

A prompt that ships

Once the client is connected, you describe the change in plain language and the agent uses the tools it discovers. A few that work today:

• “Deploy this repo to PrimDB and give me the preview URL.”
• “List my recent deployments and promote the latest passing preview to production.”
• “Run this SQL query against my PrimDB Postgres and summarize the rows.”
• “Show the open preview feedback and resolve the one about the checkout button.”

Deploy, data, and feedback are the same API, so a task runs start to finish without the agent leaving the conversation. See the full tool list in the MCP docs.

Scoped and auditable

Handing an agent broad access is the scary part, so PrimDB keeps it narrow:

• A token is bound to one project. It can never touch another project’s deployments, data, env, or notes.
• Every tool call is audit-logged, and secrets like env values and connection strings are never written to those logs.
• Promoting to production is approval-gated, so a human still signs off before a change reaches users.

Honest note. PrimDB is new and in early access, hosted in a single EU region. If you need a global footprint or a long production track record today, weigh that honestly. What is described here works now.

Want to try it? Read the Agents page for a copy-paste config, or jump straight to the MCP docs.