Connect your agent (MCP)

The Destesi MCP server lets an AI agent operate your Destesi data directly. It turns Destesi products into tools that any Model Context Protocol client can call — branch a database, open a preview environment, run a Connect action against GitHub or Slack, and (when you self-host the server) reach Drive, Deploy, Commerce, TTS, Studio, and more.

Every call is authenticated and scoped to your workspace, so the agent only ever sees data you can see.

Two servers, two surfaces

There are two ways to connect, and they expose different tool sets. Pick the one that matches what you need before you wire anything up.

| | Hosted endpoint | Self-run binary | |---|---|---| | How you get it | dst mcp — points your client at Destesi's hosted /mcp endpoint | You run the mcp binary yourself, locally or on your own host | | Setup | One command, nothing to operate | You configure it and keep it running | | Tools | Snap branching (including a SQL query tool), preview environments, and Connect actions | Twelve families: Snap (with fixtures and lineage), Review, Preview, Connect actions, Drive, Deploy, ImageGen, Commerce, TTS, Chat, Studio, Design — about 75 first-party tools. No query tool | | Transport | Streamable HTTP, bearer token per request | stdio (default) or HTTP/SSE |

If you just want an agent that can branch your database and open previews, use the hosted endpoint. If you want an agent that can drive the whole suite, run the binary.

Note

The two servers are separate programs with separate tool registries. A tool documented for the self-run binary is not silently available on the hosted endpoint — see the reference for the exact catalog of each.

Connect Claude Code in one command

If you have the dst CLI and Claude Code installed, this is the fastest path to first value. It wires you to the hosted endpoint.

1. Install the CLI (if you haven't already):

   curl -fsSL https://get.destesi.io/install.sh | sh

2. Add the Destesi MCP server to Claude Code:

   dst mcp

   If you aren't logged in yet, dst mcp opens your browser to authenticate first (just like dst login), then registers the server. It wires Claude Code to the hosted MCP endpoint over HTTP with your bearer token attached.

3. Use it. Open Claude Code and ask your agent to work with your data:

   Clone the production remote into a branch called agent/test, wait for it to be ready, and give me the connection string.

   The agent calls clone → wait → remote_url and hands you a live Postgres DSN — backed by copy-on-write, fully isolated from your source.

Tip

dst mcp registers the server at user scope, so it's available across all your Claude Code projects. Re-run it any time to refresh the token.

What the hosted endpoint gives your agent

• Branch your data — clone a remote or fork a branch into an isolated, copy-on-write copy; branch_list, status, branch_delete; wait blocks until provisioning finishes and remote_url returns the live DSN.
• Register sources — remote_add a Postgres database or S3 bucket; remote_list, remote_show, remote_remove.
• Query a branch — query runs SQL against a ready branch and returns JSON. It runs in a read-only transaction unless you explicitly ask for writes.
• Read history — log lists a remote's checkpoints (point-in-time history).
• Drive preview environments — preview_create, preview_list, preview_status, preview_events, preview_redeploy, preview_delete.
• Act on your connected apps — one tool per action in Connect's catalog (over 200 actions across around 40 providers), each named connect_{provider}_{action}.

Connect any other MCP client

The self-run binary is what you point a desktop client at. It runs over stdio for clients like Claude Desktop and Cursor, and over HTTP/SSE when you want one server to handle several callers. Give it an identity personal access token (PAT) — the same idn_pat_… token dst login stores.

Claude Desktop

{
  "mcpServers": {
    "destesi": {
      "command": "/path/to/mcp",
      "env": {
        "DESTESI_API_KEY": "idn_pat_…",
        "DESTESI_API_URL": "https://api.destesi.io"
      }
    }
  }
}

Local (any stdio client)

DESTESI_API_URL=https://api.destesi.io \
  DESTESI_API_KEY=idn_pat_… \
  mcp

Switch on the product families you want

The self-run binary registers Snap, Review, and Preview tools unconditionally. Every other family is switched on by setting that product's base URL. Leave a variable unset and the family is simply skipped — the server logs a line and every other family keeps working.

DESTESI_API_URL=https://api.destesi.io \
  DESTESI_API_KEY=idn_pat_… \
  CONNECT_API_URL=https://api.connect.destesi.io \
  DRIVE_API_URL=https://api.drive.destesi.io \
  DEPLOY_API_URL=https://api.deploy.destesi.io \
  mcp

See the reference for the complete variable list and the tools each family adds.

Note

Use an identity PAT (idn_pat_…). One identity PAT authorizes the entire surface — Snap branching, workspace discovery, Connect actions, and every product family. Mint one from your account (or run dst login / dst login --token) as described in CLI & API auth.

A word on trust

The MCP server has no consent gate of its own. Tools carry advisory readOnly and destructive annotations that your MCP client can surface before it runs them, but enforcing them is entirely the client's job. One token grants your agent the whole registered surface, and the token is forwarded verbatim to each product — authorization is enforced by the product, not by MCP.

Two practical consequences: treat an identity PAT the way you would treat your own password, and prefer a client that asks before running a write. See CLI & API auth for how tokens are minted and revoked.

Next steps

How it works The mental model, the two servers, and where MCP fits the suite.

Reference Both tool catalogs, transports, env vars, and every tool's parameters.

Troubleshooting Auth errors, missing tools, connection problems, and common fixes.

Snap The branching engine behind the clone and fork tools.