MCP quickstart

PerSQL ships an MCP server at https://mcp.persql.com. Drop it into any MCP client and your agent gets ~40 SQL-shaped tools out of the box.

Get a token

1. console.persql.com → API tokens → New token
2. Copy the psql_live_… value — you’ll only see it once.

Claude Code

One command, no file editing:

claude mcp add persql \
  --transport http https://mcp.persql.com \
  --header "Authorization: Bearer psql_live_xxxxxxxxxxxx"

Restart the Claude Code session. The tools appear immediately.

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "persql": {
      "url": "https://mcp.persql.com",
      "headers": {
        "Authorization": "Bearer psql_live_xxxxxxxxxxxx"
      }
    }
  }
}

Restart Claude. The tools panel now lists persql.list_databases, persql.describe_database, persql.search_schema, persql.query, and the rest.

Cursor

~/.cursor/mcp.json:

{
  "mcpServers": {
    "persql": {
      "url": "https://mcp.persql.com",
      "headers": { "Authorization": "Bearer psql_live_xxxxxxxxxxxx" }
    }
  }
}

What the agent can do

The first three tools are the discoverability stack — call them in this order on a fresh database:

1. describe_database({ database: "acme/orders" }) — full schema + docs.
2. search_schema({ database: "acme/orders", query: "..." }) — find tables by intent.
3. schema_doctor({ database: "acme/orders" }) — lint for LLM-hostile patterns.

Then mutating ones:

• query, batch — run SQL, get cost back.
• pin_branch, claim_handoff — single-use scoped delegation.
• list_snapshots, restore_to_time — point-in-time recovery.

See the MCP guide for the full tool list.

Next step

Once the agent is connected, the per-agent sandbox recipe shows how to give each task its own branch.