Connect an AI agent to Seqora

Read this when the agent in your editor keeps guessing at your schema, or asks you to paste table definitions into the chat. Seqora has an MCP server built in. Once an agent is connected it can read commerce.orders for itself, run a read-only query and ask whether an index is worth building. It never gets a password and it can't write. Seqora has no model of its own: the agent you already use brings the AI.

Install it into your agent

1. Press ⌘, or Ctrl, and open Agents, under System. The Clients list has one row per client: Claude Code, Cursor, Codex, Claude Desktop, VS Code and Other, each with its maker's mark.
2. Click Install on your client's row. The dialog names the file Seqora will write, or for Claude Code the command it will run. Click Install again to confirm.
3. In the agent, ask it to list your connections. Back in Seqora the row reads Connected, with the time the client was last seen. Before its first call it reads Installed.

What Install changes, per client:

Every entry is named seqora. It runs Seqora's own executable with the argument mcp, and carries a client token in SEQORA_MCP_CLIENT. JSON files are edited in place, so comments, other servers and indentation survive, and the old file is copied to <name>.seqora-backup-<seconds> first.

Add it to another client by hand

1. Click Copy config on the Other row. Seqora makes a token, puts the entry on the clipboard and shows it under the row.
2. Paste it into your client's MCP config. For Other it has this shape:

{
  "mcpServers": {
    "seqora": {
      "command": "<path to Seqora>",
      "args": ["mcp"],
      "env": { "SEQORA_MCP_CLIENT": "<token>" }
    }
  }
}

Copy config works on every row: Codex gets the TOML table, and Claude Code gets the claude mcp add line when claude is installed. Install and Copy config both make a new token, and the client's previous token stops working. Remove revokes the token. The entry stays in the client's file and is refused from then on.

The client starts seqora mcp, which relays the client's messages to the running app over a named pipe that only your Windows user can open, or a Unix socket with mode 0600 on macOS and Linux.

Choose what each connection shares

Under Connections, each connection has Off, Read and Propose. Every connection starts Off, and an agent that names it is told it is not shared with agents. Read opens the read tools. Propose adds the proposal tools. Setting Propose on a connection whose environment is production asks you to type the connection's name.

The EXPLAIN ANALYZE switch beside it is off by default. In 0.6.1 the agent sees the setting in list_connections, and run_query still refuses EXPLAIN ANALYZE whatever it says.

What the agent can call

run_query takes exactly one statement that Seqora classifies as read-only: SELECT, WITH, VALUES, SHOW, or EXPLAIN without ANALYZE. It returns 100 rows by default and 500 at most, cut at 256 KB, with a 15-second timeout. database_brief is the brief, built from the catalog and statistics, never table data. check_index takes up to 5 CREATE INDEX statements and tests each as a hypothetical index against the statements in pg_stat_statements. It answers build, skip, exists, replace or not_testable and creates nothing. advise_indexes is the ranking the index advisor draws.

On Free the agent still sees every tool, and the paid ones start their description with [Pro]. Propose can be set on Free as well. A propose_change there is saved in Seqora, locked, and the agent is told it will open when you start the trial. revise_proposal, get_proposal and advise_indexes are refused.

What the agent can't do

No tool writes to a database. Each run_query opens its own connection in the database's read-only mode, default_transaction_read_only on PostgreSQL, runs the one statement and closes it. That connection is separate from yours, so unlocking the connection in Seqora's window gives an agent nothing more. Writes, DDL and scripts are refused before anything runs.

The proposal tools write only to Seqora's own store. A change an agent proposes reaches the database when you apply it in Seqora: see Review an agent's proposal.

See what it asked for

Activity, at the bottom of the page, lists every call: the time, the client with its mark, the tool, the connection and a short summary, then a mark for Done, Refused or Failed. Switch between Today and 7 days. A run_query is logged as its row count and its time in milliseconds, never its rows. Calls older than 30 days are deleted when Seqora starts.