AI agents

Let AI agents work on your database without being able to wreck it: a key per agent with its own Postgres role, a query firewall, human approval for risky statements, a full query log and one-click undo of an agent session.

Agent keys

One key per agent, with its own Postgres role and limits.

Query firewall

What agents may run, checked by the real Postgres parser.

Approvals

Risky statements wait for a person.

Sessions and undo

Revert everything one agent session changed.

MCP and direct connections

Use a key with Claude, ChatGPT, Cursor or psql.

Exposure check

See what anonymous visitors can read and write.

#Why

An AI agent with your database password or service role key can do anything: drop a table, delete every row, read every user's data. Agent keys give each agent exactly what its job needs, and keep a person in the loop for the rest.

Layer	What it does
Postgres role	Each key has its own role with only the grants of its preset and table list. Row Level Security applies. This is enforced by Postgres itself, whatever SQL the agent sends.
Firewall	Every statement sent through the API or MCP is parsed with the Postgres parser. Transaction control, `SET ROLE`, `DO`, `COPY`, dynamic SQL, dangerous functions and denied tables are refused with a reason the agent can read.
Approvals	`DELETE` / `UPDATE` without `WHERE`, changes to more rows than the key allows, drops, policy and privilege changes return `202 pending_approval` until a person approves them.
Query log	Every statement, its result (ran, blocked, needs approval, error), rows changed and duration, per agent and session.
Undo	Every row an agent changes is recorded with its old and new values, every schema change with its inverse. Revert one session; other people's writes are left alone.

#Quick start

Studio → AI agents → Agent keys → New agent key. Pick a preset and the tables it may use.
Copy the key (shown once) into your agent: as the MCP token, or as Authorization: Bearer for POST /platform/v1/agent/sql.
Watch Query log, answer Approvals, and use Sessions & undo if something went wrong.

Terminal

curl https://stitchdb.com/platform/v1/agent/sql \
  -H "Authorization: Bearer sdb_agt_..." \
  -H "X-Stitch-Session: fix-signup-bug" \
  -H "Content-Type: application/json" \
  -d '{"query": "update profiles set plan = 'pro' where id = 42"}'