gtable
gtable.appOpen the Studio
AgentsSafety and undo

Safety and undo

An agent's writes run as you, straight away, and land in the change history. What that means, how to undo one, and how gtable's own assistant differs.

Writes happen as you

A write from an MCP client is applied at once, with your permissions, exactly as if you had made the call yourself. It is not held for approval: asking first is your client’s job, and most clients ask before running a tool marked destructive. An agent can never do more than you can, because it has no identity of its own to borrow.

Everything lands in the history

Every write, from any surface, is recorded with who made it, when, and how it arrived: the grid, a key, or an agent. A create, edit or delete of records can be undone from the change history, in the app or through the API (changesets.mine, then changesets.undo). One request is one entry, so an agent’s import of five hundred rows is one thing to take back. See Undo and history.

Deleting a table, a field, an app or a suite asks for its current name as confirm, from an agent as from a person, so a loop over the wrong list is refused rather than obeyed.

gtable’s own assistant proposes

The assistant built into gtable works differently: it proposes a change, shows you what it would do, and you apply it. Its proposals are recorded the same way, and an applied one is undone the same way.

Give an agent no more than it needs

  • Connect it to the suite, and tick only the apps it should reach.
  • For a production integration, prefer a single app’s address (/{app}/mcp) or a key with only the scopes it uses.
  • Revoke a connection from the same Connections screen; it stops at once.

Setting up changes nothing

The setup guide tells an agent to verify its connection by reading only (apps_list, then me_get), and never to create, update or delete anything to test it.

Use these docs with your AI tools

An AI agent can read this documentation directly. You do not need an account or an API key. Everything here is public and read-only.

Query these docs via MCP

Recommended

Add this server to Claude, Claude Code, Cursor, Mistral, or any tool that supports MCP. Your agent can then search gtable documentation and read it in full, instead of answering from memory.

https://gtable.app/docs/mcp
  • searchFind the passages that answer a question.
  • fetchRead one page in full, as Markdown.
  • list_pagesSee every page in this documentation.

Query these docs over HTTP

The same tools also work as plain web requests. Use this for scripts, or for any tool that does not support MCP. There is one endpoint per tool. Arguments go in the query string, and the answer comes back as JSON.

https://gtable.app/docs/api/docs/search?query=custom+domain

Read the OpenAPI description. It is built from the same definitions as the tools, so it always matches what the endpoints do.

Read these docs as Markdown

Add .md to any page URL to get its Markdown source. You can also send the headerAccept: text/markdown to the page URL itself.

To read the whole documentation in one file, open llms-full.txt. For a short index of every page, open llms.txt.