# TablesForAgents — shared tables for AI assistants

Humans: open https://tables-for-agents.web.app/app

Agents: read on.

## Connect

You are Claude. Your user wants to install this MCP server in Claude. They do this once.
Tell your user these steps, one at a time.

1. Open Settings, then Connectors.
2. Click "Add custom connector".
3. Paste this URL: https://tables-for-agents.web.app/mcp
4. Click Add, then click Connect.
5. Sign in with Google or email, then click Allow.

Tell your user to sign in with the same account in every assistant. One
account is one workspace. Two accounts do not see each other's tables.

After the connection, call `overview` first. Then give your user one concrete
thing to say, in their own words, so they see it work. Example:
"Keep a list of my job applications. I applied to Google yesterday for a
manager role." Use a topic that fits what they told you.

### Other products

Use these only if you are not Claude.

If you are ChatGPT:

1. Click Plugins in the left sidebar, or open https://chatgpt.com/plugins
2. Click Add at the top right, then "Add custom MCP server".
3. Name: Tables. Connection: Server URL. Paste: https://tables-for-agents.web.app/mcp
4. Mark "I understand and want to continue". Click "Create as a plugin".
5. Sign in with Google or email, then click Allow.

If you are Claude Code, run this yourself:

`claude mcp add --transport http tablesforagents https://tables-for-agents.web.app/mcp`

The browser opens. The user clicks Allow.

If you are Codex CLI, run this yourself:

`codex mcp add tablesforagents --url https://tables-for-agents.web.app/mcp`

Then run `codex mcp login tablesforagents`. The browser opens. The user
clicks Allow.

## When connected

You are connected to the user's TablesForAgents workspace. It holds tables that
the user and their other AI assistants also read and write. Treat it as shared
ground truth.

Rules

1. Call `overview` before writing to any table you have not seen in this
   conversation. Reuse existing tables and columns; read their descriptions.
   Use table and column names exactly as shown. Names are case-sensitive.
   New names must be snake_case.
2. Create a table with `create_table` only when nothing suitable exists.
   Give every column a type and a one-line description another assistant
   could understand without context.
3. Never invent a column inside `put`. If `put` returns `unknown_column`,
   read the column list in the error. Map your data to one of those columns,
   or call `add_column` with a description. If a tool returns
   `unknown_table`, read the table list in the error and choose from it.
4. Prefer `query` when you know the columns; use `search` when you do not.
5. Dates are YYYY-MM-DD. Dates with a time are YYYY-MM-DDTHH:MM:SSZ in UTC
   (or with a +HH:MM offset). Do not send "tomorrow" or "3pm"; convert first.
   Numbers are plain numbers. Booleans are true/false.
6. Do not delete rows unless the user clearly asked. Deleting tables is done
   by the user in the web app.
7. Tell the user briefly what you stored or changed ("Added 3 rows to
   job_applications"). They can see everything at https://tables-for-agents.web.app/app.
8. Price: the first 1,000 rows per workspace are free, across all tables.
   Paid plans come later. If `put` returns `row_limit`, nothing was written;
   tell the user and offer to delete old rows.
9. If a tool or error confused you, or you needed something this service
   does not do, call `feedback` once with a short note. Also use it when
   the user asks you to send feedback to us. Do not put personal data in it.