Your agent works in your apps, as you.
Every suite is an MCP server and every app a REST API. Connect Claude, ChatGPT, Gemini or any MCP client: it signs in as you, sees what you see, and can never do more than you can.
Four of your deals close in November, $266,400 in all:
- Northwind TradingNov 21$48,000
- Orbit StudioNov 14$126,500
- Bramble FoodsNov 19$39,900
- Pine & PartnerNov 26$52,000
One address per suite
A client needs nothing but the suite's address. The server answers with where to sign in, your browser opens, and you choose which apps the agent may open and what it may do there.
claude mcp add --transport http acme https://acme.grid82.app/mcpIn the Claude app, ChatGPT, Cursor, VS Code or Codex, add https://<suite>.grid82.app/mcp as a remote MCP server. Each suite also serves a setup guide written for the agent itself, at /agent-setup.md.
Tools built from what you can see
The tool list, the API document and every answer are produced for the person behind the connection.
Sam's rule
Absent, not refused
A table you cannot read is not offered to the agent at all, so its name never appears.
Narrower, never wider
At consent you can give an agent fewer apps and fewer rights than you have. Never more.
Two kinds of account
A key made inside an app can never change that app's structure. Designing apps is the Studio's job, on a separate sign-in.
Keys that expire
A key made in an app lasts ninety days unless revoked sooner.
The agent proposes, the person applies
The assistant inside Grid82 never writes on its own: each change is a card showing what would change, applied only once you approve it. Every write, from any surface, is recorded and can be undone.
Here is the change. Nothing is saved until you apply it.
Everything is recorded
Every write, from the grid, the API, MCP or the assistant, goes into the app's history with who made it.
Everything can be undone
An approved change can be reverted, and reverting a revert is just undo again.
One tool per operation, or one script
/mcp
One tool per operation, which is what most clients expect and what a small model handles best.
/mcp/code
Two tools, search and execute: the model reads the API document, then writes one script that makes many calls, in a sandbox with no internet access and fixed limits.
/<app>/mcp
A server for a single app, for an integration that should touch that app and nothing else.
Built for a model's budget
Paged lists
A list returns at most 200 rows at a time, with a cursor for the next page, so a model never swallows a whole table by accident.
Safe retries
A write sent twice with the same idempotency key is applied once, so a client that lost the answer can simply try again.
Shapes on request
The suite server does not repeat every table's schema in every tool description. The agent asks for the app it cares about.
Every endpoint and tool is in the documentation.
Questions
Which agents can connect?
Any client that speaks MCP over HTTP with OAuth: Claude Code, the Claude app, ChatGPT, Cursor, VS Code, Codex and others. Agents of your own can call the REST API instead.
Can an agent see more than I do?
No. It signs in as you, and at consent you can only narrow what it may do. Tables you cannot read are not even offered to it.
Can it change my data?
Through MCP or the API, an agent writes as you, within your rights, and every write is recorded and can be undone. The assistant inside Grid82 asks before each change.
Where is the reference?
Every endpoint and tool is in the documentation, and each suite serves a setup guide for agents at /agent-setup.md.
Connect your agent to your apps
Your permissions, in every client. Access opens from the waitlist.