Private Preview - features may change without notice
GramSpec
Architecture Spec API Reference Pricing About
Start free

Build on it · 18

Your AI assistant reads the model

MCP is the standard AI tools now use to reach outside systems, and we serve it. One line of setup in Claude, Cursor or anything else that speaks it, and their own assistant can read their model and ask questions of their database, read-only.

The MCP server. One flow connects an AI assistant once: it is told where to sign in, the person signs in and approves, and the assistant is given a token. The other is a question: the assistant calls read-only tools that read the model and run checked queries on the database.GRAMSPECYOUR DATASomeone at yourcompanyYour AI assistantClaude, Cursor, anything thatspeaks MCPSign in and approvein the browser, two-factorincludedAccessa key, or a token thatrenews; kept only as hashesThe MCP endpointone address, for anyassistantEight read-onlytoolslist, load, search,describe, check, runThe modelChecks, then runa plain read, true to themodelYour databaseread-only123412345 The MCP server. One flow connects an AI assistant once: it is told where to sign in, the person signs in and approves, and the assistant is given a token. The other is a question: the assistant calls read-only tools that read the model and run checked queries on the database.GRAMSPECYOUR DATASomeoneat yourcompanyYour AIassistantClaude,Cursor,anythingthat speaksMCPSign in andapproveAccessThe MCPendpointone address, forany assistantEightread-onlytoolsThe modelChecks, thenrunYour databaseread-only123412345
Blue circles connect the assistant, once. Green squares are a question asked of it.

Connecting it, once

  1. Someone adds GramSpec to their assistant. One address. They can give it a key made on their account page, or simply sign in when asked.
  2. The assistant's first call is told where to sign in. With no key, the answer names the sign-in address, and the assistant registers itself and opens it in the browser.
  3. The person signs in to GramSpec and approves. Through the ordinary sign-in, a second factor included where they have one. The assistant never sees their password.
  4. The assistant is given a token. Good for an hour and renewed by itself for up to ninety days. Tokens and keys are random, and only their hashes are kept. Each call after that is checked against them.

A question

  1. Someone asks their assistant a business question. In the tool they already work in. The language model doing the reasoning is theirs.
  2. The assistant calls GramSpec. Each call carries the key or token, and a call from a web page on some other site is refused.
  3. It works through eight tools, all read-only. List the models on the account, load one or get an overview of it, search it, read one thing in full, have SQL written, check SQL, and run it. The model itself can also be attached to the conversation.
  4. SQL is checked against the model before it runs. The same checks Reason's own queries pass: a plain read, nothing from the database's system tables, and tables and joins that are the ones the model declares. The assistant can ask for the check alone.
  5. It runs on their database, read-only, and the rows go back to the assistant. When the model has no path for a question, what goes back is a refusal that names what is missing, not a guess.