AI agents
An agent can build a working app in minutes. Running it safely in production is the hard part: pointed at raw data, it fills the gaps with unfounded confidence — invented joins, guessed definitions, stale data treated as current.
Credible gives your agent the right context over MCP and enforces access control on every query, so Claude, ChatGPT, Gemini, Cursor, or the agent you build answers fast, correctly, and securely from day one.
The problem
Text-to-SQL on raw tables gets you a demo. Three things stop it from becoming something you can ship.
Active users, net revenue, which window is “last month” — the definitions that make an answer correct live in docs, old SQL, and a few people's heads, not in the schema the agent can see. So it guesses, and a guessed join double-counts quietly.
An agent holding your credentials can return any row to anyone. Who sees what, row by row, and who asked for what, become code you write — and rewrite — for every agent and every app.
Login, permissions, hosting, secrets, versions. The same agent writes that code in the same session, and it's the last code you want deciding who can see revenue. Most vibe-coded apps stop here.
Get started
Start free and connect your data — Postgres, MySQL, spreadsheets, CSVs. Then run one command in a directory. It installs our open-source modeling skills, wires up MCP, and starts a local model server against your real data. Credentials stay with Credible, never on your laptop.
Your agent drafts the data model; you confirm it's right; publish. From then on, every agent and app you build gets the same model, with the same rules, over MCP.
In Claude Code, Codex, Cursor, or Gemini CLI. Tell it what revenue is, what a customer is, who sees what. The agent drafts the model with the skills; you confirm it's right.
npm install -g @credibledata/cred-cli
cred initOne versioned package: the data model, its access rules, and any app built on it. A published version never changes, so a rename next month can't break what's running today.
cred publishClaude, ChatGPT, Gemini, Cursor, Codex, or the agent you build. One URL for all of them. The engine resolves who's asking and what they're allowed to see.
{
"mcpServers": {
"credible": {
"url": "https://mcp.credibledata.com/global/"
}
}
}How it works
The data model holds what your data means. When an agent asks, the engine matches the question by meaning and returns the slice of the model it needs — not the whole schema stuffed into every prompt — and enforces who sees what on every query.
Every question enters through one gateway over MCP. There it is checked against the model's access rules and logged to a permanent audit trail. The agent only gets back the rows the caller is allowed to see.
Ask about “sports gear” and the agent finds Running Shoes and Athletic Apparel. It matches a question by meaning, not by column name, and gets back the slice of the model that answers it, plus suggested queries that are known to be correct.
The agent queries the model, not the tables. Every definition is written in one place, joined totals never double-count, and every field reference is checked against the live model before a query runs — so the answer is explainable and auditable, not merely plausible.
Small request in, small context out. A correct answer costs a fraction of the tokens — and correctness is the biggest saving of all, because wrong answers are what multiply retries.
Every retrieval is a test of whether the model surfaced the right concepts. Misses become proposed fixes — a better doc line, a missing index, a view worth declaring — as reviewable changes to the model, in git. The more you ask, the better it gets.
Apps
A data app is just files in the same package as the model. Describe the app, the agent builds it, and Credible runs it — no login code, no backend, no server to keep up.
The viewer is already signed in, and the model's access rules decide what they see. A rep and the CFO open the same link and get different rows, and no app code made that call.
Publish, and the app shows up for your team on infrastructure built for production. The app only talks to Credible, so there is no password in its files to leak.
The app, its libraries, and the data model it queries ship as one version that can't change after it's published. Roll back one and you roll back both.
The result
An agent that shows its work instead of asking to be trusted — in the tool your team already has open.
Credible is the engine that plugs into the agent you already use: Claude, ChatGPT, Gemini, Cursor, and the agents you build yourself, over MCP. One data model, the same answers, wherever the question gets asked — and nothing to rebuild when you switch.
Ready when you are
Start free. One command in the coding agent you already use, and every agent and app you build runs on context you can trust.