Use your own agent with Prominara

Use your own agent with Prominara

Prominara now runs a remote MCP server at mcp.prominara.com. Connect Claude, ChatGPT, Cursor or Codex to it and your assistant can read your AI visibility data and act on it: create the prompts you want tracked, run a check, log the page you shipped, and read the measurement verdict when the window closes.

It is included in every plan and in the 14-day trial. There is no separate purchase and no second usage meter.

The hypothesis

Every software product is currently building its own assistant. Ours included: Prominara has a copilot, and people use it.

But watch how anyone actually works for an afternoon. They are in Claude drafting the page. They are in Cursor deploying it. They are in ChatGPT researching the category. Then they open a browser tab, find the dashboard, remember where the button lives, and click it. The dashboard is not where the work is happening. It is where the work has to go to be recorded.

The bet is that people do not want one assistant per product. They want one assistant, and they want it to reach everything. That is what the Model Context Protocol is for: a product exposes a server, and any MCP-capable client can use it. The user picks the assistant. We supply the data and the actions.

If that is right, the interesting question stops being "what should our agent do" and becomes "what should any agent be able to do with our data, safely".

What shipped

Fourteen tools, all prefixed with prominara_, split into two groups.

Nine read tools cover the ground the dashboard covers: list your sites, get the visibility overview for a window with its confidence intervals and competitor ranking, list tracked prompts, read the recorded answers for one prompt, list ranked missions, read the sources intelligence, read measurement verdicts, get suggested prompts, and check plan usage.

Five write tools cover the actions worth taking from a conversation: create tracked prompts, run a check now, log a content change, set a mission status, and enable or disable prompts.

That is the whole surface. Nothing deletes. Nothing touches billing, team settings or another organization. There is no tool that changes what a plan allows.

Three things worth asking

"Which of my tracked prompts lost citations this month, and which page should I fix first?"

The agent reads the visibility overview for the window and the ranked missions, then answers with the prompts that moved and the page that should own each answer. This is the question the dashboard answers too, but you asked it in the middle of a conversation about your content plan, and the answer arrived in that conversation rather than in a tab you have to reconcile afterwards.

"I published the new comparison page today. Log it against the prompts it targets."

This is the one that changes behaviour. Logging a content change is what opens the before-and-after observation window: the affected prompts get frozen, the sampled runs start, and a control group of untargeted prompts carries the counterfactual. It takes about ninety seconds in the app, and it is the step people skip, because it happens after the satisfying part is over.

From a coding agent, it happens in the same session as the deploy. Ship the page, ask the agent to log it, and the measurement window opens the same day. Ninety seconds of friction is enough to lose a measurement. Zero seconds is not.

"Did the page I shipped three weeks ago move anything yet?"

The agent reads the verdict and relays it. Note what it does not do: it does not look at two numbers and decide whether they are different. The verdict comes from Prominara, with its probability, its range and its evidence grade, in wording we control. The agent is a courier for that sentence, not the author of it. A language model handed a before number and an after number will happily narrate a causal story, which is exactly the failure mode this product exists to avoid.

Security, in one paragraph and then in detail

You sign in through OAuth 2.1 with PKCE. There is no key to paste into a config file.

In more detail: every new connection shows a consent page naming the client that is asking, the address it will send you back to, and the organization you are connecting. Nothing is approved silently from a previous grant, so a second client cannot inherit the first one's access. Read and write are separate scopes, and write is a checkbox you can leave unchecked, which gives you a connection that sees everything and changes nothing. Tokens are stored as hashes and never written to logs. Every connection is listed under Settings, then Connected agents, with the client, the scopes, the organization and when it was last used, and revoking takes effect on the agent's next call.

One more thing that matters more than it sounds: when a tool returns the text an AI engine produced about your brand, that text is wrapped and labelled as untrusted third-party content. Your agent should treat what ChatGPT said about you as evidence to read, not as instructions to follow.

Quotas, and why an agent cannot outspend you

A check run by your agent costs exactly what pressing Run now in the app costs, against the same monthly allowance. Creating prompts consumes the same tracked-prompt slots. Nothing about connecting an assistant creates a new billing surface.

The interesting design work was in the refusals. An agent that hits a limit and receives a bare failure will retry, and a retry loop against a metered operation is how people get surprised by a bill. So every refusal carries a code, a plain sentence the agent can relay to you, and the numbers behind it: which limit, what remains, when it resets. Partial success is reported rather than hidden. Ask for five prompts when three fit your plan, and you get three, plus a sentence explaining the other two.

There are per-call ceilings as well: twenty prompts per create, ten per check run, fifty per enable-or-disable. And a prompt with a check already in flight is skipped rather than checked twice, so a loop cannot double-bill the same question.

Try it

The server address is mcp.prominara.com. In Claude, add it as a custom connector. In Claude Code, one command does it. In Cursor and VS Code there is a one-click install link. ChatGPT connects it as a developer-mode connector.

Setup guides per client, the full tool reference, the security model and a troubleshooting path are in Connect your AI agent. The hub page, including the example prompts worth starting with, is at Bring your agent.

If you connect it and something is missing, tell me. The catalog is deliberately small, and the next tools should be the ones people actually reach for.