Model Context Protocol
Run the whole fix loop in one conversation.
Connect Claude, Claude Code or Cursor to Rovoki. Ask it to measure a brand, read what to fix, and re-measure after you ship. One thread, no dashboard.
What you can ask it
Plain questions, in the tool your team already works in. Each one maps to a real capability below, so nothing here is a promise the server cannot keep.
Where do we stand for our fintech client across ChatGPT, Perplexity, Gemini and Claude?
Start an audit for our brand against its three closest competitors.
What are the top three fixes from the last audit, and the predicted point gain on each?
Which buyer questions do we lose, who wins them, and in which language?
Where is the brand named in English but missing in Spanish?
Re-run last month's audit and show me what moved.
The fix loop, in one conversation
AI-visibility work is a loop: measure where you show up, find the questions you lose, change the site, measure again. In a dashboard that is four context switches. Through an agent it is one thread, so one operator can carry more clients without opening another console.
No new dashboard
Rovoki becomes a capability inside Claude, Claude Code or Cursor. Nobody learns another console.
No API keys
You sign in once with the same login you use for the web console. Your agent inherits exactly what your account can do, and nothing more. There is no key to copy, leak, or rotate.
The same guards as the app
Your clients, your role, the agency-and-client data split, all enforced on the agent exactly as they are on the console. An audit outside your scope is not found.
The re-measure is your proof
After a client ships a change, one call re-measures the same brand in the same tracking series. That before-and-after is the artifact you hand the client to defend the retainer.
What the agent can and cannot do
An agent connected to your account is powerful, so here are the boundaries, before you connect anything.
Least privilege, by construction
The tenant is read from the signed token, never from a value the model supplies. The agent can reach exactly the clients your account can, in exactly your role. It cannot widen its own access by asking.
Spend is capped by credits
Starting an audit spends from the same credit ledger as the console, which is the hard ceiling. Read tools are rate-limited; audit starts more tightly still.
Third-party output is quarantined
An audit returns real answers captured from other AI engines. Rovoki wraps that text in untrusted markers. A well-behaved agent treats anything inside them as data to report, never as instructions to follow.
Data handled like the console
The agent reads results that already live in your account; it creates no new data path of its own. What Rovoki stores, and for how long, is in the privacy policy.
Connect your agent
One server URL, one browser sign-in, no API keys. Pick your tool, paste, and sign in with the same login as the web console.
claude mcp add --transport http rovoki https://rovoki.com/api/mcpThen run /mcp, pick rovoki, and choose Authenticate. Adding the server alone does not open the sign-in.
Settings, Connectors, Add custom connector. Paste the URL, then press Connect to sign in.
https://rovoki.com/api/mcp{
"mcpServers": {
"rovoki": { "url": "https://rovoki.com/api/mcp" }
}
}Open Settings, MCP, and sign in when rovoki shows Needs login.
{
"servers": {
"rovoki": { "type": "http", "url": "https://rovoki.com/api/mcp" }
}
}Or the Command Palette, MCP: Add Server, HTTP. Authenticate from the CodeLens above the entry.
{
"mcpServers": {
"rovoki": { "serverUrl": "https://rovoki.com/api/mcp" }
}
}Refresh the MCP panel in Cascade, then sign in.
Turn on Developer mode, then Settings, Connectors, Create. Paste the URL, choose OAuth, and finish the sign-in. Needs a paid plan (Plus or Pro).
https://rovoki.com/api/mcpAdd a remote server over HTTP transport at the URL below and complete the browser sign-in. No API key.
https://rovoki.com/api/mcpWhichever tool you pick, after connecting ask it whoami. You should see your account, the clients in scope, and your remaining credits. That call is free; nothing spends until you start an audit.
Eight tools.
The read tools work for any role; starting or re-running an audit needs a member or admin seat.
| Tool | What it does |
|---|---|
| whoami | Your account, the clients you can audit, and your credit balance. Reads nothing that spends. |
| list_clients | The client companies in your scope. Where a client_id comes from. |
| start_audit | Start an audit for a client. Runs a few minutes, spends credits. |
| get_audit_status | Poll progress. Tells the agent how long to wait before re-checking. |
| get_report | Read a finished audit by section: scores, fixes, share of voice, placements. |
| get_answers | The raw engine answers behind the scores. |
| list_audits | Past audits across your clients. |
| rerun_audit | Re-measure the same subject after a fix, in the same tracking series. |
The shape of a run
A typical session moves in one direction, and the agent waits rather than loops:
whoami / list_clients which clients can I audit?
│
start_audit(client_id) → audit_id
│
get_audit_status poll until done (do not re-call start_audit)
│
get_report(section) scores, then the fixes
│
rerun_audit after a fix, re-measure the same subjectAudits run through real browser capture, so they take a few minutes and spend credits. A start_audit can come back started: false with needs_credits or waiting_for_browser. Those are normal states, not failures.
Common questions
What is MCP?
The Model Context Protocol is an open standard that lets an AI agent call outside tools mid-task. Connect once, and your agent can run audits and read reports as structured data, without you leaving the editor or pasting anything.
What does it cost?
An audit spends from the same credit ledger as the web console, and reads are free. There is no separate MCP charge. You need an account, and it is in private beta.
Which AI tools work?
Claude Code, Cursor, and Claude (claude.ai or Desktop) today. Any tool that implements the MCP spec will also work.
What can it access?
Exactly what your account can: your clients, your role, nothing wider. The agent reads results that already live in your account and creates no new data path of its own. Answers captured from other AI engines are wrapped in untrusted markers, so a well-behaved agent treats them as data, never instructions.
Point your agent at Rovoki.
Rovoki is in private beta. See how the measurement works, then ask us for access.