There are two ways to hand an agent your workspace. The remote connector runs hosted and connects with your RankControl login. The npm package runs on your machine with an API key. Both give any MCP client 80+ tools: visibility data, the content pipeline, repurposing, link building, brand memory, and reports.
Remote connector
The hosted server lives at https://api.rctrl.com/mcp. Nothing to install
and no API keys. You sign in and approve access in the browser. Connect as
the workspace owner. A workspace still in setup must finish onboarding
first.
Claude (claude.ai)
- Open Settings → Connectors.
- Click Add custom connector.
- Name the connector
RankControl, pastehttps://api.rctrl.com/mcp, and save. - Click Connect. Your browser opens rctrl.com: sign in, review the access list, and approve.
Claude can now use the tools in any chat.
Claude Code
claude mcp add --transport http rankcontrol https://api.rctrl.com/mcpClaude Code walks you through the same sign-in and approval in the browser.
What you approve
The consent screen lists the access before anything is granted:
- Read AI citations and visibility
- Read articles and the content calendar
- Read traffic and crawler analytics
- Plan and edit content
- Manage link outreach
- Manage social drafts
- Publish approved articles
- Manage workspace settings
Only the workspace owner can approve. To disconnect, remove the connector in Claude and the connector’s access stops.
Local server (npm)
The same package that powers the CLI is an MCP server. The local server runs on your machine with an API key and acts under that key’s scopes. Use the local server for CI, headless setups, or when you want key-level scope control.
Authorize once (rankcontrol login, see Getting started),
then register the server.
claude mcp add rankcontrol -- npx rankcontrol mcp{
"mcpServers": {
"rankcontrol": {
"command": "npx",
"args": ["rankcontrol", "mcp"]
}
}
}{
"mcpServers": {
"rankcontrol": {
"command": "npx",
"args": ["rankcontrol", "mcp"],
"env": { "RANKCONTROL_API_KEY": "rctrl_pk_..." }
}
}
}Prefer a skill?
Agents that run shell commands directly (Claude Code, OpenClaw, and any skills-compatible agent) can learn the CLI itself instead of connecting over MCP:
npx skills add rankcontrol/rankcontrolThe skill carries the full command reference plus the ground rules: dry-run
semantics, the review gates, and when not to add --confirm. Source at
github.com/rankcontrol/rankcontrol.
What agents can do
| Area | Tools |
|---|---|
| Visibility | score, trend, biggest movers, share of voice, cited sources and pages, sentiment (each by topic), citation checks, tracked queries, competitors’ other names and sites, crawler access |
| Content | list, plan titles, commit, generate, publish, title and meta, reschedule, archive, restore, citability report, internal links, an article’s questions before and since it went live |
| Reports | exec summary, top wins, agent activity, page engagement, traffic, funnel |
| Analytics | traffic and its trend over time, by page and by source, Search Console totals, top searches and clicks by day, rankings with Google AI Overview shown and cited per keyword, the biggest keyword movers, Google vs AI per tracked question, page engagement, data sources, Cloudflare connect |
| Action Items | what to work on next on your site, plan articles that link to a page, save a search for Plan Content, hide an item |
| Repurposing | queue, generate drafts, edit, Buffer and Postiz channels, push, mark posted |
| Link building | backlinks and stats, outreach pipeline, contact finding, reply drafting, queueing |
| Social | thread prospects, stats, status, reply drafting |
| Brand memory | read everything, update voice and style, manage products and buyer profiles, suggest audiences and draft a buyer profile to review |
| Team | list, invite, revoke, remove |
| Guides | search the blog’s guides on SEO and AI search |
| Support | contact_support files a report to the team when something breaks |
Example prompts that work end to end:
“Which of my tracked queries cite competitors but not us? Plan articles for the top three.”
“Update my brand tone to be more direct, then show me what the writers will see.”
“Summarize last month vs the month before and draft a LinkedIn post about the biggest win.”
Prompts
Both servers also come with ready-made prompts. Assistants that support MCP
prompts list them in their prompt menu. In Claude Code they’re slash
commands, such as /mcp__rankcontrol__weekly_summary.
| Prompt | What you get |
|---|---|
| Weekly AI visibility summary | How your AI visibility moved since last week, engine by engine, and what to work on next. |
| How each AI engine sees you | Each AI engine side by side: how often it names you, how often it cites your site, and where you’re weakest. |
| Where competitors beat you | Your competitors’ win rates next to yours, and the questions where AI answers name them and not you. |
| Pages to pitch | Pages AI answers cite for your tracked questions when they don’t name you, grouped by site, to ask for a mention or a link. |
Safety model
Agents get the same guardrails as humans, enforced server-side:
- Dry run by default. Actions with a cost or a public face (writing,
publishing, social push, outreach send, team invites) return a preview
until called with
confirm: true. Tool descriptions tell the agent to confirm what you asked for. - Removals wait for you. Removing, archiving, hiding and team changes always wait for your approval in RankControl. So does every change for 2 hours after the agent reads text written outside your workspace: prospect replies, AI answers, social threads, competitor pages. In those 2 hours, contact lookups wait only if your outreach mailbox has Autopilot on. The agent gets a link to pass on, and nothing changes until you approve.
- Granted access only. The remote connector acts under what you approved
on the consent screen. The local server acts under the API key’s scopes: a key
without
write:publishcannot publish, no matter what the agent asks. - Rate limits. Every endpoint carries a per-caller cap generous enough for real work and tight enough that a looping agent cannot run away with your usage. Costly tools sit at 5 to 10 calls per minute.
- Revocable. Every key is visible in Settings → API and dies the moment you revoke it. The remote connector stops when you remove the connector in Claude.