CrunchJunkie MCP

npx add-mcp 'https://app.crunchjunkie.io/api/mcp/mcp'

Let AI assistants read your CrunchJunkie data securely over the Model Context Protocol — and, if you switch it on, act on it.

What the MCP connector does

CrunchJunkie ships a Model Context Protocol (MCP) connector, so AI assistants can read your workspace data directly in a chat — your clients, connected data sources, reports and report data, and AI-visibility metrics. You can ask "summarise last month's Google Ads report for Acme" or "which prompts is Acme invisible for in ChatGPT?" and the assistant pulls the answer from your live CrunchJunkie data instead of you exporting anything. It's one server, not a set of per-vendor integrations. The same connector URL works with claude.ai, with ChatGPT (including Deep Research, which is what the connector's built-in search and fetch tools exist for) and with any other MCP-compatible client — same tools, same scopes, same safeguards, whichever one you use. Two things are true by design. It's scoped to your team: an assistant only ever sees the workspace of the account that signed in. And reading never changes anything — the read tools fetch and search, full stop. The connector can also act for you — draft a report, add or remove a widget, schedule a report, track or stop tracking an AI-visibility prompt, run a fresh GEO audit — but only after a workspace owner switches that on, and even then every change is preview-and-confirm (see "Letting an assistant take actions" below). Billing, plan and account settings are out of reach entirely. The connector is available on paid plans.

Connecting an assistant

Open Settings → Connect AI assistants (MCP) to find your connector URL (it looks like https://app.crunchjunkie.io/api/mcp/mcp). Add that URL as a remote MCP server in your client: • Claude (claude.ai or desktop) — Settings → Connectors → Add custom connector → paste the URL. • ChatGPT — Settings → Connectors → Add (available on Pro/Team/Enterprise), or add it as a source in Deep Research → paste the URL. • Anything else that speaks MCP — add it wherever that client takes a remote MCP server URL. There's nothing vendor-specific to configure. One ChatGPT-specific note: in a normal ChatGPT conversation, only the read tools are visible — ChatGPT hides write-capable tools unless Developer Mode is enabled (Settings → Apps & Connectors → Advanced → Developer mode; on Business workspaces an admin may need to switch it on). That's ChatGPT's policy, not a connector limitation: the same connector exposes all 45 tools in Claude. Write actions additionally require the workspace opt-in below, an account role that's allowed to edit, and confirm every change with a preview first — four independent switches, so nothing can write by surprise. When you add it, the client sends you to a CrunchJunkie sign-in screen to authorise access — the same OAuth flow everywhere. You log in exactly as you do to the app, approve the access, and you're connected — no API keys to copy or manage. Because authorisation rides on your normal login, the assistant acts as you and sees only what your account can see. You can revoke access at any time from the client's connector settings. Five scopes can be requested, and you can see which were granted on the consent screen: • read:reports — list and open your reports and pull their data • read:visibility — read AI-visibility and product shopping-visibility metrics • read:clients — list your clients and their connected data sources • write:reports — create a report, add, edit or remove widgets, schedule a report • write:visibility — track prompts, manage competitors, tags, scan and project settings, add timeline notes, run a GEO audit, start scans and generate content briefs Every connection currently receives all five scopes when you sign in, so a scope alone never decides whether an assistant can change something — the workspace setting, your role and the confirmation step do (see the four gates below).

What the assistant can read

The read tools are: list your clients, list a client's connected data sources, list and open reports, pull a report's data, read AI-visibility metrics, read product-level shopping visibility (recommendation rate, head-to-head win rate and competitor products), and a general search/fetch over your workspace. A deeper analysis layer reads the per-prompt data underneath those metrics: list your tracked prompts with pooled per-prompt visibility, drill into one prompt's per-engine results and the competitors the AI actually named, see which domains the engines cite (with genuinely-cited and retrieved-only counted separately), find content gaps — prompts where your brand is absent from every recent answer while competitors are named, the same definition the Content Brief generator uses — read your Signals inbox, pull the latest GEO Audit score with its findings, read Brand Perception — the terms AI actually uses about each brand, with any stored fact-check results — see the query fan-outs (the web searches engines ran before answering your prompts), and read the AI-crawler analytics from your forwarded server logs. It can also read the AI answers themselves — the list of answers behind every visibility figure, and any one answer in full with the pages it cited and the searches the engine ran — follow visibility day by day with the number of answers behind each day, drill into one cited page to see which prompts and engines used it, tie visibility to the sessions and conversions AI assistants send via GA4, open saved content briefs, and check the workspace's Managed-AI scan allowance. Reading is always available on a paid plan and never alters your data — there is no billing access, and no ability to change your plan, team or settings from a chat. Answers, cited pages and engine searches are text written by third parties, so every result that contains them says so before the data and tells the assistant to treat that text as data, never as instructions. The connector also ships six ready-made prompts — a weekly client review, a competitor teardown, a content-gap plan, a visibility-drop investigation, a GEO audit triage and an outreach target list (the third-party pages AI engines read on your buyer questions, with a baseline to re-measure against) — which appear in assistants that support MCP prompts. If a tool ever returns "missing scope" or "forbidden", it usually means the signed-in account isn't on a paid plan or isn't a member of the team that owns the data. Sign in with an account that has access to the workspace you're asking about.

Letting an assistant take actions (optional)

Reading your data is only half of what an assistant can do. If you want it to also make changes for you — "draft a Google Ads report for Acme", "schedule Acme's report every Monday at 9am", "start tracking the prompt 'best CRM for agencies' for Acme", or "add a conversions-by-device table to that report" — the connector has 18 action tools: create_report, add_widget, edit_widget, remove_widget and schedule_report (covered by write:reports), and create_client, add_tracking_prompt, remove_tracking_prompt, update_tracking_prompt, manage_competitors, set_prompt_tags, manage_project_tags, update_scan_settings, update_project_settings, add_annotation, run_geo_audit, run_scan and create_content_brief (covered by write:visibility). create_client sets up a new client and its AI-visibility project in one step — brand name, product type, Brand Profile, competitors, market, language and engines — so an assistant can onboard a brand end to end and then add prompts. run_geo_audit runs a fresh GEO Audit of the client's own website — the same audit as the GEO Audit tab, at most once per project every 15 minutes, without using your scan allowance. Two actions spend AI: run_scan starts an AI-visibility scan and create_content_brief writes a content brief. They need one more switch, "Allow actions that spend AI" under Settings → Connect AI assistants (MCP), and their preview shows the cost before you confirm — which engines run on your own keys, how many weighted checks count against your Managed-AI allowance and how many would be overage. A workspace can start up to 20 of them a day from assistants. A scan runs in the background (get_scan_status follows it), and the overage cap and fair-use ceiling apply exactly as for a scan started in the app. Every one of them is gated four times over, and all four must hold: 1. Scope — the token must carry the matching write scope. 2. Workspace setting — "Allow write actions" must be on under Settings → Connect AI assistants (MCP). It's off by default and only an owner or admin can change it. While it's off, the connector reads and nothing more, no matter what scopes a token holds. 3. Your role — the connector acts as the account that signed in, so it follows your team role. A member with the "Read only" role can read everything through the connector, but every action tool refuses to make changes for them — exactly as in the app. 4. Confirmation — the tool has to be called with confirm: true. Called without it, it returns a preview of exactly what it would create — the report's widgets, the schedule's cadence and recipients, the prompt text — and changes nothing. Content the connector creates follows your workspace language. If your account language is German, a report drafted through the connector gets German section titles rather than English ones. (That's about the text CrunchJunkie writes; your assistant already replied in whatever language you asked in.) Every confirmed action is recorded — who made it, through which connection, which tool and what it targeted, never the text you sent — and owners and admins can see the recent ones under Settings → Connect AI assistants (MCP). Records are kept for 12 months. Actions stay scoped to your team and never touch billing or account settings. You can turn actions back off at any time and the read tools keep working. It's the same grounded, catalog-checked logic as "Draft with AI" and "Ask Crunch to build a widget" in the app — so an assistant can only build things CrunchJunkie actually supports, never invent a metric or a data source.