Close CRM MCP with Claude: add competitor moves and AI answer gaps to your sales records

Close is the system of record. Its Agent Index listing shows 55 tools: 22 reads, 29 writes and 4 the listing does not classify. Close's developer docs name reads such as list_objects, aggregate_objects, find_opportunities, activity_search and fetch_call, which returns a call's outcome and, where available, its transcript. Close's own blog post describes the same split: Claude pulls CRM data, updates records and creates leads, and can be combined with other MCP tools for multi-source work.

What Close cannot show is the outside view. Linkeddit supplies it. Compete watches up to 12 competitors weekly and returns each move with a priority, category, source link and date, with reviews from G2, Capterra, TrustRadius and Trustpilot stored separately. Answer Radar asks the five engines your buying questions and records who is named, recommended and cited. Both come with one plan at $60 a month.

Claude does the thinking. Linkeddit sends no email and writes nothing to Close, and the Close connector has no view of competitors. A B2B SaaS founder on Reddit named the problem sales teams face: "How do you get someone to entertain switching when they're not actively unhappy?" Competitor moves and AI answers are evidence for that conversation, and Close is where the conversation is tracked.

Close's help center says its MCP server supports Streamable HTTP and OAuth 2.0 Dynamic Client Registration, and recommends OAuth through a supported client. On a paid Claude plan you add it under Settings, Connectors, Organization connectors. On the Claude desktop app's free plan you add it from the chat window through Connectors. In Claude Code you pass your Close API key and a scope header.

The scope decides what Claude can do. Per Close's docs, mcp.read is read-only, mcp.write_safe can read and write, and mcp.write_destructive can also delete. Each higher scope includes the lower ones. Every workflow on this page works with mcp.write_safe, which covers create_note, create_task, create_comment and create_draft_email. Nothing here needs delete, so do not grant it.

Then add the Linkeddit connector at https://mcp.linkeddit.com/mcp. Linkeddit is a connector in Claude, and a CRM scope never gives it any access to your Close data.

Ask Claude to read your latest Answer Radar results from Linkeddit and save one note per week in Close. Linkeddit returns, per engine, presence, recommendation, share of voice, mean position, sentiment and the domains cited. On Linkeddit's test project, plausible.io, the latest run measured five of six engines: presence was 8 of 9 answers on ChatGPT, 8 of 9 on Gemini, 7 of 9 on Google AI Mode, 6 of 6 on Google AI Overviews and 5 of 9 on Claude. Perplexity returned errors on all 10 samples and was reported as unmeasured, not as zero.

That honesty is the point of the note. Sales reps get one line they can trust: where the brand is named and where a competitor is named instead. In the Claude result, Plausible held 17 percent of vendor visibility and Fathom and Matomo 12 percent each. Close's create_note attaches a text note to a lead, so choose a lead to hold the report, for example an internal record you create for market notes.

Say what the numbers are. Linkeddit measures through official provider APIs with grounded web search and calls the result a controlled proxy, not an exact copy of what each buyer sees. Put that sentence in the note.

Filter to high priority and act only on those. In the week of 26 October 2026, Linkeddit's high priority list on the test project held 2 moves: Fathom launching its Luna AI assistant and Matomo publishing critical PHP security guidance. Each came with a source link and a reason it matters. A rule that creates a Close task only for high priority moves would have created 2 tasks that week and left the rest in Linkeddit.

Close's create_task adds a to-do for a lead, create_comment adds a comment to a note, call or opportunity, and create_note adds a note. Close also lists create_workflow, but it creates a workflow with Draft status, and nothing in the tool list says Claude can publish or start one. Treat the workflow as a draft you review and switch on yourself.

Claude can also find where the move matters. Using find_opportunities, it can list the open opportunities whose lead names that competitor, and attach the task there. Check the matches before approving, because a name in a note is not proof the buyer is comparing you.

A battlecard goes stale when nobody owns the refresh. A competitive intelligence founder on Reddit put it this way about a related problem: "The problem was never a lack of information. It was a lack of decision ownership." The fix is a short, dated card with one owner, built from sources you can open.

Claude builds it from three places. Linkeddit supplies the competitor's recent moves and the stored reviews, including low-rated ones through its max_rating filter. Close supplies your own record of lost and open opportunities through find_opportunities and the call outcomes through fetch_call. Claude writes the card, marks each line as a competitor claim, a review quote or something from your own calls, and saves it with create_note on the relevant lead.

Use create_draft_email when the card should reach a rep. Close's docs say the draft is saved unsent for the user to review, edit and send from Close, and is never sent automatically. That keeps a person between a model's summary and a customer's inbox.

When an engine recommends a competitor, it usually cites pages. Linkeddit lists your open gaps and the sources cited behind each one. On the test project it showed 1 open gap, "Umami vs Matomo for self-hosted analytics", with an evidence score of 22, and the run's citation share named domains such as analytics-alternatives.com and bootstrap.build.

Ask Claude to take the cited domains for your top gaps and, for each publisher or partner you already track in Close, add a task with the cited page link. If the contact is on file, create_draft_email saves a short unsent draft you can read and edit. The outreach is yours; Linkeddit only shows what the engines cited, and Claude should not invent contacts that are not in Close.

Check that each page is somewhere a correction or inclusion is something the publisher would reasonably want. A cited page that describes a competitor accurately is not a target.

Start at mcp.read, run one workflow, then raise the scope. Close's docs describe the scope as part of the connection, so a person who only needs to read can be given a read-only connection. Scalekit's comparison of the Close MCP server and the Close API describes the scope split as finer than the API's, and the search listing for it is dated 23 September 2026.

Use a test lead before a real one. Look at the first note Claude proposes: it should name its source and date, and say which numbers are measured. If it reads like a verdict, ask for the evidence again. On the test project, Linkeddit's brief tools returned an error for the week and the competitor brief for Fathom came back empty, and a good note says that instead of smoothing it over.

Count what changed in Close, not what was written. Track how many high priority moves became a task, how many of those tasks were done in the week, and how many notes a rep opened before a call. In the test project's latest run, five of six engines were measurable, so the number to watch is the engines where a competitor is named instead of you, week over week.

Common mistakes are granting delete scope for a read job, writing a competitor claim without a source link, and treating one week's answers as a trend. That run took one sample per question, so look for a gap that repeats across runs before you change a sales message.