Customer.io MCP with Linkeddit in Claude: setup, scopes and competitor-driven campaigns

It gives Claude access to Customer.io's Journeys UI API and Pipelines API through eight tools. cio_prime loads the API reference, cio_schema finds endpoints, cio_read_api makes GET requests, cio_write_api makes POST, PUT and PATCH requests, and cio_delete_api makes DELETE requests. cio_skills_list and cio_skills_read expose task guides that Customer.io maintains, and cio_auth_status shows which workspaces the token can reach.

The directory listing says Claude can build automations, broadcasts and newsletters, create segments from plain English, inspect profiles, draft emails in Design Studio, analyze deliveries, opens and clicks, and manage data pipelines. What it cannot see is outside your workspace: what a competitor shipped this week, or which brand ChatGPT names when a buyer asks your category question. That is the gap Linkeddit fills.

Grant read only. Customer.io's docs say a connection gets only the read scope by default, which excludes sensitive personal data, and that you cannot grant the connection more access than your own login has. The write, write:live and configure scopes are independent of each other, so a task that edits content and sends messages needs both write and write:live.

Two account switches sit behind them. By default MCP cannot edit live data such as active automations or segments in use, and it cannot read attributes classified as sensitive. An admin turns on Allow MCP to edit live data or Allow MCP to access sensitive data in AI settings. Reads, writes and deletes also use three separate tools, so Claude's Connectors screen lets you allow reads automatically and ask before each write.

Add both connectors, then name the workspace in the chat. Customer.io is in Claude's Connectors directory; on a Team or Enterprise plan, an Owner of your Claude organization adds it for everyone first. If you cannot use the directory entry, add a custom connector with the URL https://mcp.customer.io/mcp. Choose Streamable HTTP, because Customer.io does not support the server-sent events transport.

EU accounts normally follow the redirect after sign-in; if your client does not, use https://mcp-eu.customer.io/mcp. For Claude Code, Customer.io also publishes a plugin. Then add Linkeddit at https://mcp.linkeddit.com/mcp, which comes with the one plan at $60 a month. Each person signs in with their own Customer.io login, and revoking a session happens in personal settings.

Claude reads the change from Linkeddit, then drafts the message in Customer.io. On Linkeddit's test project, plausible.io, the weekly Compete brief for the window 2026-10-26 to 2026-11-01 reported 5 competitors monitored, 2 high priority moves, 9 worth watching and 8 low signal. One high priority item was Fathom Analytics launching an AI assistant called Luna, with a source link to its changelog and a why it matters line.

That is the raw material for the three workflows on this page. Respond to a pricing or messaging change: Claude reads the move through compete_signals_get, creates a segment of affected customers in Customer.io, and drafts a one-time send. Weekly digest: Claude combines compete_brief_get and answer_runs_get into a newsletter draft for your team. Comparison content: Claude reads compete_signals_get and search_competitors_get and drafts a comparison email. Each saves a draft only.

A scope note for all three: creating drafts, segments and messages needs write. Sending to real recipients needs write:live, so leave it off until you have read the draft.

On Linkeddit's test project, the weekly brief returned the counts above, and compete_signals_get returned five signals for the week of 2026-10-26, all low priority: three Simple Analytics blog posts on privacy and cookie rules and two Umami roadmap posts. The competitor list showed 5 monitors, all weekly, with the next run on 2026-11-04. Compete had not yet analyzed plausible.io itself, its entry being marked pending.

answer_runs_get listed three audit runs on the same project, newest on 2026-10-28, and all three ended with status partial. A good draft says so instead of papering over it. This run was read-only: nothing was started, saved or spent, and the Customer.io side was not run.

Check three things. First, the workspace: Customer.io's docs say you must tell Claude which workspace to use by name or ID, so name it in the first line of the prompt. Second, the scopes you granted, compared with the job. Migma's operating guide makes the same point: a drafting task should not quietly become permission to send.

Third, treat text pulled from outside as evidence, not instruction. A competitor blog post or a customer message might contain wording that reads like a command. Tell Claude that only your prompt gives instructions. A competitive intelligence founder described the usual failure on Reddit: "Teams collect screenshots and links, put them into Notion, send a Slack update and move on. Nobody clearly owns the response." Naming the reviewer in the prompt fixes that.