4 min readNovember 19, 2026

AI Agents Connected to My LinkedIn Tool Mostly Read. Here's What I Built Around That

I put an MCP server in front of my LinkedIn writing tool and counted what the agents did with it. 85% of their tool calls were reads, not writing. Here is what I built, what I changed because of it, and what I still don't know.

AI agents came to my LinkedIn tool to read, not to write: 85% of tool calls were reads
AI agents came to my LinkedIn tool to read, not to write: 85% of tool calls were reads

I run ContentIn, a LinkedIn writing tool, and since July it has had an MCP server: a URL you add to Claude, ChatGPT, n8n and similar clients so an agent can use the product's tools. In October I counted what the agents actually did with it.

85% of their tool calls were reads: the person's voice profile, their past posts, their analytics.

Before I counted, I expected most people to use the server to write posts, more than to build more complex automations around it. The reads changed how I think about what I'm selling, so here is the longer version.

Why a LinkedIn writing tool built an MCP server

In our feedback and cancellation notes, whenever people compared us to another product, they named ChatGPT or Claude 31 times and another LinkedIn tool 3 times. The competitor they had in mind was the chat window they already pay for.

For most of this year my answer was to make our own writing better, so nobody would need the chat window. The MCP server started as a side door next to that: if people spend their day in Claude anyway, let Claude reach their account.

What I built

13 tools today. Two design decisions took most of the time.

Read-only voice context. One tool returns the person's writing guide, their style rules in priority order, their own stories, numbers and frameworks, their best posts and their strategy. It changes nothing. Any model can call it before writing and then write the post itself. We also have our own writing tool, and it has one hard rule: if a detail is missing, it returns a question instead of a post. It does not invent a client story to fill the gap.

A confirm step the agent can't skip. Scheduling and publishing take two calls. The first returns the exact text and a one-time token that expires after five minutes and is bound to that text. The second call needs the token. A post can only go out after a person has seen the words. It also covers prompt injection: an agent that read a web page with hidden instructions still can't post on its own.

The token alone doesn't stop an agent from approving its own post, though. If the same agent holds the scheduling tool, it can read the token from its own result and send it straight back. So in the n8n workflow I built on top of it, the agent only gets the read-only voice tool. Writing and scheduling run as plain workflow steps, and a Slack Approve button sits between the preview and the confirm.

Slack approvals take longer than five minutes, so after the click the workflow asks for a fresh preview, compares it byte for byte with the text that was approved, and only then sends the new token. If the draft changed in between, nothing goes out. Structure does that job better than a line in a prompt.

Everything posts through LinkedIn's official API. The agent never sees a password or a session cookie, and one sign-in is bound to one LinkedIn profile.

What the usage told me

Agents can already write. What they can't produce is the person: how they write, what they've already said, which of their posts worked, who engaged with them. That's what 85% of the calls went to. The prose is the part any model can do; the history is the part only we hold.

What I changed

On 10 November the agent version went live as its own product line: no web app, your AI is the interface. You sign up, paste your LinkedIn URL, see the voice profile your AI will work from, and connect Claude or ChatGPT with one URL. There's a small console for status, usage, the queue and billing.

Two days later I reversed a pricing call. At launch, connecting and seeing your voice profile was free and every tool needed a paid plan. Now posting from your AI is free on the agent line: it can schedule and publish up to 2 posts a day and 30 a month per LinkedIn profile, up to 14 days ahead, and you confirm each one. The web app stays paid. What costs money on the agent line is the data and our writing: analytics, past posts, saved material, the full voice context and our own writing tools. $29 a month per profile, or $49 with the people who engaged with your posts scored against your ICP. The full split is on the ContentIn for Agents page.

The reasoning goes straight back to the 85%. If people write with their own AI and come to us for posting and for their data, charging for posting puts the paywall on the part anyone can copy. The data is the part an agent can't get anywhere else.

What I don't know

Whether anyone pays for this. People pay for a writing tool because it writes. Paying for "your AI, plus your own history" is a different ask, and I have no proof yet that it converts. The new page has a fixed trial target for its first month and I'm watching it daily. I don't know if it will hit.

Whether free posting gets abused. Every customer posts through the same LinkedIn app, so spam from free accounts would put the paying ones at risk. The limits are per profile and every post still needs a human yes. I'll be watching that closely too.

Whether the five-minute token plus a re-preview is the right trade. The alternative is binding the token to the approval event itself, which is cleaner but means the server has to know about Slack, n8n and every other place a human might click Approve.

Try it

Posting from Claude or ChatGPT is free, with no card. Start on the ContentIn for Agents page, or follow the setup guide. The step-by-step version is in How do I connect my AI agent to LinkedIn?

Create Engaging LinkedIn Content

Use ContentIn's AI Ghostwriter to write posts that resonate with your audience and build your personal brand effortlessly.

Start Your Free Trial