Can you build your own CRM with ChatGPT, Claude or n8n?

Yes, for a small, simple version. n8n can move data and call a model, and Claude or ChatGPT can read and write records through connectors. What breaks is everything a CRM quietly does: matching the same person across tools, avoiding duplicates, permissions, a trustworthy history, and honouring opt outs. Plan to build those yourself.

What can you build with n8n, Claude or ChatGPT?

More than people expect. A weekend gets you something that looks like a CRM. The building blocks are real:

• n8n is a workflow automation tool. It has a node for HubSpot that supports contacts, companies, deals, engagements and tickets, with operations such as create, update, search and get all (n8n HubSpot node docs). Its AI Agent node lets you "connect a chat model and one or more tools, and the agent decides which tools to call to complete a task" (n8n AI Agent node docs).
• Claude can connect to tools and data sources through custom connectors that use remote MCP, available on its Free, Pro, Max, Team and Enterprise plans, with Free limited to one custom connector (Claude help centre). Through the API, the MCP connector lets you "connect to remote MCP servers directly from the Messages API without a separate MCP client" (Claude API docs), and it is marked beta there.
• ChatGPT is an option too, but we could not open OpenAI's connector help page when we wrote this, so we do not describe its features or limits here. Check OpenAI's own documentation for the plans and rules that apply to you.

A typical home built stack looks like this: a form or a LinkedIn export feeds n8n, n8n writes rows to a database or a spreadsheet, a model drafts a message, and an assistant answers "who did I talk to last week?" by reading the table. That works. For ten contacts it works very well.

What does the simple version look like?

Each box is a decision, and each decision becomes a thing you maintain. The cost of the tools themselves is small. n8n's public pricing page listed a Starter plan at 20 euros a month and Pro at 50 euros a month on annual billing, with 2.5K and 10K workflow executions, and a free self hosted Community Edition on GitHub (n8n pricing, read 19 October 2026). The cost that matters is your time when something goes wrong.

What breaks first when you build your own CRM?

Six things, in roughly the order they show up.

1. Identity: who is this person?

A real contact arrives from a form as Maria Petrova, maria@acme.com, from LinkedIn as Maria P., Acme Corp, and from an import as m.petrova@acme-group.com. Is that one person, two or three? A CRM with identity rules decides consistently. A set of workflows decides differently each time, depending on which path the data took.

2. Deduplication

Duplicates are the first symptom. The second is worse: a wrong merge that puts two people's histories on one record. Both come from the same missing piece, a rule for what counts as the same contact, applied everywhere. With a model in the loop you also get nondeterminism, so the same input can produce different matches on different days.

3. Permissions

A spreadsheet has sharing settings. It does not have "this rep sees their accounts only", "this contractor cannot export", or "this field is read only after a deal closes". The moment two people use it, you need roles. When you add an assistant with a connector, the assistant becomes another user with whatever access its key has. Claude's own help page warns that connectors are not verified by Anthropic and advises connecting only to trusted servers and reviewing tool actions carefully (Claude help centre).

4. History you can trust

A CRM's history is the messages, in order, attached to the contact, including who sent them and when. In a home built system the history is spread across the mailbox, the messaging tool, the workflow logs and the table. n8n's own pricing page shows execution logs are retained for 7 days on Starter and 30 days on Pro, with a cap on saved executions (n8n pricing). Workflow logs are for debugging, not for being your record of what you told a customer.

5. Opt outs and compliance

This is where a hobby project becomes a liability. If you contact people, you must be able to stop. Under the GDPR a person has the right to object at any time to the use of their data for direct marketing, and once they object the data "shall no longer be processed for such purposes" (GDPR Article 21). In the United States, the FTC's CAN-SPAM guide says you must honour an opt out within 10 business days, that your opt out mechanism must keep working for at least 30 days after you send, and that each violating email can carry penalties of up to $53,088 (FTC CAN-SPAM guide). That is a number to respect, not a forecast of your risk. Neither rule is about AI. Both mean you need one suppression list that every workflow checks before it sends. Most DIY builds check it in one workflow and forget the others.

6. Who is responsible when the model is wrong?

A model that drafts an email is a tool. A model that sends it with no review is an actor. When a workflow lets the agent tool call its way to "send", you need an approval step, limits and a stop. You write all of those.

Can Claude or ChatGPT be the interface to a real CRM instead?

Yes, and this is the version that holds up. Keep the data and the rules in a system built for it. Use the assistant as the way you talk to it.

That is how we use it. Sellinger offers a hosted MCP connector with 43 tools for Claude and ChatGPT, and a REST API with webhooks from the Full Stack bundle up. In practice you can ask your assistant about your pipeline, review who is waiting, or launch a campaign in plain language, while Sellinger holds the contacts, the conversations and the sending rules. You can see the available MCP recipes, such as a weekly pipeline review or an approval queue triage, and the definition of MCP. The assistant is the interface. It is not the database.

When is building your own the right call?

Use this checklist.

• [ ] You have an unusual process that no CRM fits, and you can describe it in a page.
• [ ] You have an engineer who will own it for a year, not a weekend.
• [ ] You contact fewer than a few hundred people, so a mistake is cheap.
• [ ] You do not send messages on behalf of other companies or clients.
• [ ] You are happy to maintain the suppression list, the dedupe rule and the permission model yourself.

If you cannot tick most of these, buy the record and keep your effort for the offer and the audience. The same trade shows up on the outreach side, which we cover in What DIY outbound breaks first, and the broader question in Can AI replace Salesforce or your CRM?.

What should you build and what should you buy?

A fair split for a small team:

• Build the parts that are specific to you: your scoring rule, your qualification question, the report you read on Monday.
• Buy the parts that are the same for everyone and costly to get wrong: identity and dedupe, permissions, the conversation history, the suppression list and the sending limits.

That keeps what is yours in your hands and what is risky in a system whose job it is. If you want to see what that looks like, read about the Agentic CRM and what AI CRM means in practice.