This lesson is being recorded
Every video of this course is being uploaded to my YouTube channel. Subscribe so you do not miss it.
Bruno Ashford @brunoashfordSubscribeModule 6 · Lesson 6.5 · 12 min
Choosing Trustworthy Connectors
Assess the risk of any connector before it touches company data, and decide what it may read, draft or change.
Use ← → to move. The marker highlights text and the pencil draws on the slide; navigation locks while you annotate. Undo removes the last mark, Esc exits.
1Prefer the official source
First choice: a connector listed in your claude.ai settings or an MCP server published by the vendor itself on its own domain. Second choice: a well-known open-source server with an active maintainer. Avoid anonymous servers that ask for broad access to client data.
2Score it with the checklist
Use the Connector Risk Assessment Checklist from this module's materials. It asks who built it, what data it touches, whether it can write or send, where credentials are stored, and how you would switch it off. Anything scoring high risk needs a second person to approve.
3Ask Claude to do the first pass of due diligence
Claude can read the vendor's documentation, the permissions list and the tools the server exposes, and summarise the risk for you. You still make the decision.
I am considering connecting the MCP server at [URL or package name]. Research it: who publishes it, whether it is official from the vendor, what tools it exposes, which of those can write, send or delete, what data it can access, and how authentication works. Then fill in the Connector Risk Assessment Checklist in materials/06-connector-risk-assessment-checklist.md and give me a recommendation: connect, connect read-only, or do not connect.4Keep tokens out of shared files
If a server needs an API key, never paste it into a .mcp.json file that goes to the team or into version control. The project configuration supports environment variable references such as ${API_KEY}, so the secret stays on each person's machine.
5Know how to disconnect
Every connector needs an exit. Remove a terminal-added server with the remove command, and disable claude.ai connectors in your account settings or toggle them in /mcp. Also revoke the access token inside the vendor's own admin panel.
claude mcp remove hubspotHere is the uncomfortable part. When you connect a tool, you are handing a stranger's software a key to part of your business. Some MCP servers are built and maintained by the vendor you already trust. Others are weekend projects on a code-sharing site with three stars and no maintainer.
There is a second risk that surprises owners: prompt injection. If a connector reads outside content, like emails from strangers or web pages, that content can contain hidden instructions aimed at your agent. The official Claude Code documentation is explicit: only connect servers you trust, and be careful with servers that fetch external content.
Key points
- Official vendor servers and claude.ai connectors come first.
- Prompt injection is real: content from strangers can carry instructions.
- Score every connector before connecting; high risk needs two approvers.
- Secrets stay in environment variables, never in shared config.
- Every connector needs a documented way to switch it off.
Avoid
- Installing a server because it has an impressive README.
- Giving write access when the use case only needs to read.
- Pasting API keys into files that get shared or committed.
- Removing the server from Claude Code but leaving the token active in the vendor account.
A Shopify store owner found an unofficial MCP server that promised 'full store automation' and wanted admin access. The checklist flagged it immediately: anonymous author, write access to orders and refunds, token stored in plain text. She chose the platform's official route instead, connected read-only for reporting, and kept refunds as a human task. The whole assessment took 20 minutes and avoided handing refund power to unknown code.