Free checklist · 8 pages · Build
Your MCP server is a new door into every customer's data.
Securing your MCP server before customers connect it
What can go wrong when ChatGPT and Claude call our product through an MCP server, and what stops it?
- For CTOs, VPs of Engineering and Heads of AI at U.S. B2B SaaS companies opening their product to ChatGPT and Claude through an MCP server
- Proposed approach · Checklist · 8 pages, PDF
What is inside
Know the five threats to an MCP server, the control the specification sets for each, and the tests to pass before the first customer connects.
- Juniper Dispatch: five fixes before launchA filled example
- Five threats, and the control for eachStep 1 · Know the five threats
- Review every tool on one sheetStep 2 · Review every tool
- Pass these before the first customer connectsStep 3 · Before launch
- Permissions live in code, never in the promptWhere we have done this
- Bring this page to the Product ReviewYour notes for the call
This answers
Which attacks reach an MCP server, and what stops each one?
Five threats reach an MCP server: token passthrough, the confused deputy, tool poisoning and injection, over-broad scopes, and state handles and consent. Check the token audience on every call and give your API a token of its own, require consent per client, return free text as data, grant read scopes first, and never treat a state handle as authentication.
How do we review a tool before ChatGPT and Claude can call it?
Put every tool on one review sheet: one user goal, typed and bounded inputs with no customer ID, user and customer from the token, approved output fields with free text quoted and capped, and writes in their own tool with a scope step-up and user confirmation. A tool that fails one row waits.
What must pass before the first customer connects?
A token issued for another app is refused, your API never sees the token the client sent, customer A cannot reach customer B, a planted instruction changes nothing, and revoking a user, tool or customer takes effect on the next call. Run every test on each build.
You also get
- A filled security review of a fictional connector: five findings and their fixes
- The five threats to an MCP server, with the control for each in the specification's own words
- A review sheet every tool must pass before ChatGPT and Claude can call it
- The tests to pass before the first customer connects
Free preview
Read the first 3 of 8 pages now
Free to read here, no form needed. The form opens the other 5 pages and the PDF.
Page 1 of 8Cover 
Page 2 of 8Juniper Dispatch: five fixes before launch 
Page 3 of 8Five threats, and the control for each 
5 more pages after the form
A review sheet for every tool, the tests to pass before launch, how we keep permissions in code and a page to bring to the call.
- Page 4Review every tool on one sheet
- Page 5Pass these before the first customer connects
- Page 6Permissions live in code, never in the prompt
- Page 7Bring this page to the Product Review
- Page 8Walk your riskiest MCP tool through with us. Book a free 30-minute Product Review.
Related guides
More for the Build stage
Blueprint · 8 pages
How can an AI agent act in our product without paying twice, looping or being talked into harm?
Guide: Before your AI agent touches money: limits, approvals, undo
Get the guideBlueprint · 8 pages
Property ManagementHow do we let AI book, assign and draft in our property platform without anyone acting outside their role?
Guide: Role-safe AI for rental operations software
Get the guideBlueprint · 8 pages
FintechHow do we add AI to a regulated financial platform without touching the core, or letting the AI decide?
Guide: AI beside a regulated core: documents in, people decide
Get the guide
Your customers already expect AI from your product.
Let us make it real inside the product you run today.
