All posts
Product9 min read

MCP server security review: 12 questions to ask a vendor, and our answers

For the security lead or the buyer who has to sign off on an MCP server: twelve questions in four blocks, the answer that closes each point, and DeployIt's own answer, as its security page gives it.

The DeployIt Team

We build DeployIt, an MCP server that lets a whole company ask its code from Claude and track each developer's work.

An MCP server security review comes down to four things, in this order: who can call the server, what it reads and keeps, what it can do, and what you can prove afterwards. Twelve questions cover them. Under each one, you will find the answer that closes the point, and DeployIt's own answer, as our security page gives it.

This grid is written for the security lead, the legal counsel or the buyer who can block the purchase. It applies to any vendor of an MCP server. Where our public pages say nothing, we say so.

In your Claude · example

Can a customer export their invoices to CSV?

Yes, from the billing page. The export covers the invoices of the selected period, one line per invoice.

Credit notes are not included.

Verified in the code4 files read · no line shown

This is the object under review: an answer about the product, given to a person with the Everyone profile, with a status and no line of code. The twelve questions ask what had to happen for it to exist.

Why an MCP server security review is not an API review

An MCP server gives an AI assistant a set of tools to call. The assistant, not a developer, decides at runtime what to ask. A classic integration review assumes every call was written in advance. Here nobody wrote them.

So the review asks less "what does this endpoint do" than "what can anyone make it return". If the term is new to you, start with what an MCP server is, in plain language.

What is under review here

DeployIt is an MCP server. Your company gets one link, https://mcp.deployit.ai/<company>/<code>; each person adds it to Claude (on the web, on the desktop or in Claude Code) and signs up with a name and an email, which is not verified; their questions are logged under that name. Everyone starts with the Everyone profile: questions about the product, answered from the code; a filter removes the code, the file paths and the names of team members it recognises from the answers, without proving their absence, and a leak test tells you what still gets through. An administrator gives the Developers profile (excerpts) or the Leadership profile (developer statistics) to whoever needs it. Repositories are connected from GitHub or GitLab; DeployIt makes no commit, branch or pull request, and only adds one webhook per repository.

A working copy of the code is kept while a repository is connected, hosted by DigitalOcean in London (UK); at indexing, each source file is sent to the model to be summarised, then the passages needed for an answer and the diffs of the commits that are scored; OpenAI's models by default, or the provider whose key you bring.

Block 1: who can call it

1. How does a caller authenticate? A specific answer closes the point: an access token on every request, and a refusal without one. "The address is hard to guess" leaves it open.

DeployIt: the company link carries a code, and that code only opens sign-up. Whoever adds the link to their assistant signs up themselves, on the Everyone profile. After signing up, each person's assistant presents a token that belongs to them alone, valid for an hour and renewed. See who gets in, and how, and the mechanism, step by step.

2. Does the server refuse a token issued for another service? The MCP authorization specification requires it: a server must only accept tokens issued for it, and must not pass a token on (revision 2026-07-28, checked on 17 November 2026). The point is closed when the vendor names that check.

DeployIt: our public pages do not describe this check. Ask us in writing at [email protected]: a person answers.

3. Can you cut one caller without touching the others? "Rotate the shared secret" leaves the point open.

DeployIt: removing a person cuts their access without touching anyone else's. A removed person who still has the link can sign up again, with the Everyone profile, until you change the code of the link. If the link has left the company, you change its code: nobody can sign up with the old one, and people already signed up keep their access. In an emergency, you can also sign everyone out at once.

Block 2: what it reads, keeps and sends

4. Which sources does it read, with which permission, and what does it write there? A list closes the point: each source, the permission requested, every write. "Access to your repositories" does not.

DeployIt: through the DeployIt GitHub application, or through an OAuth authorization on GitLab (gitlab.com), on the repositories you connect and no others. The webhook it adds to each repository is removed with the repository. On GitLab, the access requested (the api scope) would allow writes: DeployIt uses it only for that webhook. See where your code goes.

5. Does a credential to the source ever reach the caller? It should stay with the server.

DeployIt: signing up gives a person a token for DeployIt only. The connection to the repositories stays between DeployIt and your Git host.

6. Does a caller see a slice, or everything the server reads? A boundary held by the server closes the point; a boundary written "in the prompt" does not.

DeployIt: what an answer can show depends on the person's profile, not on their question: see profile by profile.

  • Everyone: no line, no file path, and developer names removed when DeployIt knows them.
  • Developers: excerpts and paths.
  • Leadership: no line, no file path, and the statistics of the team and of each developer.
  • Data tool: statistics only, no question about the code.

The filter removes what it recognises: code blocks and fragments, file paths, secrets, personal email addresses. It is filtered and tested, and no filter is infallible: how a leak test works shows what the test measures.

7. Where does the data live, who processes it, and is it used for training? Expect named companies, a location, and a written position on training.

DeployIt: the working copy is hosted by DigitalOcean in London (United Kingdom); it holds the content of each selected source file (secret and key files are excluded) and its summary, and you can also keep it in your own database. OpenAI reads the code by default; the provider whose key you bring does, for each role (agent, expert, mini) where you add it. We do not train or fine-tune any model on your code.

In transit, everything goes over HTTPS. Model keys, tool-server secrets and the address of your database are encrypted separately (AES-256-GCM). At rest, encryption is that of the database and the cache DigitalOcean manages, which its documentation gives as LUKS (checked on 17 November 2026). See the companies that handle your data and, for a DPO's reading, where your code goes with AI.

Block 3: what it can do

8. Can any tool change something? "Mostly reads" leaves the point open: one writing tool changes what a hijacked assistant can do.

DeployIt: five tools, server_info, ask_code, recent_changes, stats_schema and stats_query. None of them changes your repositories or your settings, and every call is written to your company's log. stats_query takes a structured query, never free SQL. A tool of your own company that you open to the agent acts with its own rights.

9. What happens when a text the server reads contains instructions? Source files and commit messages are text, and text can address a model. "The model is trained to ignore that" is not an answer.

DeployIt: our public pages make no claim on this point, and we do not claim to settle it alone: it also depends on the assistant that receives the answer. What they state is what such an exchange could reach: tools that change nothing in your repositories or settings, and an output that goes through the filter for the no-code profiles.

Block 4: what you can prove

10. Can you see who asked what? Ask what the log contains and how long it is kept, not whether logs exist "on request".

DeployIt: a log of who asked what, what was answered, and who changed which setting. It is kept 12 months by default; 3, 12 or 24 months, as you choose. The context attached to a question and the content the investigation read are not kept.

11. What does a removal leave behind? "Everything is deleted" leaves the point open until backups are mentioned.

DeployIt: for a removed repository or a deleted company, access is cut at once, then a queued job erases the data, usually within the following hours, and retries if it fails. A removed repository's statistics are erased only if you ask. The daily database backups hold the data until they rotate: DigitalOcean keeps them for seven days, according to its documentation (checked on 17 November 2026). The end of a subscription erases nothing: the company moves to the free plan.

12. What does the vendor not promise? A written list of limits is worth more than a roadmap told in the present tense.

DeployIt: three things, under what we don't promise.

  • That a model is never wrong: every answer says whether it is Verified, Backend only, Not found or Inconclusive.
  • That a shared link won't travel: anyone with the link can sign up; you see every sign-up, you can remove a person, and you can change the code.
  • That a filter or a leak test proves nothing can get out.

Two more facts belong here. Automatic sign-in through your Claude organization's identity provider is in beta, on request. And DeployIt claims no security certification.

How to score the answers

Do not add up a number. Three states are enough.

  • Settled: the answer is specific, and you can check it.
  • Open: the answer is plausible but cannot be checked yet, or the vendor's public pages are silent and you have asked in writing. Write down who will check it, and when.
  • Blocking: the vendor gives no answer when asked in writing, or its answer contradicts its own pages.

One blocking item in blocks 1 to 3 stops the purchase. An open item is a task with an owner and a date. By this grid, questions 2 and 9 stay open for DeployIt until you hold our written answer.

One commitment sits outside the grid: if there is a data breach, you are told within 72 hours at the latest. It is listed with the others under what we commit to.

The grid covers access and data. To judge what the answers themselves are worth, run the trial that evaluates an MCP server on its answers. For the reasoning behind this shape of access, read how to give AI access to your codebase safely; for the product as a whole, DeployIt: ask your code and follow each developer's work from Claude.

Next step: send the twelve questions to the vendor you are reviewing. For DeployIt, read the security page against the answers above, then add the demo, https://mcp.deployit.ai/demo, to your Claude connectors to see an answer and its status.

Price. Free plan: 15 questions a week for the whole company. A 30-day trial, no card. Plans from €290 a month excluding tax for up to 25 tracked developers: see pricing.

Ask your code from Claude

30-day trial, no card. Or try the demo first: add https://mcp.deployit.ai/demo to your Claude connectors.

Connect my code

Frequently asked questions

What should an MCP server security review cover?

Four things, in this order: who can call the server and how that access is cut, what the server reads, keeps and sends, what its tools can do beyond answering, and what evidence you get afterwards. Twelve questions cover them. For each, a specific answer you can check closes the point; a slogan leaves it open.

Is an MCP server riskier than a normal API?

Not by nature. The difference is the caller: an AI assistant decides at runtime what to ask, so the server must hold against requests nobody planned. A server whose tools write nothing, and whose output depends on the profile of the person asking, is the smallest version of that risk.

Does DeployIt have a security certification?

DeployIt claims no security certification. What it gives a reviewer is written on its security page: where the code goes, what gets out profile by profile, who gets in, its commitments, its limits and the companies that handle the data. If your policy requires a certificate, that is your policy to apply.

Where does an MCP server send my code?

Ask the vendor for three things in writing: where the code is kept, which model provider reads it, and whether it is used for training. DeployIt keeps a working copy hosted by DigitalOcean in London (United Kingdom) while the repository is connected; OpenAI reads the code by default, or the provider whose key you bring. DeployIt does not train or fine-tune any model on your code.

What happens to our data when we remove a repository or cancel?

Removing a repository from DeployIt cuts access at once, then a queued job erases its working copy, usually within the following hours; its statistics are erased only if you ask. The daily database backups, which DigitalOcean keeps for seven days, hold the data until they rotate. The end of a subscription erases nothing: the company moves to the free plan.

Continue reading

Connect my code