To give AI access to your codebase safely, do not give access to the code: give access to a place that answers questions about it. Four properties have to hold together: a connection that makes no commit, branch or pull request, a profile that decides what an answer may show, a token for each person, and access you can cut at any time.
This page is for the CTO or the security lead who has to say yes or no. A list of properties cannot be verified. What can be verified is what an access becomes between the day it is granted and the day it is withdrawn. So we follow one access across its life, then say what DeployIt states at each moment, with the section of our security page that carries it.
Does the CSV export accept ISO-formatted dates?
Yes. The export takes a start date and an end date in ISO format (year, month, day).
Any other format is refused, with a message that gives the expected one.
That is what a person with the Everyone profile receives: a sentence and a status. No line of code. The rest of this article is about what stands behind that sentence.
How to give AI access to your codebase safely: one access, five moments
The days below are illustrative: they mark the order of events, not a calendar. Each moment tests one property.
Day 0: the grant
Someone needs answers about your product, with an assistant. The meeting opens with "let's just give it the repository".
The decision is about scope, not trust. An assistant that answers questions never needs to write code. So ask what the server writes to your repositories, and with which permission: a write permission nobody uses is still a write permission.
Then comes the perimeter: who can sign up, and what a newcomer gets by default. A profile given by an administrator, never by the address someone typed, should decide what leaves.
Day 1: the first useful question
A support person asks the question above. There are two ways to answer it. One opens the files and hands them to the requester. The other returns a sentence.
In the first case the requester obtained source code; in the second, information. A support person who pastes a piece of your parser into a customer email has moved something out of the codebase, and nobody decided that.
Day 9: the question that must not be answered
Someone asks the assistant for the payment provider's API key, or for the content of a configuration file.
Here the property is a filter, and precision matters. A filter removes what it recognises, which is not everything. The right description is "filtered and tested; no filter is infallible", never "impossible". Ask which layers stand between the question and the file, and how you learn that one of them missed: how a leak test works covers that part.
Day 30: the review
Someone asks where an answer came from, and who asked. An answer that says what it rests on can be reviewed by a person who was not in the room on day 0. A log of who asked what does the same for the access itself.
Day 45: revocation
The contractor's engagement ends. This is when you find out what day 0 distributed. If each requester received a repository token, you now have tokens to hunt.
If each person holds their own token for the answer server only, you remove one person and nobody else notices. MCP authentication and revocable access gives the mechanism.
What DeployIt states at each moment
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.
Moment by moment, with the section of the security page where you can check it:
| Moment | What DeployIt states | Section |
|---|---|---|
| Day 0 | Only the repositories you connect. One webhook per repository, removed with the repository. On GitLab, the access requested (the api scope) would allow writes: DeployIt uses it only for that webhook. | Where your code goes |
| Day 1 | Everyone: no line, no file path. Developers: excerpts and paths. Leadership: no line, no file path, and the statistics. | Profile by profile |
| Day 9 | The filter removes what it recognises: code blocks and fragments, file paths, secrets, personal email addresses. A leak test attacks each no-code profile, on demand and after a settings change that affects the output (at most once an hour, 10 times a day), and tells you what gets out. | Profile by profile and What we commit to |
| Day 30 | Every answer says whether it is Verified, Backend only, Not found or Inconclusive. A log keeps who asked what, what was answered, and who changed which setting. | What we don't promise and What we commit to |
| Day 45 | One token per person, valid for an hour and renewed. 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, so you change the code of the link as well; in an emergency, you sign everyone out at once. | Who gets in, and how |
A data warehouse does not sign up: it presents a key, shown only once, that opens the statistics and nothing else. The product as a whole is described in DeployIt: ask your code and follow each developer's work from Claude.
What we do not promise is on the same page, under limits: that a model is never wrong, that a shared link won't travel, or that a filter or a leak test proves nothing can get out. The test catches the leaks it knows how to provoke. It warns you; it closes nothing for you.
What should the assistant refuse to answer?
Some questions are legitimate and still should not be answered for everyone: how the product defends itself, how internal prices are computed. In DeployIt, an administrator lists, profile by profile, the topics the agent will not discuss. When a question falls on one of them, the refusal is stated plainly to the person's assistant, which passes it on.
It is a setting, not a proof: test it with the questions you fear. For what the code does not contain at all (a roadmap, a promise made to a customer), read the status: only a Verified answer counts as a fact.
AI on a client's code under an NDA
Agencies and contractors meet a harder day 0: the code belongs to a client. In DeployIt, each client is its own company, with its repositories, its link, its profiles, its log and its plan. One account administers several, and nothing passes from one to another. A contractor signs up through that client's link, gets the Everyone profile, and is removed the day the engagement ends; change the code of the link that day too, since the contractor still has it.
One fact does not change. DeployIt keeps a working copy of the client's code, and a language model reads it. If the contract forbids sending the code to an AI service, that covers DeployIt too. Whether it is allowed is for the client to say, in writing; where the code goes, stop by stop gives you what to show them.
“We can just use GitHub's own MCP server”
It is a sound option for one case: a developer wiring their own assistant to repositories they already have rights to. GitHub's MCP server signs each user in with OAuth or a personal access token (its README, checked on 17 November 2026). So every requester needs a GitHub account with rights on the repository.
That works for five developers. It does not for support, sales or a client's contractor, and day 45 comes for each of them. The GitHub MCP server and what it answers walks through that route.
The second gap is the material. Repository access returns files; someone still has to read them and derive an answer. Claude's GitHub connector and what it cannot see shows where that difference is paid.
What this model does not solve
It says nothing about what happens next: a correct answer can be pasted into a message by the person who received it. It does not cover what lives outside the repositories, such as your contracts or your customers' data. And none of these properties is a certification: DeployIt claims none. They are design properties, to check one by one with the twelve questions of an MCP server security review.
Next step: add the demo, https://mcp.deployit.ai/demo, to your Claude connectors. Ask a product question, then ask for the code behind the answer, and see what the Everyone profile returns. Then check each row of the table on the security page.
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.
Frequently asked questions
How do you give AI access to your codebase safely?
Do not give access to the code: give access to a place that answers questions about it. Four properties have to hold together: a connection that makes no commit, branch or pull request, a profile that decides what an answer may show, a token for each person, and access you can cut at any time. Check each one on the vendor's security page before you connect a repository.
Is it safe to connect a private repository to an AI assistant?
It depends on what leaves, not only on what is read. With DeployIt, a working copy of the code is kept while the repository is connected, hosted by DigitalOcean in London (United Kingdom), and a language model reads it: OpenAI by default, or the provider whose key you bring. With the Everyone profile, a filter removes code, file paths and the names it recognises from the answers; it is filtered and tested, and no filter is infallible.
Can a contractor ask questions about our product without repository access?
A contractor signs up from Claude through your company link, with a name and an email, and gets the Everyone profile: answers about the product, from which a filter removes code and file paths; it is filtered and tested, and no filter is infallible. Signing up gives a token for DeployIt only, never a repository credential. Removing the person cuts their access without touching anyone else's. If they still have the link, they can sign up again, with the Everyone profile: changing the code of the link prevents it.
Can we use AI on a client's code under an NDA?
The setup narrows what is exposed: one isolated company per client, and people who receive answers instead of repository access. It does not change one fact: a working copy of the client's code is kept and a language model reads it. Whether the contract allows that is for the client to say.
Can we stop an AI assistant from answering on some topics?
In DeployIt, an administrator lists, profile by profile, the topics the agent will not discuss, for example security or internal pricing. A refusal is stated plainly to the person's assistant, which passes it on. Like the filter, it is a setting to test, not a proof.
Continue reading
How to connect Claude to a private GitHub repo without sharing the code
For a CTO: two ways to connect Claude to a private GitHub repo. Claude's own connector hands the files to whoever connects, with that person's GitHub access; one company link gives each person a profile and, with the Everyone profile, answers from which a filter removes the code (filtered and tested). The steps, in order.
MCP server security testing: how a leak test works, and how to read it
For a security lead or a CTO: most MCP server security testing checks who gets in. A leak test checks what gets out: trap questions fired at each profile that must not see code, a sorted report, and a warning that blocks nothing.
DeployIt: ask your code from Claude, and follow each developer's work
For the person who decides, CTO, CEO or head of product: the whole of DeployIt in one read. One link for your company's Claude, answers from the code without giving access to it, statistics on each developer, what happens to your code, and the price.