Security
Designed so a security review is short
The safest data is data we never receive. Zevqi is built around that idea, and we are explicit about the places where the design has limits.
Data flow
What goes where
Stays on the developer machine
- •Source code and diffs
- •Quiz questions and answers
- •The written explanation from a review
- •Your model API key
Sent to your model provider
- •The diff being explained, using the developer’s own key
- •Only if a key is set. Nothing is proxied through Zevqi
Synced to your team (if you join one)
- •Only what is listed in full below
- •File paths, results and pseudonymous ids
- •Never code, diffs or commit messages
The full list
Exactly what Zevqi receives and stores
This is every kind of data that reaches your team’s dashboard. It matches what is stored in our database, so a security review can check it line by line.
Sent for every developer on the team
- •Email, and a name if they give one
- •The repository name and the paths of the files each lesson touches
- •Whether each lesson is waiting, passed or failed, and the quiz result (correct out of total). Managers do not see a score per person
- •A content fingerprint (hash) of each file, so Zevqi knows when a file changed after it was reviewed
- •Their answer after a quiz: understood, needs clarification, or investigate further
- •Timestamps
Sent by the git scan (zevqi scan)
- •File paths, commit counts, and the first and last commit date of each file
- •How many commits were marked AI-assisted
- •Each author as a pseudonymous id (a one-way hash of their email), with their commit counts. No email, no name unless names are shared
Sent only if the developer turns sharing on
- •A short lesson summary, up to 900 characters
- •Open questions the developer raised, up to 5 of 300 characters each
Never sent
- •Source code, diffs and commit messages
- •Quiz questions and the text of answers
- •The explanation a developer writes in zevqi review
- •Your model API key
Check it yourself before anything is sent
The setup command prints what it will send and where, before it runs. To see a scan’s exact upload without sending it:
zevqi scan --dry-runThe command-line tool is open source (MIT), so you can read what it does: github.com/zevqi/zevqi.
Delete it any time
The team owner can delete the team and everything stored for it (lessons, scans, keys and feedback) from Settings in the dashboard. Payment records are kept separately from the team’s data. Developers can also leave a team with zevqi team leave, which stops syncing from that machine. Data already synced stays until the owner deletes the team.
Questions about your data? Email hello@zevqi.com.
Controls
How we protect it
Source code stays local
Code is read on the developer machine. With no model key it goes nowhere. If a developer adds their own key, the diff goes only to the provider they chose, under their own account, and Zevqi says so first. It never reaches us.
Minimal sync
A team record contains a file path, a content fingerprint (a git blob hash, so we can tell when a file changed), a score, the developer’s own answer, a lesson id, a status and timestamps. Summaries and notes are opt-in.
Secrets are hashed
Team keys and dashboard tokens are generated with 192 bits of randomness and stored only as SHA-256 hashes.
Locked-down storage
Team data lives in Postgres with row-level security enabled and no public policies. Only our server can read it.
No telemetry in the CLI
The open-source CLI makes no analytics calls. Network traffic is limited to your model provider and, if you opt in, your team server.
Passwordless sign-in
Owners sign in with a one-time email code that is stored only as a hash, expires in 10 minutes, works once and locks after five wrong guesses. Sessions are signed, httpOnly cookies, and admin actions also check the request origin.
Told before it sends
Before anything is installed, Zevqi states what goes to your model provider (the diff of changed files, new files in full) and that nothing else leaves your machine. You confirm, or stop.
Fail-safe hooks
Every hook exits with success even on error, so Zevqi can never block or break your coding agent.
Your choice
What each way of using Zevqi shares
You can use Zevqi with no AI at all. This is exactly what leaves your machine for each way, and which ones exist today.
| Way | What leaves your machine | What it needs from you | Status |
|---|---|---|---|
Git knowledge map zevqi scan reads your git history: who changed which files, when, and whether a commit says an AI helped. | File paths, commit counts, dates and pseudonymous author ids. Names only if you add --names. Never code, commit messages or emails. | One command. No model, no key. | Available |
Attested review zevqi review <file>: you read the real change and describe it in your own words. Not graded. | The path, a content fingerprint and your own call (understood or not yet). Your words stay on your machine. | One or two minutes. No model, no key. | Available |
Pull-request check zevqi check in CI lists changed files whose AI-written changes nobody has reviewed. | The changed file paths and the repository name. | A team key stored as a CI secret. | Available |
Written by your coding agent A block in AGENTS.md makes Codex, Cursor or Claude Code write the lesson itself and hand it to Zevqi. | File paths, fingerprints and scores. No extra model call and no key. | zevqi init adds the block. The agent has to follow it. | Early access Implemented and tested on our side. Not yet verified in live Codex or Cursor. |
Claude Code hooks Lessons are prepared in the background after each agent turn. | File paths, fingerprints and scores. With a model key, the diff goes to your provider. | zevqi init, and a model key for the questions. | Available |
Your own model key Zevqi writes questions that are specific to the diff. | The diff of the changed files goes to your provider under your own account. Never to us. Zevqi tells you before it installs anything. | An API key you provide and pay for. | Available Optional. Most of Zevqi works without it. |
Local model AI-written questions from a model running on your own machine. | Nothing leaves your machine. | Install and run a model. | Planned |
GitHub app Show who approved each change, from pull-request reviews. | Pull-request metadata. No code. | An organization admin installs the app. | Planned |
Cursor and Codex hooks Capture edits directly in each tool. | File paths. | A per-tool install. | Planned |
Lessons written by Zevqi We would read your diffs and write the questions. | Your code would reach us. | Nothing. | Not offered By design: your code never reaches us. |
Using your Claude subscription login Running Claude in the background on your behalf. | The diff would go to Anthropic through your login. | Nothing. | Not offered The provider asks products to use API keys, so we do not do this. |
- Git knowledge mapAvailable
zevqi scan reads your git history: who changed which files, when, and whether a commit says an AI helped.
- Leaves your machine
- File paths, commit counts, dates and pseudonymous author ids. Names only if you add --names. Never code, commit messages or emails.
- Needs from you
- One command. No model, no key.
- Attested reviewAvailable
zevqi review <file>: you read the real change and describe it in your own words. Not graded.
- Leaves your machine
- The path, a content fingerprint and your own call (understood or not yet). Your words stay on your machine.
- Needs from you
- One or two minutes. No model, no key.
- Pull-request checkAvailable
zevqi check in CI lists changed files whose AI-written changes nobody has reviewed.
- Leaves your machine
- The changed file paths and the repository name.
- Needs from you
- A team key stored as a CI secret.
- Written by your coding agentEarly access
A block in AGENTS.md makes Codex, Cursor or Claude Code write the lesson itself and hand it to Zevqi.
- Leaves your machine
- File paths, fingerprints and scores. No extra model call and no key.
- Needs from you
- zevqi init adds the block. The agent has to follow it.
Implemented and tested on our side. Not yet verified in live Codex or Cursor.
- Claude Code hooksAvailable
Lessons are prepared in the background after each agent turn.
- Leaves your machine
- File paths, fingerprints and scores. With a model key, the diff goes to your provider.
- Needs from you
- zevqi init, and a model key for the questions.
- Your own model keyAvailable
Zevqi writes questions that are specific to the diff.
- Leaves your machine
- The diff of the changed files goes to your provider under your own account. Never to us. Zevqi tells you before it installs anything.
- Needs from you
- An API key you provide and pay for.
Optional. Most of Zevqi works without it.
- Local modelPlanned
AI-written questions from a model running on your own machine.
- Leaves your machine
- Nothing leaves your machine.
- Needs from you
- Install and run a model.
- GitHub appPlanned
Show who approved each change, from pull-request reviews.
- Leaves your machine
- Pull-request metadata. No code.
- Needs from you
- An organization admin installs the app.
- Cursor and Codex hooksPlanned
Capture edits directly in each tool.
- Leaves your machine
- File paths.
- Needs from you
- A per-tool install.
- Lessons written by ZevqiNot offered
We would read your diffs and write the questions.
- Leaves your machine
- Your code would reach us.
- Needs from you
- Nothing.
By design: your code never reaches us.
- Using your Claude subscription loginNot offered
Running Claude in the background on your behalf.
- Leaves your machine
- The diff would go to Anthropic through your login.
- Needs from you
- Nothing.
The provider asks products to use API keys, so we do not do this.
Statuses are literal. Available means it is built and tested. Early access means built, but not yet verified in the live tool. Planned means not built. Not offered means we have chosen not to.
Honest limits
What we do not have yet
No SOC 2 or ISO 27001 yet. We will not imply otherwise. Ask us about the control mapping during a security review.
No SSO, SCIM or audit log yet. They are on the enterprise roadmap.
Scores are self-reported. A developer can pass a quiz without understanding the code. Treat results as a signal, not proof.
Dashboard links are bearer links. Anyone with the link can view the team page. Treat it like a password until per-user access ships.
Author ids are pseudonymous, not anonymous. The git scan sends each author as a short one-way hash of their email, never the email or a name (unless you choose to share names). Nobody can read an email back out of an id, but anyone who already has a list of candidate emails could hash them and compare. Keep the dashboard link private.
AI explaining AI. Lessons are generated by a model and can be wrong. They are a study aid, not an authority.
Subprocessors
Who touches team data
| Provider | Purpose |
|---|---|
| Vercel | Hosts the website and API |
| Supabase | Stores team records (Postgres, row-level security) |
| Resend | Sends sign-in codes and team summary emails |
| PostHog | Product analytics for the website and dashboard (production only). Never receives source code |
| Razorpay | Processes payments (we never see card numbers) |
| Your model provider (optional) | Writes lessons only if you add your own key. Not used by the knowledge map, attested reviews or agent-written lessons |
Disclosure
Found a vulnerability?
Email hello@zevqi.com with the details. We acknowledge reports within 2 business days, fix confirmed issues quickly, and credit reporters who want it. Please do not test against other customers’ data.
Need a security questionnaire answered?
Book a call and we will fill it in with you, including the parts that say “not yet”.