Data & Trust

No model reads your code on a free scan. Checks run on nittim's own servers, and the only outside calls about your code go to GitHub, to fetch it, and OSV, to check package versions.

An AI audit sends a prioritized code digest to Anthropic's API. Commercial terms bar training on it — nittim has no data-sharing deal and keeps nothing after. See Anthropic's commercial terms for their side of that boundary.

Anthropic deletes API inputs and outputs within 30 days, unless a zero-data-retention deal exists — nittim doesn't have one, so that window is the vendor's own guarantee, not ours. See Anthropic's data-retention policy.

nittim can never modify your code. The nittim-audit GitHub App holds read-only permissions only, no events. Administration: read-only reads settings, never changes them — it is how nittim sees what your default branch requires before a merge. GitHub enforces it regardless of nittim's code; the exact list is at github.com/settings/installations. Using a token instead? Make it fine-grained, read-only, scoped to one repo, and revoke it after.

Reports are kept until you choose to delete them. There is no automatic expiry for live reports.

When a report is deleted, it immediately disappears from all read paths — the report link returns not-found and the API returns 404. The underlying row is then permanently removed from the live database on the next hourly retention pass. Rows deleted more than 30 days prior are purged on that pass; a row deleted today will be purged on or after the corresponding date 30 days out.

That 30-day window governs the live database — everything a report link, the API, or nittim itself can query. A separate, nightly encrypted backup of the whole database also exists, for disaster recovery only, and a deleted report's data can remain inside that backup for up to 120 more days after the 30-day window closes. Backups are never queried for an individual record — they expire on a fixed schedule and are only ever restored to recover the entire system, never one report.

You own every report you generate. To delete one:

DELETE /api/v1/audit/<id>
Authorization: Bearer <your-api-key>

Dispute any finding — on the report page, or via your agent over MCP with evidence from your code. A person reviews every dispute. An accepted one gets corrected on the record, and you're paid in credits — more for a pipeline flaw than a wrong finding. We pay to be proven wrong because it's worth more than the credits.

Public reports and scans are open to anyone with the link — that's the whole point. Delete one to stop it being reachable.

Private or posted source is locked to its owner or org member by default — everyone else gets 404. Only the owner or an org admin can create, see, or turn off a share link (the Share button always says what it hands over); an ordinary member can see one exists but never gets the link. A share link uses its token, never the report's id, so it can't widen what any other page shows. It works for anyone who has it, no account needed, until it's turned off — then it stops working immediately, for everyone.

Everyone in an organization can read a report run inside it; an owner or admin can delete it. Reports stay scoped to their own organization.

Reports live in Supabase (Postgres, us-east-1), encrypted in transit (TLS 1.2+) and at rest (AES-256). A Vercel-held service-role key reaches the database — never sent to your browser, never committed to source.

Over MCP, nittim pulls nothing: only the files your assistant posts arrive, and the server cannot ask for more. Where they land is the same server and the same 6 places as every other path — MCP changes the doorway, not the map.

This list narrows where a copy could be. It cannot prove our code never writes to one of these places — only an independent review of the code path, or running both halves inside your own cloud, can close that. We'd rather say so than let the table imply more than it shows.

Only the AI tiers send to a model. This is what goes.

• Source files in the languages we read — the repository's own code.
• Dependency manifests and lockfiles — package.json, requirements.txt, go.mod, Cargo.lock and their peers.
• Configuration and infrastructure — Dockerfiles, CI workflows, Terraform, .env files and their examples.
• Documentation the repository ships — Markdown and plain text.
• Executable scripts with no extension, admitted by their shebang line.

Audit · File count500

files fetched from the repository, at most.

Full Audit · File count2000

files fetched in ONE window — and a Full Audit fetches window after window until every eligible file has been read, so this is not a total.

Audit · Per-file size200 KB

per file — a bigger file is skipped, never partially guessed at.

Full Audit · Per-file size4 MB

per file on this tier — the lockfile cap, applied to every file. A big source file an Audit skips is fetched whole here.

Both · Lockfile size4 MB

for a lockfile, which is long by nature and is where the dependency truth lives.

Audit · Total size5 MB

of source in total, across everything fetched.

Full Audit · Total size20 MB

of source per window, not in total — see the sentence under this list for what a Full Audit adds up to.

Audit · Per-file characters16K characters

of any single file reach the model; past that the file is cut and marked as cut.

Full Audit · Per-file characters200K characters

of any single file reach the model on this tier; past that the file is cut and marked as cut.

A Full Audit is sold on reading every eligible file, so the numbers per window above are not a total: it fetches one window, then the next, until every eligible file in the repository has been read. What travels over a whole Full Audit is your repository's own eligible source — the file kinds listed above, at the per-file caps above — and nothing else.

Nothing is stripped. Comments, strings and secrets-looking text travel exactly as they are written in the file — the only changes are the cut at the per-file cap above, and the file markers around each body being escaped so a file cannot forge one. Images, fonts, binaries, vendored dependencies and generated bundles never travel at all.

A report is one row. Every column, by what it holds.

3 columns can hold text from your repository: the report itself, because every finding carries the lines it points at, and the two working-state columns a multi-pass or batch audit writes as it goes and clears the moment it finishes. Every other column is an id, a verdict, or machinery.