Privacy & code custody
Your code stays yours.
Mapparoo's audience is handing over their most sensitive asset. These are the custody commitments — read-only access, temporary clones, evidence excerpts only, and every processor named before you connect anything.
Read-only repository access
Repository access is granted through a GitHub App you install on the repositories you select, using short-lived per-run tokens. Runs happen only when you invoke them.
Temporary clones deleted after each run
The full repository exists in exactly two places during a run — GitHub and the isolated sandbox — and the temporary clone is deleted when the run ends. We do not retain raw clones.
Guides keep evidence excerpts only
Analysis artifacts — registries and guides — are retained as your work product until you delete them. They contain only the limited code and document excerpts cited as evidence.
Self-serve deletion
Full project deletion and full account deletion are always available from your account. Deletion removes projects, runs, registries, guides, and revokes hosted share links.
No model training on your content
Your content is processed by our AI provider under terms that do not use it for model training. The sandbox vendor (paid Pro) has model training off by default; that setting is confirmed for our team. The current arrangement is disclosed here and kept current.
No scores, ever
You will never see "your app is 78% complete." Findings carry evidence, not false-precision percentages.
What "independent" means
Independent means independent analysis — not a different vendor. Mapparoo runs on Anthropic’s Claude models, the same family many founders build with. What makes the audit independent is how it is built, not who made the model. Your documents and your repository are read separately, with nothing shared between the two readings. The stages that judge and verify see only the validated results of those readings — never your raw documents, never your code, and never any session of the agent that wrote it. The verification pass tries to prove findings wrong, and the guide says which findings it examined. And every run records which model and prompt version each stage used. So no model grades work it produced, in a context it produced it in.
Processors of customer IP
Every third-party processor of your repository and document content is named here. Adding a processor is a change we announce — not something that happens silently.
| Processor | Role |
|---|---|
| GitHub | Source of repository content (read-only App you install) |
| Anthropic | AI analysis (API processing of run content) |
| Vercel Inc. | Isolated tier-2 execution (clone at a pinned commit; declared install/build/test) in a Firecracker microVM per run, US region iad1 on AWS. Processed under Vercel’s Data Processing Addendum and Terms; model training off on our paid Pro plan. DPA and subprocessors: security.vercel.com. |
| Cloudflare (incl. R2) | Object storage of uploads and artifacts; network transit and edge security |
What we disclose honestly about the sandbox processor
No persistence
Each run’s sandbox is non-persistent: its filesystem is discarded when the run ends, and no snapshot of your code is retained by the vendor. Verified in our qualification testing, including the crash path.
Execution records
The vendor retains a record of which commands were run and when (not the output those commands produced) in its operational logs, per its observability retention schedule. We keep your code’s content out of that channel by design: command output is written to a file inside the sandbox and read back over a separate control channel.
Single region
Sandbox execution occurs only in the US (iad1); no regional choice is currently offered at this tier.
No execution SLA
The vendor provides no uptime SLA for sandbox execution at our plan tier; on vendor outage we degrade to tier-1 (static) analysis rather than block runs.
Process limits
We apply a per-phase process ulimit inside the sandbox. On the current vendor substrate, the kernel may not refuse fork storms under that limit (cgroup process ceiling remains high); wall-clock caps and forced teardown remain the backstop.
Processors of account & payment data
| Processor | Role |
|---|---|
| Stripe | Payments, receipts, invoices |
| Email provider | Transactional email (run complete, receipts, security events) |
| GitHub | Authentication when you sign in with GitHub |
Legal
The full Terms of Service and Privacy Policy include additional detail on data handling, retention, and your rights.