What leaves your machine

The DataParade CLI is local-first: the structural scan pipeline — ingest, analysis, classification, data-flow detection, and writing dataflow.json — runs entirely on your machine and sends nothing anywhere. Every network interaction is tied to a feature you explicitly configure. This page is the complete list.

By feature

Feature	When it makes network calls	What is sent	How to avoid it
Structural scan (`npx @dataparade/cli scan .`)	Never	Nothing	— (this is the default)
Fetching the CLI itself	`npx` downloads the package from the npm registry on first use	Standard npm request metadata	Install once (`npm install -g @dataparade/cli`) or run from source
Auto-upload after scan	Only when `DATAPARADE_WORKSPACE_API_KEY` / `--workspace-api-key` is set	`dataflow.json` — component names, properties, and file paths with line ranges; raw source snippets are stripped before writing the file	Don’t set a key, or pass `--skip-auto-upload` / `DATAPARADE_SKIP_AUTO_UPLOAD=true`
AI inference, BYOK (`--ai-inference` + `SCAN_BYOK_*`)	During the optional AI pass	Prompts containing bounded excerpts of scanned files (never `.env` contents) go to your chosen provider at its official endpoint	Leave AI inference off (the default)
AI inference, platform (`--ai-inference` + workspace key)	During the optional AI pass	Quota preflight (file count and ingested bytes), inference tasks with the same bounded excerpts, and a completion report to the DataParade API	Leave AI inference off, or use BYOK
Custom AI endpoint (`--ai-endpoint` / `SCAN_AI_ENDPOINT`)	During the optional AI pass	Same prompts, to any URL you specify — only point it at endpoints you trust with code excerpts	Don’t set a custom endpoint
Crash reporting (Sentry)	Only when a scan fails with an error	The error message and failure phase, plus the scan-root path; no file contents (`sendDefaultPii` is off)	`SCAN_SENTRY_ENABLED=false`
LangSmith tracing	Only when you set `LANGSMITH_API_KEY` / `LANGCHAIN_API_KEY` and enable tracing	Summary-only run metadata (counts and ids) — no file contents or full scan output	Don’t set LangSmith variables (off by default)

A quota preflight is only called for platform AI: a workspace key with AI inference off runs the scan fully locally and contacts the API only for the upload step described above.

Guarantees that hold regardless of mode

.env is never read. .env and .env.* files are excluded from ingest and are never embedded in AI provider prompts. Pointing scan at a single .env file is refused with a security warning. Other credential files are not auto-excluded — add them with --exclude if your repo has any.
No raw source in the output. dataflow.json keeps file paths and line ranges only; sourceLocation.code snippets are stripped before the file is written. The same discipline applies to AI insight evidence in metadata.
AI suggestions must cite evidence. Provider proposals must reference exact repo-relative paths that exist in the scan — no fuzzy path matching. Rejected proposals are not merged.
Keys are never printed. config output redacts SCAN_BYOK_API_KEY and DATAPARADE_WORKSPACE_API_KEY as <redacted>.

Fully offline checklist

For a scan with zero network traffic:

Install the CLI once (or vendor it), so npx doesn’t hit the registry.
Don’t set DATAPARADE_WORKSPACE_API_KEY.
Don’t enable --ai-inference (off by default) — or use SCAN_BYOK_PROVIDER=local against your own Ollama-compatible endpoint.
Set SCAN_SENTRY_ENABLED=false so even a crashing scan reports nothing.

Then scan reads your repository, writes dataflow.json, and touches nothing else.