PANTOUM runs locally against your SPFx repositories. The only outbound traffic an upgrade generates is:

• Claude API — PANTOUM's AI features route through the Claude Agent SDK, using either your Claude Code subscription or an ANTHROPIC_API_KEY in your environment. This is the only AI-side traffic.
• npm registry — the solution's own npm install during the upgrade, plus optional version lookups for the third-party dependency updater and Studio's SPFx version picker.

M365 CLI's m365 spfx project upgrade runs locally against your project files to generate the upgrade report — no Microsoft 365 tenant login is required for this command. PANTOUM itself adds no telemetry, no analytics endpoint, and doesn't phone home.

When PANTOUM hands a task to Claude — a build fix, a PnP migration, a template-driven migration — the tool set is an explicit whitelist: Read, Edit, Write, MultiEdit, Grep, Glob, LS, Bash. WebSearch and WebFetch are never enabled.

Claude gets its instructions from the templates shipped with PANTOUM (src/templates/*.md) — it does not search the web for the error, pull an answer off StackOverflow, or browse package docs. Every prompt is self-contained, every tool call is logged in the run report, and you can audit the entire AI interaction before you commit anything.

If you want to verify this yourself:

grep -rn allowTools src/ pantoum-webapp/server/

Every call site lists its tools explicitly — none include web tools.

A few habits worth having regardless of the tool:

1. Run against a branch, not main. Every run writes to pantoum_run_{runId}/ so you can diff and cherry-pick.
2. Prefer Claude Code's native auth over a hardcoded ANTHROPIC_API_KEY in scripts or CI.
3. Read the Markdown report before you commit. Every AI action is tracked as a patch with an id, a description, and a before/after.
4. Isolate untrusted codebases. For unknown or sensitive repositories, run in a fresh container or a disposable workspace.