Privacy and Public Evidence

LightClaw has no product analytics, telemetry endpoint, hosted control plane, or automatic showcase uploader. Runtime data stays under the configured local home unless the operator explicitly sends a prompt to a chosen provider or performs a separate publication action.

Data and network boundaries

Data	Default location or recipient	Automatic public upload
Config and credentials	owner-readable app config on the host	never
Memory, skills, logs, jobs, receipts	local .lightclaw/ state	never
Task source and artifacts	local owned workspace	never
Telegram messages	Telegram Bot API and the local process	never by LightClaw
Configured model context	the explicitly selected provider	not public, but leaves the host
Delegated task context	the explicitly selected local coding-agent CLI	governed by that CLI
Sanitized Run Card	explicit local destination after run export --apply	never
Showcase entry	manually reviewed Git commit/pull request	only by that explicit action

Private receipts may contain goals, commands, workspace-relative details, recovery context, and timestamps. They are mode 0600 and are not suitable for direct publication. The Run Card exporter uses an allowlist and omits commands, handoffs, checkpoints, undo paths, and timestamps, but its output still requires human review.

The public showcase accepts only synthetic or explicitly publishable input, declared provenance and consent, a sanitized Run Card, and a token-free reproducible recipe or permission-manifest skill. CI checks known secret/path patterns and replays recipes. No automated sanitizer can prove anonymity, so a maintainer also reviews the full contribution.

GitHub Actions artifacts are restricted to package distributions, coverage, security results, and generated benchmark/runtime evidence. A regression test rejects workflow artifact blocks that mention receipts, prompts, runtime homes, workspaces, or repositories.