Your prompts.
Your control.

• v8-prompt runs only when you press its key. It reads the text in the agent's input box and a few facts about the repository it runs in (next section) — not the contents of your files, apart from the narrow exceptions described there.
• Your coding agent receives the improved text only after you press Enter; the same key restores your original text.
• Its key in ~/.v8-prompt is random and not tied to you; we store only a keyed hash of it.
• Setup backs up every file it changes, and v8-prompt uninstall puts them back.

• With each prompt, the CLI sends a context pack collected from git and your manifests: the repository folder's name, the branch, whether there are uncommitted changes, the stack (languages and frameworks read from manifests such as package.json, pyproject.toml, Cargo.toml, go.mod), the test, build, lint, typecheck and run commands, the names of instruction files present (CLAUDE.md, AGENTS.md, .cursor/rules, .github/copilot-instructions.md), up to 8 repository paths that match your request by name or by content (ignored files never), the test files that belong to them, up to 2 existing files to follow as a pattern, the names of the instruction and rule files that apply to them, up to 4 commands your CI runs (with URLs, hosts, secrets and variables removed), the subject lines of the last 5 commits, and answers you confirmed earlier for that repository.
• No file contents leave your machine. File contents are searched on your device (tracked files only; binaries and files over 512 KB skipped) to rank paths, and only the paths are sent, plus the names of up to 8 functions or classes defined in those files (names only, never code; v8-prompt context symbols off stops them). The exceptions: when your manifests give no commands, up to 4 backticked command lines that look like test, build or lint commands are quoted from CLAUDE.md or AGENTS.md; and the CI commands above, redacted. The whole pack is capped at 4 KB.
• It is sent only when you press the key and is not stored on our servers. Run v8-prompt context show to print exactly what would be sent; v8-prompt context off or V8_PROMPT_NO_CONTEXT=1 sends the prompt alone.
• If Claude Code shows its last reply in the editor (as # comment lines), v8 sends at most its last 4,000 characters as context for a follow-up prompt and leaves the block in place. It is handled like prompt text: sent only when you press the key, not stored.
• Answers you give to v8's questions are remembered only on your device, in ~/.v8-prompt/projects/ (at most 8 per repository). v8-prompt forget clears the current repository; v8-prompt forget --all clears all of them.

• We process prompt text and repository context in memory and do not store them on our servers. The tool keeps your recent originals and their rewrites (at most 20, for one hour) locally in ~/.v8-prompt/rewrites/ for undo. Text that looks like a password or API key is removed before it reaches an AI provider.
• The Live page publishes aggregate successful-generation counts. Recent activity uses keyed, pseudonymous identifiers; activity older than one day is removed by daily maintenance. Prompt text is never included.
• Quotas and payment access are enforced on the server. Failed AI attempts, retries and new-device registrations have separate abuse limits. Modifying the local package does not grant more server capacity.
• No passwords: you sign in with Google or a one-time email code.
• Sessions and tokens are stored only as hashes. Payment details never touch our servers.
• See and disconnect every connected terminal, download your data, or delete your account in one click.

• HTTPS everywhere with HSTS and a strict Content Security Policy.
• Isolated, hardened containers: read-only, unprivileged, resource-limited.
• Payment webhooks are signature-verified; every change is tested before deployment.
• Encrypted secrets at rest and daily consistent backups.

Email support@v8labs.app. We respond within 72 hours and won't take legal action against good-faith research. See security.txt.