Effective 2026-11-13 · Pinta v0.13.0
The short version: Pinta is a single-user developer tool. We don't operate any servers, collect any data, or use any analytics. Your annotations and project files stay on your machine. The only requests the companion makes to the internet are the AuditFlow dependency checks, and only when you click (section 7). The Chrome extension talks to a companion process on localhost; the companion talks to your AI coding agent (Claude Code, Cursor, Aider, etc.) which is also running locally.
Pinta lets you annotate your running web app — drawings, element selections, comments — and hand the annotations to an AI coding agent that edits your project's source files. There are two pieces:
pinta-companion on npm) — runs on your machine on localhost, started by you.outerHTML snippets (max 2 KB), nearby text, computed styles, your free-form comments, and full-page screenshots.chrome.storage.local.pinta-standalone database), keyed by the URL's origin.When you click Submit, the extension sends the session payload over a WebSocket to 127.0.0.1 on the port the companion is listening on. The companion writes submitted sessions to .pinta/sessions/ inside your project root. Both ends are on your machine.
The companion exposes the session over an HTTP API and an MCP server, both on localhost. Your AI coding agent (running locally) reads the session and decides what to edit. Pinta has no connection to the AI provider — the agent's own privacy policy governs how it handles the prompt.
Captured outerHTML is sanitized client-side before it leaves the page: inline event handlers (onclick etc.), integrity, nonce, CSRF / token / auth / JWT / bearer attributes, and password input values are all stripped from the captured HTML. The capture is also truncated to 2 KB to bound exposure.
If you annotate inside a tab that has authenticated session state (cookies, tokens), the screenshot will of course render whatever's on the page — Pinta is showing the agent what you see. Treat your agent the same way you'd treat anyone you screen-share with.
<all_urls> host permission — Pinta is a developer tool you invoke on your own running app at any URL. No site is targeted automatically.tabs — to read the active tab URL for routing the session to the matching project companion.activeTab + scripting — to inject the annotation overlay and to capture screenshots of the user-activated tab.sidePanel — to host the side-panel UI.storage — for theme preference; standalone sessions use IndexedDB.offscreen — only for the optional Voice Command feature (off by default, enabled per-project in Settings): it hosts the microphone and the browser's built-in Web Speech recognition in an offscreen document. Speech recognition is done by the browser's built-in speech service, which may send audio to the browser vendor; Pinta itself does not send or store the audio.The extension bundles its fonts (Poppins) — no runtime fetch from Google Fonts. The landing page at pinta-companion.vercel.app does load Google Fonts; that's a static site, not the extension. The extension itself makes no network calls beyond localhost; the companion's outbound requests are listed in section 7.
Starting in v0.3.0, Pinta supports opt-in modules — agent-side integrations the user explicitly enables in Settings and explicitly ticks per submit. When ticked, the agent runs the module after applying source edits. Modules are the only path by which Pinta-related actions can send data off your machine. Most run only when you opt in twice (Settings + per-submit checkbox); the AuditFlow Dependencies checks run only when you click their button.
glab CLI. Pinta does not handle, store, or transmit GitLab tokens. glab auth login is the user's responsibility; Pinta only invokes glab as a subprocess./uploads endpoint (via glab api) and embedded as markdown in each issue body. Screenshot capture is locked on whenever this module is ticked so issues never go out without visual context.AuditFlow reads your project's manifests and lock files on your machine (npm, pnpm, yarn, Maven, Composer, Python). Reading and the local rule-engine scans send nothing anywhere. Two actions make requests, and only after you click them: the vulnerability and stale-version checks in the Dependencies category, and Update definitions in Threat Watch. They contact these hosts and no others:
api.osv.dev — the OSV vulnerability database. The vulnerability check sends the public package name, ecosystem and installed version of each package, and the advisory ids it needs details for. Update definitions sends package names and ecosystems only, with no versions.registry.npmjs.org — npm packages. The stale-version check sends the package name in the request address.search.maven.org — Maven Central. The stale-version check sends the group and artifact name.repo.packagist.org — Packagist (Composer). The stale-version check sends the package name.pypi.org — PyPI. The stale-version check sends the package name.Pinta only asks about public packages. Never sent: private, workspace, local, git and path packages; pnpm and yarn 2+ packages (listed as not covered); your source files or any file contents; secrets of any kind; your project name or path. A package with a name or version that does not look like a public one is skipped. Requests carry no cookies or credentials, only the name pinta-companion as the user agent. Those services see your IP address and what is listed above, and handle it under their own policies. Pinta does not run these checks on a timer.
Stored locally: check history, definitions fetched with Update definitions, and false-positive marks are written under .pinta/audit/ in your project root. False-positive marks are keyed with a random key kept in ~/.pinta/audit-fp.key on your machine, so a finding's text is not stored. Delete either to clear it. Secret-looking values are redacted from findings before they go to Fix, Discuss or File issue.
Resolve with Maven is off until you press it. It then runs the Maven on your machine (mvn) in your project, and Maven may then contact Maven Central and any repositories your project or your Maven settings name, and may download plugins and dependencies. Pinta itself sends nothing in that step and does not control those hosts.
Threat Watch's file watcher is local only: it watches a few project folders for changes to manifests, lock files, .mcp.json and CLAUDE.md, and re-scans on your machine. It opens no network connection.
These checks look for known patterns and published advisories. They can miss problems, and a clean result does not mean a project is safe.
pinta-standalone..pinta/sessions/ from inside your project root..pinta/audit/ from your project root, and ~/.pinta/audit-fp.key to reset false-positive marks.Pinta is a developer tool. It is not directed at children under 13.
Material changes will be reflected by bumping the effective date at the top of this page and noted in the CHANGELOG.
Open an issue at github.com/kevzlou7979/pinta/issues or email kevzlou7979@gmail.com.