Last reviewed 2026-06-03 · Pinta v0.4.0
The short version: Pinta is bring-your-own-Claude. It runs as a skill inside your own interactive Claude Code terminal. It never proxies, stores, or shares your Anthropic credentials, never uses “Login with Claude.ai” OAuth, and is not a headless / Agent-SDK / claude -p / cron tool. You use Claude under your own Anthropic plan and terms — Pinta is only the UI that hands annotations to the agent you already run.
Pinta has three pieces and none of them talk to Anthropic:
pinta-companion) — runs on localhost, writes sessions to your project's .pinta/ folder./pinta) — runs inside your already-authenticated, interactive Claude Code terminal and reads the session from the companion.The model call happens in your terminal, under your account, exactly as if you'd typed the request yourself. Pinta is a consumer of Claude Code, not a replacement for it, and it has no network path to any AI provider.
/pinta skill is event-driven, runs in an interactive terminal, and pauses itself after ~30 minutes idle. It is not an always-on bot, not Agent-SDK, not claude -p, not a cron job.Interactive, individual use of Claude Code — the way Pinta runs — is the supported lane under Anthropic's Pro/Max subscription plans. If your usage is heavy or automated, that belongs on API billing, not a subscription. Pinta doesn't change which billing applies; it inherits whatever you've signed in with in your own terminal.
In April 2026 Anthropic stopped letting Claude subscription credentials be used inside third-party harnesses (the OpenClaw situation). The pattern Anthropic acted against is specific — and it's the opposite of how Pinta works:
| The banned pattern | How Pinta works |
|---|---|
| Logs in with the user's Claude.ai subscription OAuth | Never touches Claude.ai OAuth. You run your own Claude Code |
| Acts as a third-party harness that calls Claude programmatically | Not a harness. It hands work to the interactive Claude Code you run |
| Headless / automated high-volume pipelines on a subscription | Interactive, human-in-the-loop, idle-paused, with the model deciding edits with your confirmation |
| Proxies / shares credentials through a central service | No central account; credentials never leave your machine |
| Disrupts Anthropic's prompt caching | Sits outside the API path entirely; Claude Code manages its own caching |
The distinction that matters: OpenClaw was the thing talking to the model. Pinta is a UI that hands work to Anthropic's own officially-supported, interactive surface. The behaviors that got OpenClaw banned are exactly the ones Pinta is designed to avoid.
Pinta's optional modules (e.g. GitLab Issues) run agent-side, after edits land, using your own local CLIs (glab auth login, etc.). They don't add any new path to Anthropic, don't store tokens, and only run when you opt in twice (Settings + per-submit checkbox). They don't change the compliance posture described above.
If you deploy Pinta to a team, each user must supply their own Claude account/key — Pinta will not route multiple users through one subscription. Before any commercial or team deployment, review the current Anthropic terms linked below; you are responsible for your own usage under your own plan.
Questions about how Pinta uses Claude? Open an issue at github.com/kevzlou7979/pinta/issues or email kevzlou7979@gmail.com.