Vibe coding level:

You don't think of yourself as a "builder" but you'd like to become one. You want to move from "just talking in meetings" to actually building real software. The terminal is a terrifying black box.

How Sparkle improves Sparkle

Sparkle ships with a built-in improvement agent: a special agent that works on Sparkle itself instead of your project. It watches how the app actually behaves on real machines — the errors people hit, the flows that stall, the slow paths — and turns what it learns into small, scoped pull requests against the open-source Sparkle client.

The whole point is to make the app better for everyone without anyone having to file a bug report. But because it learns from your usage, you decide whether it runs at all, and nothing ever leaves your machine without your say-so.

At the top of the Sparkle pane there's a single control — "Can we use your anonymous logs to automatically improve Sparkle?" — with three settings:

  • Always — Sparkle evaluates your logs hourly and, when it finds a clear, privacy-safe fix, submits the pull request automatically. No action required from you. (Submission only happens if the automatic privacy scan passes — see below.)
  • Case by case (default) — Sparkle drafts the pull request and shows it to you. Nothing is submitted until you approve it. This is the recommended setting.
  • Never — Sparkle does not evaluate your logs at all. They stay on your device, and the improvement agent never reads them. You can switch this back on whenever you like.

How it works, step by step

  1. Once per hour, Sparkle uses a small amount of your Claude Code subscription to review your recent session logs for failures and performance problems.
  2. If it spots something worth fixing, it drafts a focused, single-purpose change against an app-owned clone of the open-source Sparkle repository — never your own project.
  3. It writes a short spec (problem, evidence, proposed change) and the code change, then scrubs everything for anything sensitive.
  4. Depending on your setting, it either shows you the PR to approve (Case by case) or submits it automatically (Always).
  5. The PR goes to the open-source Sparkle project, where maintainers review it like any other contribution.

What we send — and what we never send

Privacy is enforced at two levels: by how the agent is wired, and by what it's allowed to write.

Structural guarantee. The improvement agent can only edit the app-owned clone of the open-source client. It physically cannot touch your project's files — so the code in any pull request is Sparkle's own open-source code, never yours.

Scrubbing. Before a pull request is submitted, Sparkle scrubs the title, description, commit messages, and diff. We do not send:

  • Personally identifiable information (names, emails, anything that identifies you)
  • Secrets, API keys, or tokens
  • Snippets of your code, prompts, or file contents
  • File paths from your machine, project names, or URLs

The agent derives only anonymized, aggregated insights — for example "the worktree step intermittently fails under index-lock contention" — and never pastes raw log lines. If a fix could only be justified by including something sensitive, the agent doesn't open the PR at all.

Your logs stay local. The raw logs are read on your device and are never uploaded. Only the scrubbed, anonymized pull request ever leaves your machine.

How pull requests reach the open-source project

You don't need a GitHub account. By default, approved improvements are submitted to the open-source Sparkle project through Sparkle on your behalf, so contributing is zero-friction. If you'd rather contribute under your own GitHub identity, you can opt in, and Sparkle will open the pull request from your account instead — your call.

You're always in control

  • Change the Always / Case by case / Never setting any time from the Sparkle pane.
  • On Case by case (the default), every pull request waits for your approval.
  • Switch to Never to turn the whole thing off — your logs are never evaluated again until you turn it back on.

Sparkle's client is MIT-licensed open source. Improving it from real-world usage — with your consent, and with your data kept private — is how it gets better for everyone.