Toph 0.0.8 — Bring Your Own Brain

Toph 0.0.8 — Bring Your Own Brain

Toph no longer insists on one model vendor. Plug in whatever you've got; I'll do the talking.

New

• OpenAI API-key provider. Drop in an API key — or point Toph at any OpenAI-compatible base URL — and use it for transcription and polishing. Your self-hosted endpoint counts.
• Pick who does each job. Transcription and polishing are now routed separately, so one provider can hear you while another cleans up after you.
• Per-provider model settings. Each provider exposes its own fields — transcription model, polishing model, reasoning effort, Chat Completions vs Responses — right in Settings.
• Provider picker during onboarding. Fresh installs no longer assume you want a ChatGPT subscription. Nothing is preselected; connect whichever one you actually have, and it claims the roles still going spare.

Improved

• Connection checks before you commit. Toph verifies credentials when you connect, and a provider without working credentials can't be routed work it will only fail later.
• Cost estimates follow your models. Pricing now resolves per provider and per model, with sensible fallbacks when the catalog has never heard of what you're running.
• Settings and onboarding share one provider UI — same cards, same status badges, same secret inputs, so connecting somewhere new looks the same wherever you do it.

Under the hood

• New provider architecture: a registry, provider definitions, and a provider service replace the old hardcoded auth/transcription/inference paths.
• Credentials moved to dedicated per-provider storage with restricted file permissions.
• Substantial new test coverage across the registry, provider service, credential storage, the OpenAI clients, pricing, and the settings UI.

Contributors

Thanks to @YourTechBud for shipping this one.

Full changelog: v0.0.7...v0.0.8

Toph 0.0.7 — Polishing While You're Still Talking

I used to wait politely until you stopped speaking. I've stopped waiting.

New

• Incremental polish. Your transcript is now polished in chunks while you're still dictating, so when you hit stop there's far less left to do before the text lands. Long rambles no longer mean long waits.
• Smarter polish prompt. The rewrite instructions were rebuilt from the ground up: keeps every idea and caveat in the order you said it, preserves negations and hedges, strips fillers and false starts, and groups sentences into real paragraphs instead of one-line-per-thought.
• The Engineer preset learned structure. Spoken headings, subheadings, questions and answers, and enumerated lists ("the first one is… then there could be…") are now rendered as proper Markdown, with clearer rules on when technical tokens get backticks and when they stay prose.
• Deeper reasoning for polish. The polish model now runs at medium reasoning effort instead of low, which buys noticeably better rewrites on messy dictation.

Fixed

• Live transcription lag is now diagnosable. Some batches recorded mid-session weren't being transcribed until after the session ended. Toph now writes a privacy-safe diagnostics log (ids and timestamps only, no transcript text) so we can pin down the handover stall instead of guessing.

Improved

• Built-in presets upgrade themselves. If you never edited a shipped rule preset, you get the new text automatically. If you did edit it, your words stay untouched.
• Rule presets can be three times longer. The body limit went from 4,000 to 12,000 characters, so the new Engineer preset fits with room to spare.
• Every polish call is accounted for. Incremental chunk calls record their usage and cost against the session, and a rerun supersedes them cleanly instead of double-counting.

Under the hood

• Dependency and GitHub Actions bumps across the workspace to clear pnpm audit.
• Reruns always re-polish single-shot and never inherit incremental state from the live recording.
• Test coverage for the polish coordinator, prompt assembly, polished-text blocks, preset upgrades, and transcription diagnostics.

Contributors

Thanks to @YourTechBud for shipping this one.

Full changelog: v0.0.6...v0.0.7

Toph 0.0.6 — Now Assimilating Windows

Toph 0.0.6 — Now Assimilating Windows

Phase two of world domination begins on the world's most-installed OS. Toph now speaks fluent Windows — and got a smarter tongue while it was at it.

New

• Windows clipboard manager — Toph can now paste your polished dictation on Windows, not just eye it enviously from across the aisle.
• Windows builds in CI — a GitHub workflow that cranks out Windows binaries automatically, so Windows releases stop being a bring-your-own-machine affair.

Fixed

• Text inputs in Settings no longer surrender focus after every single keystroke. You can now type more than one character at a time. Revolutionary, I know.

Improved

• The default polish model jumped from gpt-5.4-mini to gpt-5.6-luna — sharper cleanups on your dictation, same zero effort from you.

Under the hood

• Ran the engineering reviewer over the new Windows clipboard code and cleaned up what it flagged.
• Sorted out cross-platform env handling and some build-tooling housekeeping so the Windows pipeline behaves.

Contributors

Thanks to @YourTechBud for shipping this one.

Full changelog: v0.0.5...v0.0.6

Toph 0.0.5 — Second Take

Toph 0.0.5 — Second Take

Your first attempt missed. Toph's retry no longer just shrugs and files the text away — it finishes the job and drops the words where you wanted them.

New

• Retrying a failed dictation from the pill now copies the transcription and pastes it for you, exactly like a successful run would — no more digging through recent sessions to fish out the text.
• New "Transcription copied" pill state, with a calm green glow, that lets you know the words landed on your clipboard before quietly stepping aside.

Under the hood

• Added a copied dictation phase end to end, plus delivery test coverage for re-runs so the retry path keeps behaving.

Contributors

Thanks to @YourTechBud for shipping this one.

Full changelog: v0.0.4...v0.0.5

Toph 0.0.4 — Try Again, Smarter

Toph 0.0.4 — Try Again, Smarter

Retries used to redo the entire pipeline from scratch. Now I only redo the parts that actually failed. Same result, a lot less waiting.

New

• Smarter retries for failed dictations, with a retry button right on the overlay. Instead of re-running the whole pipeline, I now retry only the failed transcription batches — and if every batch already transcribed fine, I just regenerate the output from the transcripts I already have. A full rerun only happens when your transcription settings changed. Less waiting, same result.

Fixed

• macOS now uses the proper Toph app icon instead of the generic one. The dock looks like it's supposed to.

Under the hood

• Tightened the auto-updater's internal timer and state typing. No behavior change — just fewer ways for future me to trip over my own scheduling.

Contributors

Thanks to @YourTechBud for shipping this one.

Full changelog: v0.0.3...v0.0.4

Toph 0.0.3 — Penguins welcome

The minions of world domination now run on Linux. Bring your tux.

New

• Linux support is official. Toph runs on Linux. Hotkeys, audio capture, tray, overlay — all wired up. The titlebar got slimmed down to match the platform too.
• Pick your mic and speakers. Input and output audio sources are now selectable from settings. No more guessing which device the OS handed us.
• Cancel a dictation by tapping the shortcut twice. Started talking, regretted everything? Press the transcription shortcut again and it's gone.
• Tiny version stamp in the UI so you can tell at a glance which Toph you're talking to.

Fixed

• Dictionary modal no longer plays "which button does what" with you — the UI confusion is gone, and Toph now ships with a sensible default set of dictionary entries out of the box.
• The system tray icon on Linux stops leaving ghost copies of itself behind.
• The close button on the rule-switcher overlay actually closes things now.
• Audio input refresh works on Ubuntu — there's a refresh option in the dropdown when the OS forgets your devices exist.
• Shortcut manager on Linux behaves itself.
• Overlay spacing stopped looking like it was assembled at 2am.
• The home page no longer shows a stale status from a past life.

Improved

• Fallback status is finally visible. When Toph can't use your chosen mic and falls back to another input, both the overlay and the settings page tell you about it clearly instead of pretending everything's fine.
• Settings page reorg. A new side-nav makes the page actually navigable — sections are easier to find and the layout breathes.

Under the hood

• Laid the groundwork for the auto-updater — coordinator, IPC, UI, contracts, tests. Not fully battle-tested yet, but the bones are in. Linux gets a polished fallback to manual update commands in the meantime.
• GPU acceleration is back on. You probably won't notice. We will.
• Release workflow fixes so future versions actually ship from master with the right pnpm.
• Format, lint, gitignore, and readme housekeeping.

Contributors

Thanks to @YourTechBud for shipping this one.

Full changelog: v0.0.2...v0.0.3

Toph 0.0.2 — Failure Has Receipts

Toph 0.0.2 — Failure Has Receipts

When the pipeline faceplants, Toph now keeps the receipts without doxxing your filesystem.

New

• Failed dictation sessions now stay in Recent with retry support and a copyable, sanitized debug report.
• The Engineer polish preset now understands spoken structure commands like heading, subheading, question, and answer.
• The README now explains Toph’s vibe, setup path, and how to refresh built-in Polish presets without stomping your edits.

Fixed

• No dedicated fix: commits in this release.

Improved

• No dedicated refactor, performance, or user-facing style commits in this release.

Under the hood

• Recent history now tracks retained sessions instead of only completed conversions.
• Electron Builder packaging commands explicitly disable artifact publishing.
• Package versions were synchronized to 0.0.2.

Contributors

Thanks to @YourTechBud for shipping this one.

Full changelog: v0.0.1...v0.0.2

Toph 0.0.1 — First Contact

The first cut. Voice in, polished text out, pasted where your cursor was. The loop works end-to-end — everything else is iteration from here.

New

• Voice dictation with live segmentation, backed by Silero VAD (and an energy-based fallback for batch mode).
• Transcription via OpenAI OAuth — no API keys to manage.
• Auto-paste into the focused app.
• Post-processing pipeline with raw_concat, prompt presets, and a polish-rule switcher.
• Configurable global shortcut.
• Persistent overlay and system tray.
• First-run onboarding flow.
• History page with a rerun button for re-processing past recordings.
• Cancel in-flight dictation sessions.
• Per-session cost reporting.
• macOS build + release workflow on GitHub Actions.

Under the hood

• Engineering guidance docs and a reviewer agent.
• Single source of truth for runtime settings.
• desktop-ui package with a tight public surface.
• Drizzle schema and migrations.
• oxfmt + oxlint + oxlint-tailwind in the lint pipeline.
• GitHub issue templates, license files, wtp config, and AGENTS.md.

Contributors

Shipped by @YourTechBud.

Full changelog: https://github.com/YourTechBudStudio/Toph/commits/v0.0.1