Changelog

Every release, in order.

Pulled live from GitHub Releases on our repo. Refreshed at most once an hour. Latest first.

Recent desktop highlights

the headline changes per build, newest first

v1.0.99 — Cloud Video quality-hardening

July 9, 2026

Grok first-frame generation for Cloud Video: both Economy and Premium now start each clip from a sharp, aspect-matched first frame, for cleaner and more on-subject motion.
Cleaner clips: the image prompt behind that first frame is sanitized so stray “title card / captions / credits” wording can't get painted into the opening frame, and the video prompt is length-clamped so the motion direction is never silently cut off.
Honest, measured quality hardening — not a claim of perfection: first-frame provenance recorded per job (local-only), privacy-safe prompt logging on the cloud gateway (no raw prompts), and observe-only first-frame QA scaffolding. Verified end-to-end with real clips.
Windows: Microsoft Store (recommended) or direct .exe. macOS (Apple Silicon & Intel). Direct builds are unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.98 — Steadier sign-in + cleaner Cloud Video

July 9, 2026

Cloud Video clips come out cleaner: removed a prompt guardrail that could ironically cause garbled on-screen text, and Premium clips render faster and more consistently.
Honest time estimates: the “time remaining” on a Cloud Video generation now reflects real measured speeds instead of under-promising.
Sign-in that stays signed in: a dead-refresh-token edge case that could log you out or show “cloud rejected” now clears itself narrowly — only on a confirmed-dead session, never on a momentary network blip.
Windows: Microsoft Store (recommended) or direct .exe. macOS (Apple Silicon & Intel). Direct builds are unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.97 — Cloud Video motion + faster Premium

July 9, 2026

Your storyboard's per-scene motion now reaches the Cloud Video model: the camera and subject movement the brain plans for each scene actually drives the generated clip, instead of a generic pan.
Premium Cloud Video is faster and the in-app time estimate matches reality (a stale estimate showed ~21 min for work that takes ~7).
Auto-length clips: a Cloud Video clip follows the scene's narration length — the gateway splits a long scene into parts and returns one stitched clip — so you're not capped at short segments.
Windows: Microsoft Store (recommended) or direct .exe. macOS (Apple Silicon & Intel). Direct builds are unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.91 — TubeForge Cloud Video + the $15 Starter Bundle

July 8, 2026

TubeForge Cloud Video (new, recommended video engine): generate animated clips on TubeForge's own GPUs with no API key of your own, billed by credits — Economy (low-cost, up to 20s) and Premium (higher-detail, up to 40s, returned as one clip).
The $15 Starter Bundle: a one-time $15 unlocks the desktop app for life (no subscription) and includes 500 Cloud credits to start. Existing license buyers keep everything — local + bring-your-own-key access is unchanged.
Bring-your-own-key OpenRouter / Grok video stays as the Advanced fallback and never uses credits, and local rendering is still free on your own GPU — no cloud render farm, ever.
Windows: the Microsoft Store is now the recommended install (Microsoft-verified, no SmartScreen, auto-updates); the direct .exe remains available. macOS (Apple Silicon & Intel) unsigned.

v1.0.89 — Stay on the latest build

July 5, 2026

Out-of-date installs now get a clear “time to update” prompt, so no one is stuck on an old build that's missing the latest fixes. Your license carries over automatically — nothing to re-buy or re-link.
A maintenance release on top of v1.0.88. As always, the byte-identical render guarantee held — your existing projects render exactly the same.
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.88 — Your license, verified — with offline grace

July 4, 2026

Your paid access is now verified reliably against your account, with a built-in offline grace window: keep working for up to 12 hours between checks, so a short trip off Wi-Fi never interrupts you while your license stays tied to your account.
Sign-in is steadier across restarts, and the account gate no longer trips on a slow or momentary connection.
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.87 — Hardened top to bottom: your money, your uploads, your data

July 4, 2026

Cost control that actually holds: your Profit-Mode spend cap now enforces on image, voice, and thumbnail generation (it could previously be exceeded), and cancelling an AI-video job immediately stops the spend instead of finishing and billing every remaining clip.
Your uploads stay private: an upload you scheduled as private or unlisted is no longer silently published to the world, and the same finished video can't be posted to YouTube twice.
Your keys and data stay yours: your AI API key is now sent only to the AI provider (never to a download URL it hands back), backups are taken as a consistent database snapshot that can't be corrupt, and the auto-update feed and sign-in were hardened against SSRF/CSRF.
Fewer things break: large YouTube uploads no longer fail near the end (the access token refreshes mid-upload), crashed AI-video jobs recover on restart instead of getting stuck, and cloud sync reconnects on its own after a Wi-Fi drop. Dozens of fixes from a deep internal audit, each covered by an automated test; the byte-identical render guarantee held throughout.
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.86 — Deep-audit blocker fixes: renders, cost, cloud sync, security

July 3, 2026

Renders you can trust: layered / keyframe scenes that used to come out solid black now render correctly, and mixing video formats in one project no longer produces a corrupt file — the two ways a video could silently fail while still reporting success.
Paid work isn't wasted: a completed AI-video clip is no longer thrown away and replaced with a still image, and a hiccup mid-generation no longer double-charges you.
Cloud sync + entitlement: deletes now propagate across your devices (no more 'zombie' rows reappearing), and paid access can no longer be granted from the client — it is enforced on the server.
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.85 — Storyboards carry the whole script

July 3, 2026

The storyboard can no longer shrink your video. If the AI summarises your script instead of cutting it word-for-word (a real 1,698-word script came back as 24 scenes totalling ~430 words — a 2-minute video from an 11-minute script), TubeForge now detects the shortfall and re-cuts the narration deterministically from your actual script at sentence boundaries. The AI's scene count and image prompts are kept; the words are always yours, complete, in order.
Scene durations now always derive from the real narration (150 words per minute), so the storyboard's total runtime matches the script's length — on the manual Storyboard page and in the guided director alike.
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.84 — Scripts hit their length + real YouTube titles

July 3, 2026

A 15-minute video now gets a 15-minute script. Length is a hard requirement in the generator, and a new length-guarantee loop keeps the same model writing seamless continuations until the word target is met — a 15-min target means ~2,250 words, not the ~350 a single pass could return (which produced a 2-minute video).
Ideas now come with real YouTube titles: Title Case, payoff-driven, and the duration can never leak into the title again (no more “night sky surrender — 15 min”). A server-side sanitizer guarantees clean titles no matter which model you use — the duration lives in the idea's length field where it belongs.
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.83 — See what's failing + a brain that learns your visuals

July 2, 2026

New Known-issues panel in Diagnostics: recent warnings and errors from your machine's local log, grouped by event with counts and the latest detail — so you can see exactly which subsystems need attention instead of problems hiding in a log file. Local-only; nothing leaves your computer.
The channel brain grew a visual memory: when a thumbnail wins for real — you apply an A/B comparison winner, or a rotation finalizes with a measured CTR lift — the winning composition (angle + prompt) is remembered and automatically applied to future thumbnail and image prompt generations. Each channel's visuals converge on what its audience demonstrably clicks.
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.82 — One brain, one key: OpenRouter-only + your money, your choice + smooth motion

July 2, 2026

One-click Connect OpenRouter: sign in on openrouter.ai and every AI role (text, image, video, voice) is keyed instantly — no more pasting four keys. All your usage shows in one OpenRouter dashboard.
The separate direct-Anthropic (Claude) provider was removed entirely — OpenRouter serves every Claude model, so one key powers the whole brain. A leftover Anthropic key is cleaned from your keychain automatically, and saved “Claude” preferences keep working via OpenRouter with zero action.
New channel brain memory: the top evidence-backed learnings from each Improve analysis are remembered per channel and silently applied to every future script, storyboard and metadata generation — the more videos you analyze, the sharper the next one gets.
A Mode can never silently put you on a paid provider again. If your channel is set to free Pollinations or Edge-TTS, Max-Quality mode keeps it free — paid image/voice/video providers are reached ONLY through your explicit pick or an explicitly-saved paid channel default. (Previously, Quality mode overrode your free choice and billed premium models without asking.)
The hidden image-to-video init frame is now ALWAYS generated with the free provider. It's invisible to you, so it never spends your money — under any mode, cap on or off.
Brain failover got a cost ceiling: when a cheap model rate-limits or fails, the fallback can only be a same-price, cheaper, or free model — never a silent jump to premium-priced models.
Ken Burns shimmer fixed: motion filters now supersample 2× before the zoom (the same proven fix Shorts always had), so rendered scenes glide smoothly instead of vibrating.
More fake-timeout fixes: YouTube package generation, quality review, Shorts metadata/hooks/thumbnails, channel branding and the safety check no longer show a false “timed out” error while they finish in the background.
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.81 — No more fake timeouts on long AI generations

July 2, 2026

Big AI operations — storyboards with 20+ scenes, full scripts, motion suggestions for all scenes, upgrade-all-prompts, thumbnail batches, retention analysis, Shorts repurposing, Hook Lab, Improve, the Channel Assistant and Director plans — previously hit a flat 2-minute timeout and showed “This took too long and timed out” while the generation actually finished fine in the background (the result appeared on the next refresh).
Each of these now has a proper multi-minute timeout, so the app stays in its loading state and delivers the result instead of a false error.
Cleaned up stale “mock provider” notes in the Storyboard, Ideas and Analytics dialogs (the mock system was removed long ago).
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.80 — AI-motion scenes actually render with their clips

July 2, 2026

Fixed a race in the guided “make a video” director: it started the render the instant AI-video clips were queued, so the “AI motion” scenes you selected (and paid for) at the cost gate almost always rendered as static images while the clips finished uselessly in the background. The director now waits for every clip before rendering.
Fail-open: if a clip times out or fails, the build still completes using the scene's still image — a build never dies waiting.
The build's cost now includes the real AI-video spend (it previously showed $0 for the clips).
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.79 — Accurate AI Bill, smoother AI video, fresher internals

July 2, 2026

Your AI Bill is now complete: every image and voice generation is recorded (they previously logged nothing, so the in-app total under-reported real spend and the Profit-Mode cost cap under-enforced on media).
Profit Mode can't be bypassed anymore: the hidden image-to-video init frame switches to the free image provider when the cap is on, and its cost shows on the video job.
Chained AI clips no longer come out in slow motion — duration fitting caps slow-down at 1.15× and trims large overshoots instead of stretching.
A clip that fails mid-chain no longer wastes the clips you already paid for — the video completes with what was generated; clip timeouts are retried.
Clear errors: “brain not configured” now says exactly that (with what to do) instead of an opaque Internal Error, and invalid AI-Mode values are rejected instead of silently degrading to Balanced.
All backend + desktop dependencies refreshed (0 known vulnerabilities). Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.78 — Captions that work + a real Shorts upgrade

June 26, 2026

Captions fixed: “Generate timing” is now instant and reliable — it previously ran a heavy speech model that took minutes per scene and timed out, so every scene showed “0/24, failed”. And the caption style you pick in the preview now matches exactly what gets burned into the video (real fonts + animation), for long videos and Shorts.
Shorts are now a real clip, not one static image: each Short is built from a fast-cut sequence of fresh 9:16 images (about one every 5 seconds), gently animated, with the parent video’s caption style and your channel’s background music mixed in automatically.
One-click matching Short thumbnail: regenerate your video’s thumbnail design in vertical 9:16 with the Short’s title burned on — so your Short looks like the video it came from, optimized for the Shorts feed.
Punchier publish details: Short titles, descriptions and hashtags are now tuned for maximum impressions (curiosity-gap titles, search-friendly first line, broad + niche hashtag mix).
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.77 — Accurate AI billing + cheaper, better AI video

June 25, 2026

Accurate cost tracking: the in-app AI Bill now shows the real model for each AI clip (not a generic “stub”) and OpenRouter’s actual billed cost, so it matches your provider invoice.
Smarter, cheaper AI video: when a scene has no image, the app now auto-generates a high-quality init frame in the background and animates THAT (image-to-video) instead of having the video model imagine the whole scene — better, on-model clips with far fewer expensive regenerations.
Background music now auto-attaches from the channel when a render doesn’t specify a track (it was being silently dropped).
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.76 — Smooth clips + images: no audio bleed, captions burned

June 25, 2026

Fixed mixing AI clips with images: a clip shorter than its narration no longer lets the previous scene's voiceover bleed into the next scene. Every scene is now frame-exact (audio and video the same length), so playback stays perfectly aligned.
Captions are now burned on scenes that use a clip and/or image (the Flexible Render Studio path previously dropped subtitles).
Visuals page: a scene covered by an AI clip no longer asks you to also generate an image — selecting either one is enough — and a finished clip appears in the selection bar instantly (no tab-switch needed).
Voice model picker now filters out music/audio-gen models that the speech endpoint rejects. Windows + macOS (Apple Silicon & Intel) builds — still unsigned (SmartScreen / Gatekeeper prompts expected).

v1.0.75 — Model defaults + a no-surprise-charge fix

June 25, 2026

New default brain model: Z.ai GLM 5.2 via OpenRouter (1M-token context, strong reasoning, low cost) for scripts, ideas, hooks and other writing — you can still pick any model per job or per channel.
Fixed: explicitly choosing a free model on the Script page no longer silently falls over to a paid model if the free one is rate-limited — a free pick now stays free (or fails honestly), never a surprise charge.
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.74 — One AI model system — pick any model, per job

June 25, 2026

Unified model selection across every role (script, image, video, voice): a clear priority order — your per-job pick → the active Mode (Max Quality / Cost / Speed / Balanced) → the channel default → provider default. Modes now steer image, video and voice too, not just text.
Manually pick any model from a live OpenRouter catalog for images, video clips and voice — the lists refresh automatically every 24 hours, with a one-click refresh and per-channel defaults.
Channel consistency: a channel's brand voice (tone, audience, narrator persona) and brand bible are now woven into script and storyboard generation, so every video stays on-brand.
Cleaner status: the bottom-right backend/cloud indicator is a compact button that no longer overlaps the page. Windows + macOS (Apple Silicon & Intel) builds — still unsigned (SmartScreen / Gatekeeper prompts expected).

v1.0.73 — Security & code-quality hardening

June 25, 2026

Security: a DNS-rebinding guard on the local backend, a stricter Content-Security-Policy, and all known dependency vulnerabilities cleared — plus a locked dependency set and a single Python toolchain version so every build is reproducible.
Reliability fixes found while making the code fully type-safe: the YouTube-disconnect confirmation, the Profit-Mode cap message, and the AI Director's “resume” link now read their server responses correctly (a few error messages could previously come through blank or mislabeled).
Under the hood: the whole desktop codebase is now type-safe (zero `any`, enforced in CI) and the large API client was split into focused modules — no behavior change, just a steadier base for future releases.
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.72 — Startup + YouTube connect fixes

June 21, 2026

Fixed a “backend didn’t start” error some users saw on the first launch right after an update — the health check now allows up to 90s (a freshly-installed build can be slow while antivirus scans it) instead of giving up at 30s.
Connect YouTube needs zero credentials again: token refresh, account health, and upload now use the built-in verified Google client as a fallback, so you’re no longer wrongly asked to paste an OAuth client ID / secret.
Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.71 — Security hardening + reliability

June 21, 2026

The app window now runs in the Chromium sandbox and external links are restricted to http(s) with an off-origin navigation guard — defense-in-depth on top of the existing context isolation.
Fixed a Windows logging crash: the backend now forces UTF-8 console output, so non-ASCII characters in logs no longer spam errors mid-run.
Quality: the full test suite is green (backend 1950, render ship-blocker 74/74, desktop 347), with new CI gates running the desktop suite and a dependency audit on every change.
Windows + macOS (Apple Silicon & Intel) builds. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.70 — Desktop hardening + cloud-sync fix

June 19, 2026

Cloud sync fixed end-to-end — sessions now load correctly (a keychain encoding bug had left the sync daemon idle) and the upload queue no longer wedges, so your channels, scripts and renders reach the web dashboard.
Window fits the screen on first launch, the status popover opens upward, and bulk voiceover / visuals / captions progress now survives navigating between pages (and can't double-start).
Connect YouTube needs zero setup (built-in Google client), a free AI model stays free, Shorts and the Word-Pop / Bounce captions render correctly, and the thumbnail engine gains a per-variant re-roll.
Reliability: crashed renders no longer show a phantom “running” job, calmer error handling, and a branded launch splash. Still unsigned — SmartScreen / Gatekeeper prompts expected.

v1.0.69 — Cross-platform release builds + stability

June 17, 2026

Rolling Windows + macOS release packaging on top of v1.0.65 — automated macOS CI builds (v1.0.66, v1.0.68, v1.0.69) keep both platforms on the same version.
Maintenance and stability updates; the live GitHub Releases list below always has the exact latest build and its assets.

v1.0.65 — Sync + IPC correctness, new web pages

June 14, 2026

Cloud-sync correctness fixes and Electron IPC fixes for a more reliable desktop session.
New marketing pages — a Solutions SEO cluster and the dedicated pricing page — plus CI hardening for steadier releases.

v1.0.64 — $1 paywall app-gate + email & support

June 13, 2026

The $1 license is now enforced in-app with a fail-open gate — it never traps a signed-in user, and only shows the paywall on a confirmed-unpaid account.
New email system (Resend): payment receipt, welcome, and support-ticket acknowledgement/notification, plus a /support ticket center on the website.
Bundles the v1.0.63 stacked fixes — Jamendo music client, toast z-index above drawers, status-beacon glow, and AI-config handling.

v1.0.62 — AI Router gets a control panel — pick your mode, see your spend

June 10, 2026

New AI Router screen: choose how models are picked — Max Quality, Balanced, Cost Saver, or Speed — and optionally pin a specific model for any single job.
New AI Analytics dashboard: cost-by-model chart, model reliability, failover rate, and spend-by-task, so you can see exactly where your AI budget goes.
The smart routing engine added in v1.0.61 is now fully steerable from inside the app instead of running silently on its default.

v1.0.61 — AI Router v2 — smarter model selection with deep failover

June 10, 2026

Every brain feature now routes through a single AI engine that picks the best model for each task — cheap models for quick jobs, stronger models where quality matters — across your Anthropic and OpenRouter keys.
N-deep failover: if a provider errors, hits a budget cap, or a model is blocked by a guardrail, the router automatically reroutes to the next-best candidate instead of failing.
A live model registry keeps the candidate list current from OpenRouter, with a bundled offline snapshot so model selection still works without a network round-trip.

v1.0.60 — Works on Anthropic OR OpenRouter — with failover

June 6, 2026

Every AI feature now works on whichever brain key you have. AI Motion Suggestions and AI music no longer require an Anthropic key — they use your OpenRouter key too.
Brain failover: if you set both keys, Claude is primary and the app automatically falls back to OpenRouter (or vice-versa) when one provider errors or your account blocks a model.
Clearer errors: when OpenRouter blocks a model (guardrail / data policy), the app tells you exactly that and how to fix it — for scripts and image/voice/video. Image generation also retries once on a transient error.

v1.0.59 — Claude Opus 4.8 (Fast) is the new brain default

June 6, 2026

Script writing — and every text step — on the OpenRouter path now defaults to Anthropic's top model, Claude Opus 4.8 (Fast), for the highest-quality output.
It's a paid model, so your OpenRouter key needs credits and a data policy that allows the Anthropic provider. Prefer free or cheaper? Pick any model per-generation via Provider → OpenRouter → “Use a custom model id…”.

v1.0.58 — Analytics channel summary + key management

June 6, 2026

Analytics now shows an overall Channel summary — subscribers, lifetime views and total videos — for every connected channel, including channels whose videos all pre-date TubeForge.
Settings → AI: a stuck API key left over from an older version now deletes cleanly, plus a new “Delete all keys” button wipes every stored key at once.
Script page polish — the script content sits above the production checklist, and the Delete/Save bar no longer floats over the page while scrolling. (The account-delete error on the website is fixed too.)

v1.0.57 — Captions in the pipeline + sturdier updater

June 6, 2026

Captions are now a first-class pipeline step with their own sidebar entry and landing page; the guided flow runs Visuals → Captions → Compose.
Caption timing is generated per scene with live progress, so a long video no longer gets stuck with “untimed” captions.
The updater logs and retries its launch check, the database-clean action gives clear toasts, and Diagnostics gains a full “Erase all data” factory-reset.

v1.0.56 — In-app auto-update

June 6, 2026

An Update button now lives in the top bar; the app checks for updates on launch and every few hours.
On Windows a new release downloads in the background and prompts “Relaunch to update.” macOS shows an update-available prompt with a download link (in-place Mac auto-update is gated on code-signing).

v1.0.55 — AV1 / H.265 codec tier + animated captions

June 5, 2026

The adaptive render core auto-prefers AV1 / H.265 when your machine can encode them at quality (with the libx264 floor as a guarantee) — noticeably smaller files at the same quality.
New optional animated-captions step burns word-synced subtitles in 5 styles (Karaoke, Word Pop, Typewriter, Fade-Up, Bounce), aligned fully offline by a bundled Whisper model — no cloud, no key.

v1.0.54 — Encoder picker + multi-clip per scene

June 5, 2026

Manual encoder picker in Render Diagnostics — auto-best stays the default, but click “Use” on any verified encoder to override (or “Auto (fastest)” to reset). The choice persists across hardware re-probes and drives renders directly.
Multiple selectable clips per scene — each is its own card in the Asset Strip (Clip 1 / 2 / 3) with a per-clip delete and a segment-count label, and clips weight the proportional split by their true duration.
“Regenerate” is now “Generate another clip” — it appends a new take and keeps the old one instead of replacing it.

v1.0.53 — Adaptive Rendering Engine

June 5, 2026

A device-adaptive engine that detects your GPU (NVIDIA / Intel / AMD), CPU, RAM and ffmpeg hwaccels, then benchmarks and auto-selects the fastest stable encoder — NVENC → QSV → AMF → CPU, with a guaranteed libx264 floor.
Real 1-frame validation renders mark each encoder verified or failed, so a machine that only claims to support an encoder never gets one that breaks mid-render.
Render-time fail-open: if a GPU encode misbehaves, it falls back to the byte-identical libx264 command — a render never fails because of your GPU. New GET /render/diagnostics + a Render Diagnostics panel show what each device detected and selected.

v1.0.52 — Flexible Render Studio

June 5, 2026

A per-scene asset studio on the Visuals page: every generated clip is a playable mini-preview, and you check or uncheck the image and each clip per scene.
Mix images and clips across scenes, sequence multiple clips within one scene, and drag to reorder — the selected assets are fit proportionally to each scene’s narration.
Fully additive — untouched scenes render exactly as before, with a “Reset to auto” escape hatch.

v1.0.51 — Render audio continuity

June 4, 2026

Fixed narration + background music dropping out in the final render (especially after a clip scene) by pinning identical audio params on every scene-clip branch, so the concatenation keeps one continuous voiceover + music track.
Speed-ramped scenes now keep their audio time-stretched in sync instead of silently dropping it.

v1.0.50 — OpenRouter voice + video hotfixes

June 4, 2026

Switched the voice model to a working Grok voice (voices Eve / Ara / Rex / Sal / Leo) after the previous model 400’d.
AI video clip downloads now poll until the signed URLs appear (OpenRouter is eventually-consistent) and the job poll budget was raised to ~15 minutes.

v1.0.49 — AI provider redesign → OpenRouter

June 4, 2026

Replaced the multi-provider AI plumbing with a single unified OpenRouter client driving text, image, video and voice — with role-based API keys and a role-card Settings UI to paste each.
Added a guided “make a video” director that runs the whole pipeline end-to-end for new users, plus a per-scene AI-video cost guard under Profit Mode.

v1.0.48 — AI-clip download validation

June 4, 2026

AI video clip downloads now validate the bytes are a real video before saving, so an unreadable clip can no longer slip through and fail the render.

Published releases

pulled live from GitHub Releases

1.1.0

July 11, 2026

TubeForge AI Studio v1.1.0

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

1.0.99

July 9, 2026

TubeForge AI Studio v1.0.99

Cloud Video quality-hardening release. This is *measured* quality hardening for the self-hosted Cloud Video pipeline — steady, verified improvements, not a claim of perfection.

What's in it:

•Grok first-frame generation for Cloud Video — both Economy and Premium tiers now seed each clip from an aspect-matched OpenRouter Grok still.
•Prompt token clamp — keeps the video prompt within the model's text-encoder budget so the motion direction and on-frame guidance are never silently truncated.
•First-frame provenance — each Cloud Video job records which model produced its first frame (local-only).
•Safer image-prompt sanitization — the hidden-init image prompt is cleaned of text-inducing instructions (title cards / captions / credits) before it reaches the image model, so they can't be painted into the opening frame.
•Gateway prompt logging — privacy-safe structured logging on the owner gateway (token estimates + a prompt hash; no raw prompts).
•First-frame QA scaffolding — observe-only, disabled by default.

Downloads (below): Windows .exe, macOS .dmg (Apple Silicon + Intel).

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

1.0.98

July 9, 2026

View on GitHub

1.0.97

July 9, 2026

View on GitHub

v1.0.94

Pre-releaseJuly 8, 2026

TubeForge AI Studio v1.0.94

The paid-account lockout fix plus a cleaner AI setup and a Cloud-Video-first Visuals flow.

Fixed

•Paid-account lockout (for good). The paid-check no longer competes for your single-use sign-in token, so a paid account stays unlocked. An expired sign-in now sends you to Sign in again — never a false paywall.

New

•Simpler AI setup. The AI Keys page leads with one-click Connect OpenRouter (keys every role at once). The old per-role paste boxes are gone; the free-provider tokens (Pollinations, Google Imagen, Jamendo) stay for those who want them.
•Cloud Video, front and centre. Each scene in Visuals now opens on TubeForge Cloud Video with an Economy / Premium picker, a live credit cost, and your balance — all before you hit Generate. No API key needed. A still image is one toggle away, and OpenRouter/Grok is a quiet "use my own key" option.

Includes the $15 Starter Bundle + 500 Cloud credits, free local rendering, and the hybrid model from v1.0.91+.

Downloads (below): Windows .exe; macOS .dmg (Apple Silicon + Intel) attach shortly.

> Unsigned build — SmartScreen / Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.93

Pre-releaseJuly 8, 2026

TubeForge AI Studio v1.0.93

Hotfix over v1.0.91/1.0.92 — fixes the paid-account lockout for good.

•The unlock screen could wrongly appear for accounts that had already paid. Root cause: the desktop's paid-check refreshed your sign-in token on every poll, which burned the single-use token and left the app unable to verify you → a false "Unlock" wall.
•Now the paid-check no longer competes for the token (the sync worker owns it), and an expired sign-in sends you to Sign in again — never a false paywall.

Everything from v1.0.91/1.0.92 is included: TubeForge Cloud Video (Economy/Premium, no API key), the $15 Starter Bundle + 500 Cloud credits, OpenRouter/Grok BYOK fallback, free local rendering.

Downloads (below): Windows .exe; macOS .dmg (Apple Silicon + Intel) attach shortly.

> Unsigned build — SmartScreen / Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.92

Pre-releaseJuly 8, 2026

TubeForge AI Studio v1.0.92

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.91

July 8, 2026

TubeForge AI Studio v1.0.91

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.89

July 4, 2026

TubeForge AI Studio v1.0.89

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.88

July 4, 2026

TubeForge now requires a signed-in, licensed account to open the app. If you bought a license, sign in with that account and it unlocks — existing license holders are unaffected. This release also hardens payment handling and account linking (reliability + security).

Windows + macOS (Apple Silicon & Intel), all with the same licensing.

Builds are unsigned — Windows SmartScreen / macOS Gatekeeper warnings are expected (choose "Run anyway" / right-click Open).

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.87

July 4, 2026

TubeForge AI Studio v1.0.87

A major reliability, security, and cost-control hardening release — dozens of fixes from a deep internal audit, each verified with automated tests. The render engine's byte-identical output guarantee held through every change.

💵 Cost control

•The Profit-Mode spend cap now correctly enforces on image, voice, and thumbnail generation (it could previously be exceeded).
•Cancelling an AI-video job now stops the spend instead of finishing and billing every remaining clip.
•Failed usage records no longer silently vanish from your AI Bill or the cap.

🔒 Privacy & security

•Uploads you set to private/unlisted are no longer accidentally published public when scheduled.
•Fixed a case where the same render could be uploaded to YouTube twice.
•Your AI API key is now sent only to the AI provider, never to download URLs it returns.
•Hardened the auto-update feed (SSRF), OAuth sign-in (CSRF protection), and app-internal messaging.
•Backups are now a consistent database snapshot — a backup taken while you're working can't be corrupt.

🔧 Reliability

•Large YouTube uploads no longer fail near the end (the access token refreshes mid-upload).
•Crashed AI-video jobs are recovered on restart instead of getting stuck forever.
•Cloud sync reconnects automatically after a Wi-Fi drop.
•Fixed several memory/handle leaks and UI hangs.

---

Downloads (below): Windows .exe; macOS .dmg for Apple Silicon and Intel.

> ⚠️ Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch (Windows: *More info → Run anyway*; macOS: right-click → *Open*).

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.86

July 3, 2026

TubeForge AI Studio v1.0.86

Adversarial deep audit (docs/audit/DEEP-AUDIT-2026-07-03.md, 124 findings) remediation.

Render pipeline:

•H1: layered render no longer degrades to a black clip (undefined successful_layers

-> len(processed)); the layered branch is now covered by an EXECUTING test.

•H2: mixed video-codec concat (adaptive AV1/HEVC image scene + hardcoded-libx264

composite/layered/black scene) re-encodes via the concat filter instead of a

corrupt -c copy; uniform sets stay byte-identical.

•H6: a completed (paid) AI-video clip now wins over the adaptive image encode

(dispatch reorder) instead of being silently discarded.

Cost / AI:

•H3: the live brain path (openrouter.py::complete) now retries transient 429/5xx

(the shipped retry lived in a zero-caller function).

•H4: billing POSTs retry only the safe 429 (not ambiguous 5xx that may have

charged) + send a stable Idempotency-Key -> no double-charge.

•Failover: back off before the single transient retry (was instant re-hit).

Cloud sync:

•H5: cloud->local DELETE apply reads identity from old_record (was new-only ->

deletes dropped). Companion migration sets REPLICA IDENTITY FULL on the 12

mirror tables (applied to prod).

Security:

•C1: profiles entitlement columns locked down (REVOKE + trigger + fail-closed

verify_user_status) — paywall self-grant closed (migration applied to prod).

•C2: auto-updater no longer silently installs on quit (autoInstallOnAppQuit=false);

code-signing cert still required for full close.

Gates: ship-blocker 74/74 green throughout; 18 new regression tests that EXECUTE

the render/sync branches the string-only ship-blocker never touched.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.85

July 2, 2026

TubeForge AI Studio v1.0.85

User-reported blocker: a 1,698-word (~11 min) script produced a storyboard

of 24 scenes totalling ~430 words / ~2:24 — the model summarised the script

instead of cutting it verbatim, and the server accepted it silently. The

script-length fix (v1.0.84) was being undone one stage later.

•storyboard_generator prompt v2: states the script's exact word count and

makes full verbatim coverage a hard requirement ("you are cutting the

script into pieces, NOT rewriting it")

•server-side coverage guarantee in routers/scenes.py: after parsing, the

scenes' total narration words are compared to the script's; outside the

0.90–1.10 band the narration is re-cut deterministically from the ACTUAL

script at sentence boundaries (_split_script_verbatim, ~equal words per

scene), keeping the model's scene count and image prompts paired in

order — coverage is exact by construction, for any model

•scene durations now always derive from the final narration at 150 wpm

(words * 0.4s), so the storyboard total always matches the script length

•covers both the manual Storyboard page and the guided director (it calls

the same endpoint); 502 with a clear message when the model returns no

usable scenes

Tests: +5 (splitter verbatim/balance/edges; endpoint repairs a summarised

storyboard to exact coverage; verbatim storyboards pass through untouched);

template test updated for the v2 variable.

Gates: full backend suite 2035 passed / 6 skipped / 0 failed (ship-blocker

74/74), marketing tsc clean, changelog card verified rendering.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.84

July 2, 2026

TubeForge AI Studio v1.0.84

User-reported from a live run: a "15 min" idea produced a 357-word script

(~2 minutes of narration), and idea titles were lowercase fragments with

the duration baked in ("night sky surrender — 15 min") that flowed

unchanged through script → render → upload.

Script length guarantee:

•script_generator prompt v2 makes the word target a HARD requirement

with a length-planning scaffold (hook ~10% / body ~80% / resolution ~10%)

•new _extend_script_to_target() loop in routers/scripts.py: when the

draft lands under 75% of the target, the SAME model (pinned via

overrides so the voice can't shift) writes seamless continuations —

"continue from where it stops, no repetition, no summary, no new hook" —

up to 3 bounded rounds; every extension bills through the router

(template script_extender); fail-open keeps the draft on any error;

response/log cost is the true total (draft + extensions)

•benefits the guided director automatically (it calls the same endpoint)

Real YouTube titles:

•idea_generator prompt v3: titles must be complete, Title Case,

payoff-driven YouTube titles; duration may NEVER appear in the title

(it belongs in estimated_length_min) — with explicit BAD/GOOD examples

•belt-and-suspenders _sanitize_idea_title() in routers/ideas.py strips

trailing duration suffixes ("— 15 min", "(12 minutes)", "~30 mins",

": 45 min") and capitalizes lowercase fragments for ANY model, while

leaving legit in-title durations ("The 10 Minute Rule ...") untouched

Tests: +6 (sanitizer strip/preserve; extend loop reaches target, no-ops

when long enough, bounded + fail-open on empty additions).

Gates: full backend suite 2030 passed / 6 skipped / 0 failed (ship-blocker

74/74), marketing tsc clean.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.83

July 2, 2026

TubeForge AI Studio v1.0.83

Known-issues panel (Diagnostics):

•services/known_issues.py tails the local structured log, keeps

warning/error lines, groups by event with counts + latest detail +

last-seen; GET /diagnostics/issues; KnownIssuesPanel card with level

badges, counts, refresh, and a green all-clear state. Local-only —

nothing leaves the machine. 3 tests.

Graphical brain (visual memory — the channel brain learns what WINS):

•brain_memory.record_visual_win() formats a "[visual] Winning thumbnail

style (evidence): angle — prompt excerpt" note into the channel memory

•captured at the two moments a thumbnail PROVES itself: the user applies

an A/B comparison winner (thumbnail_compare.apply_comparison), and a

rotation finalizes with a measured CTR lift (thumbnail_rotation) —

the strongest visual evidence the app ever gets

•notes flow into every future thumbnail/image prompt generation via the

existing inject_brand_bible chokepoint, so each channel's visuals

converge on what its audience demonstrably clicks; fail-open everywhere

•+1 test (format/store/no-op-on-empty)

Version bump + README latest section + website changelog card.

Gates: full backend suite 2025 passed / 6 skipped / 0 failed (ship-blocker

74/74), FE tsc clean + 347 vitest, marketing tsc clean.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.82

July 2, 2026

TubeForge AI Studio v1.0.82

COMPLETE Claude/direct-Anthropic removal (user directive: one AI gateway):

•deleted services/ai/claude.py + ai_router/providers/text_anthropic.py
•factory is OpenRouter-only; legacy force="claude" prefs map transparently

to Claude-via-OpenRouter (anthropic/claude-sonnet-4.6); brain_available()

= OPENROUTER_TEXT_KEY only

•registry: direct-anthropic curated seeds removed (OpenRouter anthropic/*

seeds remain); stray legacy "claude" refs resolve to the OpenRouter text

adapter; _has_key drops the anthropic mapping

•secrets: anthropic_api_key removed from the allowlist and added to

LEGACY_SECRETS — any stored Anthropic key is purged from the keychain on

startup; anthropic dropped from requirements.txt + requirements.lock

•providers_roles/setup/ai routers + ALL frontend surfaces cleaned: Settings

key row, AIHub, SetupWizard (single OpenRouter option), Dashboard status,

ProviderPicker (claude option removed; type value kept for legacy prefs),

and every "Claude" copy string across 29 files

•17 test files re-locked to OpenRouter-only expectations

One-click Connect OpenRouter (OAuth PKCE):

•services/openrouter_oauth.py — loopback + code→key exchange, key saved to

all four role slots; POST/GET /providers/openrouter/oauth endpoints;

AIHub "Connect OpenRouter" card with live status polling; 4 tests

Spend-consent rules (silent premium billing killed):

•modes steer DOWN only (role_models); hidden i2v init frame is ALWAYS the

free provider (video_jobs); brain failover cost ceiling — never fail over

to a pricier model (selection._cap_failover_cost); tests re-locked + 3 new

Smooth motion:

•2x supersample before zoompan (motion/filters._ZP_SS) ends the Ken Burns

shimmer/vibration — same proven fix Shorts always had; ship-blocker

deliberately re-locked (scale/crop 3840x2160, output s= unchanged);

verified with a real ffmpeg encode

Channel brain memory (self-improvement loop, v1):

•brain_memory_notes table + services/brain_memory.py (dedupe, 40-note cap,

fail-open); top-3 Improve suggestions recorded per channel and injected

into EVERY brain system prompt via inject_brand_bible — each analyzed

video sharpens future generations; 4 tests

Plus: 22+ frontend per-call timeouts on long AI generations (kills the fake

"timed out" errors), fake-timeout copy cleanup, README + website changelog.

Gates: full backend suite green after re-locks (2020+1 passed / 0 failed,

ship-blocker 74/74), FE tsc clean + 347 vitest + 0 lint errors, marketing

tsc clean.

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.81

July 2, 2026

TubeForge AI Studio v1.0.81

User-reported (with OpenRouter log evidence): a 29-scene storyboard is one

~16k-output-token completion at ~155 tok/s ≈ 2 minutes of pure generation.

The app's flat 120s axios timeout fired, showed "This took too long and

timed out. Try again." — while the backend finished fine and the scenes

appeared on the next refetch. Same false-failure hit "Suggest motion for

all scenes" and every other long single-request AI operation.

Fix: per-call timeout overrides on all known-long single-request AI calls

(frontend/src/lib/api-methods.ts):

•600s: generateStoryboard, generateScript, aiSuggestMotion,

upgradeAllImagePrompts, generateAllThumbnailImages, generateStarterChannel

•300s: generateIdeas, analyzeRetention, generateShorts, generateHooks,

improveBible, createProductionPlan, runImproveAnalysis, runShortImprove

Short per-scene ops keep the 120s default (a hung call should still die).

Also replaced the stale "mock provider can't be used here" notes in the

Storyboard / Ideas / Analytics dialogs (the mock system was deleted in

v1.0.49) with accurate copy that sets expectations for generation time.

Version bump (backend/frontend/marketing) + README + website changelog.

Gates: FE tsc clean + 347 vitest; marketing tsc clean.

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.80

July 2, 2026

TubeForge AI Studio v1.0.80

Version bump (backend/frontend/marketing) + README latest section +

website changelog highlight card. Ships bd54eef: the guided director

now waits for enqueued AI-video clips before submitting the render

(previously the render raced them and shipped stills while the paid

clips generated uselessly), fail-open on timeout/failure, and the

build cost includes the real AI-video spend.

Gates: full backend suite 2011 passed / 0 failed (ship-blocker 74/74),

director suite 7/7, marketing tsc clean + 142 vitest.

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.79

July 1, 2026

TubeForge AI Studio v1.0.79

Version bump (backend/frontend/marketing) + README latest-release section.

Dependencies refreshed across the desktop app:

•backend: in-range upgrade + requirements.lock regenerated via

pip-compile --upgrade (sqlalchemy 2.0.51, anthropic 0.115.1,

supabase client family 2.31, cryptography 48, uvicorn 0.49, pywin32 312)

•frontend: npm update in-range, 0 runtime vulnerabilities

(electron/electron-builder/ffmpeg-static stay deliberately exact-pinned;

ffmpeg-static is render-critical and never bumps casually)

Gates on the upgraded tree: full backend suite 2009 passed / 6 skipped / 0

failed (ship-blocker 74/74 included), FE tsc clean + 347 vitest.

Ships the 980da2d deep-audit fix batch to users (accurate AI Bill for

image/voice, Profit-Mode cap can't be bypassed by the hidden i2v image,

no more slow-motion chained clips, mid-chain clip salvage, clear

brain-not-configured errors, validated AI Mode).

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub

v1.0.78

June 25, 2026

TubeForge AI Studio v1.0.78

Captions:

•Timing reliability: scene_word_timings defaults to instant synthetic spread;

whisper is opt-in (TUBEFORGE_CAPTION_WHISPER=1). Whisper was the default and

hung/timed out per scene (~minutes), so the timing step always failed.

•Preset preview now matches the burn (real bundled Anton/Inter fonts + matched

animation degrees); ass_builder.build_scene_ass gains a play_res param.

•Shorts captions now use the parent channel's chosen preset (animated .ass).

Shorts render overhaul:

•Length-based multi-image 9:16 visuals (ceil(dur/5s), bounded 12) via cheapest

provider, Ken-Burns animated, concatenated, voice+captions muxed, channel

default music auto-mixed. Legacy single-image path is the fail-open fallback.

•New POST /shorts/{id}/thumbnails/from-video: regenerates the parent video

thumbnail concept vertically + burns the Short title (overlay_short_headline),

surfaced as "Match video thumbnail" in ShortThumbnailModal.

•short_metadata prompt -> v2 (viral titles/description/hashtags).

Gates: ship-blocker 74/74; caption+shorts sweep 235 green incl. new

test_caption_overhaul (9) + test_shorts_overhaul (8, incl. real e2e

render_short_sync); FE tsc clean.

Downloads (below): Windows .exe, macOS .dmg for Apple Silicon + Intel.

> Builds are unsigned — Windows SmartScreen / macOS Gatekeeper will prompt on first launch.

Full changelog: https://www.tubefrogeai.app/changelog

View on GitHub