Magic Master News
Release 3.14: livelier sound at streaming loudness, breath control and AI advice
Release 3.14 is about livelier sound and honest hints.
What's new
- Livelier at streaming loudness. At −13…−16 LUFS targets multiband protection no longer over-squeezes a track: the threshold comes from the track itself. Same loudness, more dynamics. Loud targets (EDM, rock) work as before.
- Breath control. Loud breaths between phrases are brought down to a natural level, not cut out. For podcasts, audiobooks and voice-over — in the restoration settings.
- AI advice for your track. The genre hint and the advice are one block now and are based on what is measured in the file.
- Gentler AI-trace cleanup. Pitch is adjusted only when needed, and after mastering you can clean the same track right away — no token charged with an account.
- Small things. Already Loud recovers loudness into free headroom; a “song or speech?” hint; another format right at the result; free runs left shown in advance.
Release 3.13: clipping repair, downloads that finish and honest stems
Release 3.13 is about getting your result to you intact: no clipping in the sound, no downloads that break off, and no archives lost because a tab was closed.
What's new
- Clipping repair. Mastering finds flattened peaks at any level — even in a track turned down afterwards or saved as MP3 — and rebuilds their shape. Clipping becomes much softer. Part of regular mastering, no extra charge.
- Free clipping check on the De-clipper page — no account, no tokens.
- Downloads that finish. The browser saves a large WAV itself, the ↓ button in your profile history works again, and the file keeps your name.
- Honest stems. You see how long is left; you can close the tab — the archive stays in your profile for 24 hours. A second click no longer charges a second token.
- Small things. A site update no longer resets the chosen file, and the password reset link also arrives in Telegram if it is linked.
Details in the article “Magic Master 3.13”.
Release 3.12: Already Loud keeps dynamics, AI clean-up removes the fingerprint
Release 3.12 is about sound that behaves more honestly.
What's new
- Already Loud no longer squashes a dynamic track. The target yields to the material: loudness is set so peaks stay clear of the ceiling, never quieter than −16 LUFS. A dense track still gets −12.
- AI clean-up removes the generator's spectral fingerprint that open detectors use to recognise a track. The change is not audible.
- Next track. A button on the result sends the next track with the same settings.
- Small things. Jazz and classical without an added hall, click removal without new peaks, and changing the download format works even after a server restart.
Details in the article “Magic Master 3.12”.
Release 3.11: AMR and OGG, reliable Telegram delivery
Release 3.11 is about making sure a file gets through: accepted, processed and delivered.
What's new
- New formats. WAV, MP3, FLAC, M4A, OGG and AMR — phone voice recordings — are accepted. Long RF64 WAV files and files with unusual headers are no longer rejected: if there is audio inside, it is processed.
- Telegram without losses. A large master arrives in seconds, and if Telegram refuses the file, the bot immediately sends a download link valid for 24 hours.
- Faster AI advice. An answer in a few seconds; tracks of a batch wait for their turn instead of a “not enough memory” rejection.
- Complaint tracks only with your consent. Kept for 30 days if you tick the box in the rating form; deleted together with your account.
- Small things. Sign in with Google and a clear “file expired” mark in history.
Details in the article “Magic Master 3.11”.
5 000 tracks mastered with no signup
The counter of free, no-signup masterings has passed 5 000.
5 062 tracks were processed by guests — people who simply dropped a file on the page and picked up a finished master a couple of minutes later. No account, no card.
What every guest gets
- one free mastering a month — same engine as the paid plans;
- 32-bit float processing; WAV 16/24 bit, FLAC, MP3, OPUS and AAC exports;
- AI genre detection and the mastering assistant's recommendations.
With an account you get more: a mastering every day, work history, tokens for the free tools and delivery straight to Telegram.
Thank you for trusting us with your tracks. 🎧
Master a Suno track from Claude or ChatGPT in one message
Your AI agent can now master on its own. Add Magic Master as a connector (one address — magicmaster.pro/mcp, nothing to install), give it a link to the track and write one message: "check the loudness, remove the Suno fingerprint, master for Spotify and show me the passport".
What the agent does without you:
• measures LUFS, true peak, duration and channel correlation — free;
• removes the Suno or Udio digital fingerprint without touching loudness — free;
• picks a preset and a target for the destination (Spotify −14 LUFS, true peak ≤ −1 dBTP) and starts the master — one token;
• 20–60 seconds later shows the mastering passport right in the chat: loudness before and after, true peak, dynamics, distance from the target.Works in claude.ai, Claude Desktop, ChatGPT (developer mode), Cursor, VS Code and Claude Code. One-click install buttons for Cursor and VS Code are on magicmaster.pro/mcp; recipes for every other client are there and on GitHub: github.com/denisok-ai/magicmaster-mcp.
A registered agent gets three trial masters with no card. The agent cannot pay by itself — it prepares the checkout, you confirm.
Step-by-step guide with an example message: magicmaster.pro/en/blog/master-suno-track-from-claude-mcp
Release 3.9: peaks are no longer clipped mid-chain
Version 3.9 is out. It is about sound: four changes to the processing chain, each of which came out of a specific complaint and each measured before it was switched on.
• Honest dynamics. The chain used to hard-clip peaks between modules, so every kick lost its attack before reaching the final limiter. Peaks now travel intact and loudness is set by the limiter alone. Measured on the control file: crest 11.41 to 12.41 dB, with loudness and ceiling unchanged (−12.05 LUFS against a −12 target, −1.01 dBTP against a −1 ceiling). Your track did not get louder or quieter — what changed is HOW loudness is reached. Nothing to configure, and it applies to every preset.
• The unlimiter got more precise. Its threshold is now relative rather than absolute: derived from the material instead of being the same for a quiet recording and a brickwalled master. Tested on a corpus of six real tracks: dynamics improved on all six, zero regressions.
• Pauses after stem separation are quiet now. The bubbling is not clicks but a floating residue of other instruments: only 8 per cent of the residual energy matches the isolated-peak criterion, and the old cleanup never touched the other 92. The residual gate pushes deep pauses down 15 dB while phrase boundaries drop 0.2 dB and vocals 0.125 dB.
• Tone correction leaves dense material alone. On saturated tracks it made things worse — all four user ratings of that feature were one star. It now steps aside on such input and works as before where it helps.
• A door for AI agents. If you work through an assistant, Magic Master is now available to it directly: an MCP server at /mcp, OAuth 2.1 sign-in and trial tokens. The assistant sends a track for mastering and returns the quality passport as a card inside your chat. Nothing changes for using the website.
Full write-up with the numbers: https://magicmaster.pro/blog/magic-master-3-9-whats-new
We are watching bass dynamics separately: the change also affects what comes after it — the exciter and imager now receive a hot signal. If you hear a difference, better or worse, the rating under your mastering result reaches us and gets read.
Magic Master now works with AI agents
Mastering is now a tool your AI agent calls on its own.
We put up an MCP server at https://magicmaster.pro/mcp. The Model Context Protocol is how agents in Claude, ChatGPT, Cursor and other clients connect external tools. One address, nothing to install. The server is listed in the official MCP registry as pro.magicmaster/mastering.
What an agent can do without you in the loop:
• measure a track — LUFS, true peak, duration, correlation, a genre hint;
• strip the Suno and Udio digital fingerprint without touching loudness;
• master for a destination: 24 presets, a LUFS target, six export formats;
• show the result passport right in the conversation — loudness before and after, true peak, range and how far it landed from the target.Analysis and fingerprint cleanup are free. A full PRO master costs one token, same as in the app. Register your agent once and get an API key plus three trial masters — no card.
The rules an agent should follow live in one file: https://magicmaster.pro/agents.md — what is free, what costs a token, how to pick a preset, and what to do with each error. The machine-readable manifest is https://magicmaster.pro/agent.json, and its prices and limits come from live settings rather than from text someone forgot to update.
Details and connection examples: https://magicmaster.pro/mcp
This week's fixes: sound, queue and payment reliability
A week without flashy features, but with fixes you can actually feel. Being straight with you: these are not new buttons — these are places where the service was letting you down.
🔊 The "No limiter" preset no longer squashes your track
It did the exact opposite of what it promised: you ended up with less dynamics than after regular mastering. The dynamics restorer sat twice as far as the point where its effect already saturates — beyond that point only side effects kept growing.
The setting is back in its working range. In measurements, crest factor rose from 7.1 to 9.4 on a tonal signal and from 12.1 to 14.6 on a beat. In plain words: transients breathe again.
⏳ Batch processing shows the queue instead of "0%"
When you upload more files than there are free slots, the extras wait. The batch list used to show a frozen "0%" for those — indistinguishable from a hang. One of you refreshed the page 260 times before giving up; every file had actually been processed.
Now a queued row says it plainly: "⏳ In queue: 3 · ~4 min", with the number of jobs ahead in the tooltip. Waiting no longer looks like breakage.
🎚 Loudness defect under heavy load is gone
At peak hours work is spread across several machines. On the extra nodes some processing settings were not applied, and a master could come out quieter than the target. Settings now travel with the job — the result is the same no matter which machine picks up your track.
⭐ Telegram Stars: payments can no longer vanish silently
There was a rare path where Telegram had already charged the stars while the service finished handling the payment without a single log record. Refunding or crediting by hand was next to impossible. Any such case is now logged and flagged immediately — with the charge id and the amount.
🌍 The English section is reachable now
The language switcher was a button that only worked inside the browser: search engines simply could not find the English pages. It is a normal link now, and the English section is open to people and to search alike.
If you ran into any of this — thank you for putting up with it. Almost everything on this list came from your complaints and from digging through logs, not from our imagination.
New: split a track into stems — vocals, bass, drums
Any track can now be split into four layers right in the browser: vocals, drums, bass, other.
🎛 How to try
- Upload a track in the app and press "Split into stems" in the PRO modules block.
- Processing ≈1.5× the track length; under load you see the queue and an ETA.
- The zip downloads automatically, and with Telegram linked it arrives in your chat too.
Separation runs on the Demucs neural network. Price — 1 token, charged only after success: failures cost nothing.
Use it for instrumentals and acapellas from a finished mix, remix material, or learning parts. Details — in the blog and on the feature page.
2 000 tracks mastered with no signup
The counter of free, no-signup masterings has passed 2 000.
2 000 tracks were processed by guests — people who simply dropped a file on the page and picked up a finished master a couple of minutes later. No account, no card.
What every guest gets
- one free mastering a month — same engine as the paid plans;
- 32-bit float processing; WAV 16/24 bit, FLAC, MP3, OPUS and AAC exports;
- AI genre detection and the mastering assistant's recommendations.
With an account you get more: a mastering every day, work history, tokens for the free tools and delivery straight to Telegram.
Thank you for trusting us with your tracks. 🎧
Update 3.8: room echo removal and an honest queue
Three highlights of this week are live in the app.
🚿 De-reverb: no more "bathroom" sound
The new restoration PRO module reduces roomy echo: direct sound and attack are preserved while the late reverb tail fades. One toggle, one strength slider. Deep dive — in the blog.
⏳ An honest queue with ETA
When the server is busy you see your position and start time: "Queued 3rd, starting in ~4 min". The forecast uses real processing speed.
🤖 The assistant got closer
- The track breakdown appears right after upload — no extra buttons.
- In Telegram, the finished master arrives with an assistant card: verdict, measures, tonal curve.
- Genre detection progress is pinned to the top, and past 95% you see an honest "finalizing" stage.
Full details in the 3.8 overview. Try it on your track — free for guests, no signup.
Mastering quality update
We’ve upgraded the mastering engine — the result is now cleaner and more consistent. What’s new:
🎚 Smart tonal balancing. The “master assistant” gently evens out the frequency balance so your track sits right on both headphones and speakers.
🎧 Reference stereo matching. Stereo width now adapts to your reference track — closer to the sound of the releases you love.
📱 Mono-compatibility protection. Your mix no longer collapses in mono — phones, Bluetooth speakers, club systems.
✨ “AI traces removed” badge. Shows when processing has cleaned up the artifacts typical of AI-generated audio.
We deliberately left loudness and the overall character untouched — we just made the result cleaner and more predictable. Upload a track and compare.
Magic Master 3.6.0 — reliable and faster on long tracks
What's new in 3.6
- Fixed the false "Mastering error" — long tracks (5+ minutes) with heavy modules legitimately take up to ~10 minutes; the progress stream used to hard-cap at 600s and fabricate an error even though the master had completed. The ceiling is now 20 minutes, with a heartbeat, and on timeout it falls back to the authoritative job status
- No need to re-run and waste a token — processing reaches the finish line reliably
- Faster spectral processing — denoise, AI-separation cleanup and vocal HF-resonance cleanup process stereo channels in parallel; up to ~1.3× end-to-end on tracks with cleanup/denoise (27.3s → 21.0s in testing)
- Byte-identical sound — the result is identical to before, verified with golden SHA-256 hashes on real Suno tracks; the sound does not change at all
24-bit WAV, 800 MB batches and fair limits
One update covering both the things that got in the way and the things you asked for.
• 24-bit WAV export is here. That is the master format for distribution: processing always ran in 32-bit float, but export stepped down to 16-bit with dithering. Now you can take the result without that step. 16-bit stays the default — the file is one and a half times smaller, and on a finished master the two are indistinguishable by ear. Why 24-bit matters anyway, and when you actually need it, is covered in a new blog post, "16 vs 24 bit: which master to deliver".
• Batch processing grew from 240 to 800 MB in total. The old budget was chosen around a proxy limit that has not been there for a while — we re-measured and raised it. We also fixed the pricing table: it claimed Free and Pro had no batch at all, while Free gets 10 files and Pro and Studio get 50 each. The app now shows both the file count and the space used before you upload, instead of failing afterwards.
• A failure on our side no longer eats your attempt. The daily counter counted every run, so a failed master took a whole day away on the free tier: the token came back, the day did not. Now only successful and running jobs count, and guests get their monthly attempt back.
• The blog and news pages work on a phone. On a narrow screen the header hid its links, and the burger menu had never been wired into those pages — the logo was the only way out. Fixed.
• Buttons are readable again. White text on the light half of the gradient measured 3.96 contrast against a 4.5 standard, so the label disappeared. We recalculated and fixed 56 buttons across the site: landing, pricing, wallet, sign-in, referral programme.
• Limit and sign-in protection hardened: a client address can no longer be spoofed with a request header.
Thanks to everyone who writes in about rough edges — half of this list came from your messages.
Tool achievements: progress you can see, tokens counted honestly
We went through how tokens for the free tools actually work. We found two defects, fixed them, and made the progress visible along the way.
• Progress now shows inside the tool itself. Until now you could only learn about achievements on the Tools page. If you arrived from search straight into the loudness analyzer or the calculator, the gamification did not exist for you. Now a strip appears under your result: which achievement is closest, how much is left, and what it pays.
• There is a single progress screen — in your profile, in the dashboard and on the tools hub. It shows how many checks you have run, which tools you used, your streak of consecutive days and how many tokens you have earned.
• An achievement counts for work, not for opening a tab. Four tools out of five marked usage simply when the page loaded. "Explorer" — four different tools — could be earned by flipping through tabs without checking a single track. The event now counts on the real action: analyzing a file, pressing Calculate, answering in the ear trainer.
• "Regular" no longer depends on the time zone. Days were counted in UTC rather than the Moscow calendar, so ten in the evening and one at night fell on the same day. For people who mix at night the three-day achievement built up slower than promised. A day is now a Moscow day.
• We added short rules, "How it works": what earns tokens, why an account is needed and how much you can earn in total.
A reminder: the free tools can earn you 4 mastering tokens. The reward is credited to an account only — guest progress is not saved.
Magic Master 3.3: faster, more accurate, easier to use
One release with everything you asked for in feedback.
• Mastering got faster. The standard preset now processes a four-minute track in 24 seconds instead of 69 — nearly three times quicker. Progress used to stall at 85–90% because the limiter ran three full passes over the track at eight-times oversampling. The redundant passes are gone and the result is identical to the decibel.
• Presets hit the loudness they promise. If the card says −9 LUFS, that is what you get: worst deviation is now about 1 LU instead of 3.5. And one value applies everywhere — app, API and AI hints (Hip-Hop used to promise −13 while the API delivered −10).
• Loud sources no longer go dull. Tracks from Suno and Udio, or exports with a limiter on the master bus, used to get over-compressed: the midrange collapsed, the sound went boxy, the stereo image drifted. Level is now matched before compression, so tonality survives — and quiet tracks are untouched.
• All 23 presets are selectable. Trap, Pop, Rock, Metal, R&B, Country, Afrobeats, K-Pop and Jazz worked in the engine but could not be picked. They are back, and the cards are grouped into Genres, Voice, Character and Special.
• The AI assistant actually works now. It had been down after the provider retired our model and was silently answering from templates. It now reads your track's metrics and explains its choice: "crest factor 6.4 dB — the track is over-compressed, use AI Unlimiter". Auto-mastering got smarter too.
• The English interface is fully translated. 86 elements were still in Russian: batch processing, PRO modules, the request form and result buttons.
• The interface is easier to read. We raised the contrast of cards and muted text in both themes, unified the logo across all pages and fixed its proportions on narrow screens.
Thanks to everyone who leaves detailed feedback after a master — that is what started this work.
The technical write-up: magicmaster.pro/blog/why-my-track-sounds-dull-after-mastering
Major quality update: loudness, sound, presets and fair limits
We ran an independent audit of the service, going through it as an ordinary user would: what you pay for, what you see in the app, and what actually comes out. Here is an honest list of what we found and fixed.
• Mastering no longer crushes loud sources. When a track arrived already loud — exported with a limiter on the master bus, or generated by Suno/Udio — the compressor engaged far deeper than intended: the midrange collapsed, the sound turned dull and distant, the top end got harsh and the stereo image drifted. The cause was ordering: level was matched to the target after compression. Now it happens before. On a brickwalled mix the midrange shift dropped from 3.3 dB to 0.8 dB. Quiet and normal-level tracks are unchanged — staging only attenuates.
• Presets now hit the loudness they promise. Picking EDM at −9 LUFS used to return −11.5, and Metal missed by more than 3 dB, because an overly long limiter release held gain at the track's minimum. The worst deviation is now 1.06 LU instead of 3.46, and 11 of 15 tested presets land within 0.11 LU. Industry distortion tests (SMPTE RP120, aliasing, multitone) confirm the extra loudness came with no added grit.
• One preset, one loudness. The app and the engine advertised different targets for the same genres: the Hip-Hop card promised −13 LUFS while the same preset via the public API delivered −10. The value is now identical everywhere — app, API and AI assistant.
• All 23 presets are selectable in the app. Trap, Pop, Rock, Metal, R&B, Country, Afrobeats, K-Pop and Jazz worked in the engine and the API, but there was no way to pick them in the interface.
• The AI assistant stopped repeating itself. It was two versions behind: it knew 17 presets out of 23 and silently replaced newer ones — Voice, Audiobook, Warm, Bright, Bass and AI Unlimiter — with the default, and carried outdated prices. It now knows the full set, the Russian platforms (VK Music, Yandex Music, Zvuk), current pricing, and picks a preset from your track's actual metrics.
• Fair limits for paying customers. Buying tokens did not technically change the account status, so paying customers were treated as free in many places: no uploads above 100 MB, standard queue instead of priority, API key creation refused, 5 AI requests a day instead of 50, and the limits page reporting zero masterings available with a full wallet. Access is now determined by actual payment.
• Tokens no longer disappear for nothing. A token was deducted before the file was validated, so a "file too large" rejection consumed it without any mastering, and there was no refund at all on the web when processing failed. Deduction now happens after validation and a failed master refunds the token automatically. The Telegram bot refund is fixed too — it used to look at account type rather than what was actually charged.
• We closed the leaks in the other direction. Reference Mastering, upscaling and vocal isolation were open to everyone including guests despite being part of the paid set; stem mastering had no rate limit and auto-mastering deducted no token.
A special thank you to everyone who writes detailed feedback after a master. One report describing a boxy sound with stereo drifting from ear to ear gave us the exact lead that started all of this. If a result disappoints you, hit thumbs down and tell us what is wrong — we read every message.
Full technical breakdown: magicmaster.pro/blog/why-my-track-sounds-dull-after-mastering
Magic Master 3.2: AI Unlimiter, RU platforms, 3x faster
A big update. Highlights:
• AI Unlimiter — opens up over-compressed 'brickwalled' and AI tracks (Suno, Udio): restores punch, no distortion.
• Russian platforms — VK Music, Yandex Music, Zvuk and Reels/Shorts in one pick (−14 LUFS).
• Voice (−16 LUFS) and Audiobook (−19 LUFS) — presets for podcasts and voiceover, no reverb.
• Character presets — Warm, Bright, Bass.
• 3× faster — parallel processing of up to 4 tracks at once.
• The Telegram bot now has a /cancel command to stop an in-progress master.23 presets in total. Full write-up of every feature: magicmaster.pro/blog/magic-master-3-2-release-en
Try it: magicmaster.pro/app — your first file each day is free.New presets and platforms: AI Unlimiter, VK, Yandex Music
We've added new presets and platform targets — pick them on the mastering page.
AI Unlimiter opens up over-compressed 'brickwalled' and AI tracks (Suno, Udio): it brings back punch and crest factor by re-emphasizing transients, with no added distortion. What a limiter clipped can't be fully recovered, but the sense of impact and openness can.
Russian platforms in the target picker: VK Music, Yandex Music, Zvuk, plus Reels / Shorts / Clips (−14 LUFS, True Peak ≤ −1 dBTP). One master fits them all.
Voice and Audiobook are speech presets: podcasts (−16 LUFS) and voiceover/audiobooks (−19 LUFS) with rumble cleanup and an even, intelligible voice with no reverb.
Character presets: Warm (analog/bass), Bright (open/air), Bass (sub for EDM and rap).
All new presets are part of PRO mastering. Your first basic master each day is free.
Magic Master 3.1.0: share cards, honest loudness, Cardlink
What's new in 3.1:
Share cards — a premium PNG with a QR code and Telegram OG preview for social, plus a short link to your result.
Loudness convergence — honest LUFS iteration progress: if the target loudness is physically unreachable, we explain why.
Result UI — a quality radar, an A/B spectrogram curtain and a Share button next to export.
International payments (Cardlink, USD and EUR) — integration ready, launching soon; not enabled yet.
Transparent free limits — guest 1 master per month, account 1 per day (UTC), the same copy everywhere.
Full 3.1 overview: magicmaster.pro/blog/magic-master-3-1-release-en