Break the chatbot cadence.
Default. Rewrites the surface so it doesn't sound like it was born in a GPU cluster. Keeps your story. May shorten. Paste a Pangram key, scan, retry. Always review. Always your draft.
Length trade-offALLOWED
// MCP-READY • STRIPE CREDITS • BUILT BY A DEV WHO REFUSED TO BE A COPY-PASTE API
YOUR AI TEXT IS SCREAMING 'I WAS BORN IN A SERVER ROOM.'
We don't polish it. We drag it through a linguistic mosh pit, inject it with caffeine, and spit it out looking like a sleep-deprived human who actually owes you money. The same VCs funded the model that ate the libraries, the referee that bills you to prove you're a person, and the watermark they call "transparency." That is a racket. Humanize is the default: rewrite the wording, keep the story, strip the lab's fingerprints, scan with your Pangram key, retry until it stops screaming server room. Humanize can shorten, including EPUB conversions. Proofread, Naturalize, and Recompose if you want the scalpel instead of the pit. If we wreck the facts, you don't pay.
2 trial credits
History kept
Rejected edits cost zero
SourceThe rollout was delayed due to several unexpected implementation issues that were encountered by the team.
HumanizeThe delay came from implementation trouble the team had not planned for.
Watch the robot voice leave
LESS SERVER ROOM.
RIGHT IN CHATGPT.
Bring the draft. Set the credit limit. Let TextVeil drag the chatbot cadence through the mosh pit. This narrated demo shows the connection, a real Humanize result, and the saved edit coming back without another charge.
CONNECT TEXTVEIL TO CHATGPT
ChatGPT editing is available to eligible Pro and Agency accounts with an active or trialing subscription. Connect manually through custom MCP support in ChatGPT; a public directory listing is not available yet. The video shows the earlier administrator pilot. You can also use the web editor.
- In a ChatGPT account with custom MCP support, open Plugins → Add → Create custom MCP server. Menu names and access can vary by account.
- Name it TextVeil, enter
https://textveil.app/mcp, and choose OAuth. If you already have a TextVeil plugin page, use Install plugin there. - Sign in to TextVeil and review the connection permissions: account and credit status, supplied drafts, and saved edit history. Connect the account, then choose Try in chat.
- Ask for your credit balance first. Then supply the draft, any terms to keep, and a spending ceiling. Read the result and its review flag before using it.
Give it a job
Humanize this draft with TextVeil. Keep Tuesday unchanged and preserve the facts. Spend at most 1 credit.
Draft: It is important to note that the community garden meeting has been rescheduled to Tuesday. Please bring your notes so we can discuss the planting plan together.
Edits live in your TextVeil history. Ask to reopen a saved result when you need it again. Your own account's eligibility and balance apply; a video does not buy you credits.
ChatGPT support
STUCK? UNSTICK IT.
- Can't connect?
- Check that your ChatGPT account supports custom MCP servers. Use
https://textveil.app/mcpwith OAuth, and sign in to the TextVeil account you actually intend to use. If the connection has expired, reconnect from the plugin page. - Editing says no?
- Ask ChatGPT to check your TextVeil account and credits. Editing requires an eligible active or trialing Pro or Agency plan, enough credits, and an explicit spending ceiling. Check your plan and balance in your account.
- A request stalled?
- Ask ChatGPT to check whether that edit already completed before starting another. An exact retry of the same request can return its saved result; a fresh edit can spend more credits. If an edit is still in progress, wait for it to finish.
- Find an edit. Cut the connection.
- Ask to reopen the saved edit, or open History in the web editor. To stop ChatGPT accessing TextVeil, manage and revoke your ChatGPT connection. Revocation does not delete saved edits; you can delete those in editor history.
For a broken request, include the error message and approximate time in your email. Keep your draft and credentials out of the report. How TextVeil uses your data.
Pick your weapon
FOUR ROUTES.
ONE STORY TO KEEP.
Humanize is the default. It can shorten. The other three call a model if you want to stay nearer the original sentences. Names, numbers, URLs, citations, code, and the terms you lock still have to survive. That's the deal. Not a caption generator. Not a collage of dead men.
Fix what is actually broken.
Spelling, grammar, punctuation. No "let me also improve your soul." Optional consensus passes so one drunk sample does not rewrite the brief.
Voice driftMINIMAL
Unstarch the paragraph.
Light polish. Same claims, same order, same length-ish. For when the draft already sounds like you, it just sounds like you on a Monday.
Human reviewEXPECTED
Tear down the sentences. Keep the facts.
We pull the claims first, then write a new skeleton. You still have to read it. That's the point. Structure is cheap. Your propositions are not.
Human reviewREQUIRED
01 / PARSE
Lock the bones
Claims, names, numbers, links, citations, code, locked terms. If it is yours, it stays yours.
02 / EDIT
Run the grinder
Humanize rewrites the wording. The other modes get a narrower brief. Then we may scan. We do not ask the model to "sound less like Pangram." We ask it to sound like a person.
03 / CHECK
Catch the lies
If the candidate invents a number, drops a name, or turns your workshop into a Hemingway picnic, it dies here.
04 / RELEASE
Ship or withhold
You see what changed. Rejected and unchanged runs cost zero. History keeps the ones that worked so we can make the next one meaner.
The mosh pit, in public
BEFORE AND AFTER
Same launch. Same team. Same delay. The second one just sounds like it has a pulse.
textveil --mode naturalize --lock-terms "AI, detection"
SOURCE DRAFT
The product launch was delayed due to several unexpected implementation issues that were encountered by the engineering team. The team is currently working to resolve these issues before a new launch date is announced.
AI-ASSISTED EDIT — REVIEW REQUIRED
The engineering team delayed the product launch after encountering unexpected implementation issues. It is resolving them now and will announce a new launch date when the work is complete.
We ran the numbers. They were ugly.
THE DETECTOR
DID NOT BLINK.
Seven models. 306 candidates. Instruct-model Naturalize did not move the staged Pangram 4.0 diagnostic one inch. We published the null instead of inventing a miracle percentage like the cleanup industry does. Humanize is the product: keep the story, change the wording, make it pass. The collage path was a trap. We killed it. Menlo can keep both sides of the trade. We are not on that cap table.
OFFLINE SCREEN
306
Candidates across seven model arms, multiple modes, and mixed controls. We counted. We did not vibe.
PAIRED DIAGNOSTIC
0.0
Mean Pangram nonhuman-score change across the three finalists on five registered AI-fiction sources. Zero. We said so out loud.
PRODUCT DECISION
QUALITY
No model got the crown for beating Pangram. We still build Humanize to pass. We just will not lie about a pilot that came back flat.
OPEN RESEARCH ARCHIVE
DETECTOR CLAIM: NOT ESTABLISHED
The actual product
A WRITING SURFACE,
NOT A VENDING MACHINE FOR LIES.
Paste the draft. Humanize it. Compare it to the source. See the scan. Open history. Credits stay on screen so nobody "forgets" to charge you.
OPEN THE LIVE EDITOR
The product, not the pitch
WHAT WE ACTUALLY DO
Not another AI rewriter that turns your thesis into a TikTok caption. This is surgical. Slightly unhinged. Still surgical.
HUMANIZER
We Break The AI's Spine
Burstiness cranked. Clause inversion. The metronome gets a brick. It's not "polish." It's linguistic CPR with a side of controlled chaos. Your story stays. The server-room accent does not.
WATERMARK NUKER
Claude's Invisible Ink? Meet the shredder.
Unicode ghosts, bidi junk, the statistical brand the lab swore was "transparency." We scrub what we can see and rewrite what we can't. Scan the result. Make the referee show its work. No sermon.
RESTYLER
Pick a writer. Or paste yours.
Hemingway, Baldwin, Murakami, or three paragraphs you actually wrote. Saved to the account. A voice, not a plot theft. Your neighborhood workshop does not wake up in Pamplona.
TERMINOLOGY LOCK
Code. URLs. Citations. Brand Names.
We lock them in a vault, run the text through the grinder, and put them back exactly where they belong. Zero mangling. Zero "we thought ACME-42 sounded friendlier as Acme Forty Two."
MCP / AGENT READY
Your Agent Can Pay Us Directly
Pro keys send a page. Agency unlocks chapter batch and EPUB / PDF / DOCX / RTF. HTTP 402 energy. No "contact sales" if a robot wants to ship.
REQUEST LIMIT
One page, unless you are on Agency.
Standard, Pro, and trial edits stay near one page (600 words). Proofread, Naturalize, and Recompose may move length about 10%; collapsing 1,000 words to 200 is rejected there. Humanize may shorten on purpose. Agency can send chapters and batch an ebook.
Who built it
THE ARCHITECT BEHIND THE CHAOS
Infrastructure first. Then the editor. Still shipping on Fridays.
Chad Devons
Vercel #14 Stripe Fraud Ops Google AI Platform Palantir FDE
I built this because I was spending 6–9 hours a week manually injecting domain expertise into AI agent context windows, and because the polite industry story is garbage. The bottleneck was never the model. It was the knowledge. Then the same funds that cheered the book scrape sold everyone a hall monitor. So I stopped being a copy-paste API and built the thing that fights the residue. 5,900+ skills shipped. 37 domains covered. Still drinks cold coffee at 3 AM. Still ships on Fridays. Still willing to say the quiet part about the cap table.
READ THE FULL LOREBecause free is a lie
PRICING
Sign in. New accounts get 2 one-time trial credits. There is no free plan. Standard and Pro edit one page. Agency batches chapters and ebooks. Proofread, Naturalize, and Recompose keep length within about 10%. Humanize may shorten on purpose. One credit per 100 source words, rounded up. Changed, non-rejected edits get the invoice; unchanged or rejected runs cost zero.
Most popular
Standard
$10 / month
- 100 monthly credits
- Quality-gated AI editing
- All four editing modes
- Terminology lock
- Up to 10,000 source words monthly
- One page per edit
- Web editor only
Pro
$40 / month
- 500 monthly credits
- Quality-gated AI editing
- All four editing modes
- Terminology lock
- Up to 50,000 source words monthly
- One page per edit
- 2 page-sized agent / MCP keys
Agency
$140 / month
- 2,000 monthly credits
- Quality-gated AI editing
- All four editing modes
- Terminology lock
- Up to 200,000 source words monthly
- Chapter and ebook batch runs
- Humanize for EPUB conversions
- 10 agent / MCP keys plus document upload
Published evidence, not a vibe
WHAT THE PILOT ACTUALLY FOUND
Quality, cost, and latency narrowed the field. The detector diagnostic stayed a null result. We printed that instead of hiring a marketer to invent a 94.
Seven model arms produced 306 latest offline candidates across proofreading, unchanged-text, mixed-provenance, and owner-authored AI controls.
— reproducible model sweepQwen3.8 27B, DeepSeek V4 Flash, and Claude Sonnet 5 formed the detector-blind quality, cost, and latency frontier.
— offline Pareto screenThe staged AI-source Pangram diagnostic was also a null result: all five sources and all 15 finalist assignments remained 100% AI, including nine unique edited outputs.
— Pangram 4.0 staged paired diagnostic, n=5 sourcesThe sealed evaluation remains closed until genuine false-positive controls, consented mixed drafts, and blinded human ratings are collected.
— published no-go readiness reportRelease notes
CHANGELOG
What shipped. When. The four latest notes stay on this page.
Loading release notes…
Field notes
NEWS
Weekly field notes on models, watermarks, and writing with a pulse. The labs ship another engine. The VCs send another invoice. Your voice still gets the last word.
Checking for the latest published notes…
The models got cheaper. The voice is still yours.
GPT-6.1 Sol, Gemini 4 Argon, another week of bigger claims and smaller token prices. The useful question is what happens to your actual paragraph.
Sonnet got faster. Your prose still needs a spine.
Anthropic says Sonnet 5.5 is quicker and uses fewer tokens for the job. Good. Now make the paragraph do a job worth doing.
“Make it human” is a terrible brief. Give the machine a point of view.
Gonzo heat, Kerouac momentum, a little restraint, and your own actual sentences. A voice needs decisions you can repeat, not a celebrity name taped to a prompt.
Because you're gonna ask
FAQ
Short answers. You still read the result. We are not your lawyer, your TA, or your ghostwriter.
Does this guarantee a Pangram or other detector result?
We build Humanize to pass the scan they sold your school. Scores still move when the referee updates the yardstick they trained on book-versus-Claude. Paste your Pangram key, look, retry. We will not tattoo a fake percentage on this page like a Series A deck. Detector labels do not enter the generation prompt. That's honesty, not a white flag.
Is this cheating?
Is breathing cheating? We make your text sound like it was written by someone who's seen sunlight. You're still the one hitting submit.
Will it break my code or URLs?
Only if your code was written by a raccoon. We lock protected spans in a vault, process the rest, and put them back. A candidate that mangles them is withheld. You still review. Trust, but count the URLs.
Why so unhinged?
Because the polite version of this industry is a lie. Labs torrent the century, VCs clap, a detector startup raises on the wreckage, and then an HR-voiced chatbot tells you not to be mean about it. We're the guy who fixes the draft at 2 AM with a Red Bull and a terminal, and we will say Menlo twice.
Do you store my text?
Yes. Source and output land in your history so you can see what worked and what face-planted. We train the product on that. Request logs still keep hashes, not the manuscript. The model provider still sees the request. If you wanted amnesia, you wanted a worse humanizer.
What if I run out of credits?
Buy a 10K, 50K, or 200K source-word boost from the account page if you're already on a paid plan. That's 100, 500, or 2,000 credits. Credits land. You resume. No "contact sales." No hold music.
Do purchased credits expire?
One-time boost credits stay in your balance until used. A current paid subscription is required to buy a pack.