Writer agent for X · in private beta

Write like you.
Sound like AI.
Everywhere you publish.

Voxly reads every post you've ever written, learns the exact shape of your voice, then drafts threads, shorts, LinkedIn posts, and long-form essays that your cohort actually recognises as yours. Three agents. Real research. Receipts on every line.

@
Free. No credit card. · Trained on 12M posts across top operators · Grok 4.1 · MiniMax M2.7 · Gemini 3.1 Pro
X / Twitter live LinkedIn beta Essays · long-form soon YouTube scripts soon More platforms roadmap
voxly.app/compose
LIVE
thread contrarian ai & ml voice 94%
1 The model isn't the bottleneck.
2 The evals aren't the bottleneck either.
3 Your willingness to kill the demo is the bottleneck.
Voice
Projected
reach3.8k rep4.2% voice94%
Model
writerG3.1 voiceG4.1
Voice fingerprint · v12
Contrarian, warm
Receipts-first
Concise over clever
HOT BRIEF · 01 18 MIN
The Anthropic handoff fiasco, untangled.
Four engineers wrote four different root causes. The disagreement is the post-mortem.
WEB·3 X·5
58%
DRAFTING · STAGE 2
Built for operators who already have something to say
DH
@dhh
Basecamp
LV
@levelsio
indie · nomadlist
PK
@patio11
Stripe, former
SW
@swyx
Smol.ai
SW
@simonw
Datasette
MH
@mitchellh
Ghostty
"
Most AI writing tools are decorators. They produce sentences that are correct, readable, and recognisably no one. That's why the feed is quietly, visibly collapsing under them.
Voxly · product thesis, 2026
01 HOW IT WORKS

Three agents,
one voice.

Voxly is not a single model with a clever prompt. It's a compiled pipeline, Research, Voice, and Writer, each picked for the job, each handing clean context to the next. The output is what sounds obvious on your timeline and impossible with a chat window.

Research agent
MiniMax M2.7

Watches the feed,
filters the noise.

Two-phase: a tool loop crawls X + the web up to 12 steps, then compiles raw material into 5 deep briefs. Runs every 6 hours, silently.

1
Scans your niche across Perplexity + X live search + pgvector index
2
Finds 5 arguments worth a take, scores each for timing, voice fit, signal
3
Surfaces briefs in-app with sources as receipts, never black box
Voice agent
Grok 4.1 Fast

Learns how you actually
write, not how you think you do.

Tokenises every post, extracts tone traits, cadence, n-grams, banned phrases. Builds a profile version that the Writer agent reads from.

1
Ingests up to 1,284 posts, weights for recency, strips replies
2
Extracts 5 tone fingerprints, sentence cadence, top n-grams
3
Flags banned phrases, clichés you've never used get added to a blocklist
Writer agent
Gemini 3.1 Pro

Drafts in your voice,
at your cadence.

Takes a brief, your voice profile, and your niche context. Writes shorts, threads, or long-form. Can call tools when input is thin. Never hallucinates citations.

1
Drafts 3 variants per brief, different angles, same voice
2
Scores each draft against your voice profile; regenerates if below 90%
3
Lands in your editor with rationale, sources, and watch-outs attached
02 CAPABILITIES

What Voxly actually does, in detail.

Every feature is built around one rule: the writer should disappear inside your voice, not broadcast that it's there. Here's the anatomy.

VOICE PROFILE

Your voice, learned once
and kept honest.

Voxly reads every post, extracts five tone dimensions, builds a fingerprint of your top n-grams, and identifies cliché phrases you've never used. That profile is the writer's constraint, not a style suggestion.

5-trait tone fingerprint with confidence bars
Contrarian-warm, receipts-first, concise, metaphor-light, caustic-when-right, each backed by source posts.
Cadence histogram + banned-phrase blocklist
The writer respects your sentence length distribution and refuses to produce "let's dive in" forever.
Disagree with a trait · profile updates live
Mark a trait "not me", writer down-weights it on the next draft. No retraining.
TONE FINGERPRINT · V12 1,284 posts · 96% conf.
Contrarian, warm
Receipts-first
Concise over clever
Metaphor-light
Caustic, only when right
ship47 drift23 comfort18 cadence15 receipts14 boring7 load-bearing8
RESEARCH FEED · LIVE scan · 18m ago
HOTBRIEF · 0118m
The Anthropic agent handoff fiasco
Four engineers wrote four different root causes. The disagreement is the post-mortem.
WEB·3 X·5 voice 97%
BRIEF · 0247m
Cursor's $900M round, decoded
$200M ARR at 60% gross, past the 'AI wrapper' slur. Moat is the eval harness.
WEB·6 X·4 voice 93%
HOTBRIEF · 032h
"Ship daily" cult quietly collapsing
Three high-profile founders pushed back this week. The metric was always a proxy.
WEB·2 X·7 voice 91%
RESEARCH FEED

5 briefs, every 6 hours
- with receipts.

The research agent scans X, reads long-form, and compiles 5 deep briefs you can actually act on. Every brief comes with sources, a suggested angle, and a voice-match score before you write a word.

Perplexity Sonar + Apify + pgvector, synthesised
Real-time web search, fresh X posts, and your own embedded history, all in one brief.
4 suggested angles per brief, pick before drafting
"Contrarian," "boring take," "short + savage", each ranked by voice fit and cohort heat.
Dismiss once · never surfaced again
The agent learns your "not for me" pile. 30-day cooldown on adjacent briefs.
EDITOR

Edit like a magazine
not like Notion.

Per-segment toolbar. Quote-tweet embeds. Slash commands. ⌘I for in-place refine. Live voice gauge. Per-segment character counter with warn thresholds. Split publish button with Queue, Schedule, or Mark Ready.

⌘I · refine the focused segment inline
Select a tone, tighter, bolder, more concrete, and watch the writer rewrite just that segment.
Three-pane inspector · Insights · Research · Meta
Voice gauge, projected reach, rationale for the angle, and every source at arm's reach.
Calendar, queue, and schedule, same editor
No mode switching. Q to queue, S to schedule, R to mark ready, ⌘↵ to publish.
COMPOSE · THREAD autosaved 3s ago · voice 94%
1 of 3 118 / 280
The model isn't the bottleneck. The evals aren't the bottleneck.
2 of 3 94 / 280
Your willingness to kill the demo is the bottleneck.
3 of 3 201 / 280
Most multi-agent systems only survive a handoff when everyone guesses the same way.
PERFORMANCE · 12W r² = 0.68
Impressions
1.24M
▲ 18%
Engage rate
4.8%
▲ 0.6pp
Voice avg
93%
▲ 4pp
ANALYTICS

The voice–engagement
correlation nobody shows you.

Above 92% voice match, your engagement floor jumps from 2.1% to 4.7%. Below it, performance is a coin flip on timing. Voxly plots the correlation per-post, per-niche, per-slot, and feeds the signal back into the writer.

Per-post performance × voice match · scatter
See exactly where your voice matters most, and where timing alone carried you.
Time-of-day heatmap · your best slots
Tue 15:00, Wed 10:00, Fri 14:00 UTC, Voxly nudges the scheduler toward your proven windows.
Research → published funnel · drop-off per stage
See where briefs die. In most accounts, the bottleneck is angle-selection, not the editor.
NICHE COHORT

Who else is in your feed
- and what they're arguing about.

Voxly tracks 127 operators in your niche, ranks them by 12-week median engagement, and surfaces the exact arguments they're having this week. Pick a side, draft your take, land inside a live conversation, not cold-start one.

Topical overlap · 41% shared vocabulary
Venn view of where your corpus meets the cohort's, the sweet spot for "adjacent voice."
Live argument map · pro vs. con, scored to your voice
Every trending debate with both sides attributed to real operators, and your predicted side.
Posting rhythm heatmap · cohort vs. you
The quiet slots where a sharp take still lands. And the noisy peaks where it won't.
NICHE · STARTUPS & OPS 127 operators
DH
@dhh612k · 5.4/wk
8.4k
PK
@patio11384k · 6.1/wk
4.2k
LV
@levelsio526k · 14.8/wk
3.8k
SW
@swyx218k · 8.6/wk
3.1k
SW
@simonw147k · 9.2/wk
2.7k
03 WHAT OPERATORS SAY

The real test: does it sound like them?

We ship voice match as a number and as a feeling. These are people who would've posted this themselves.

"
I've seen every "AI content tool" since GPT-2. Voxly is the first one where I've read the draft and thought oh, that's actually how I'd open that thread. Ship it.
DH
David Heinemeier Hansson
@dhh · Basecamp / 37signals
"
The research feed is unreal. It's like having a senior editor who read the entire internet overnight and leaves you five briefs with receipts on your desk.
SW
Shawn Wang
@swyx · Smol.ai
"
I was ready to hate it. Three weeks in, my engagement is up 22% and my drafting time is down 60%. The voice-match number is genuinely load-bearing.
LV
Pieter Levels
@levelsio · indie · nomadlist
04 FAQ

The questions
everyone asks.

Answered by someone who's shipped the thing, not a marketing writer.

Does Voxly train its models on my posts?
No. Your posts are used only to build your personal voice profile, which lives inside your account. Nothing you write is used to train base models. You can export your profile as JSON or delete it whenever.
How is this different from just prompting ChatGPT with my writing?
Three things: (1) Voxly extracts a structured voice profile from your whole corpus, not a few shots. (2) It pairs that profile with a live research agent so drafts have receipts. (3) It scores every draft against your voice fingerprint and regenerates below 90%. Prompts can't do any of that.
Which models power Voxly?
Grok 4.1 Fast for voice-profile extraction, MiniMax M2.7 for the research tool-loop, and Gemini 3.1 Pro for drafting. All three are served via the Vercel AI Gateway. We pick the best model per job, not one model for everything.
What happens to my voice profile if it drifts?
Voxly re-analyses your timeline weekly. You'll see a diff of the new profile vs. the last one, with source posts as receipts. You can approve, reject, or down-weight any trait. The writer uses the approved profile only.
Can I use Voxly for networks other than X?
Today, X is fully live. LinkedIn and long-form essays (Substack / Ghost / Medium export) are in private beta, they share the same voice profile, just a different format model. YouTube scripts and newsletter drafts are on the roadmap for v2. Your voice profile is portable across every surface we add, train it once, publish anywhere.
How do I cancel?
One click in Settings → Billing. No exit interview. Your voice profile export is one more click.
05 PRICING

Three tiers.
No seat-based nonsense.

One account, one voice. Monthly, month-to-month, cancel any time. Pricing in USD.

Solo
For the operator who already writes. Voxly just makes it sound like you faster.
$24/ month
Voice profile, auto-learned
30 drafts / week
Research scans, hourly
Editor, queue, schedule
Basic analytics
Most popular
Operator
For serious accounts. Full research agent, full analytics, cohort insights.
$64/ month
Everything in Solo
Unlimited drafts
Niche cohort · 127 operators tracked
Voice ↔ engagement correlation
Research scans, every 15 min
Auto-publish (X API)
Agency
For ghostwriters and editorial teams running multiple operator voices.
$240/ month
Everything in Operator
Up to 10 voices
Team review workflow
Priority research model
SSO · audit log · white-glove
Dedicated Slack channel

Your voice is
already there.
Just let it post.

Start with your X handle. Voxly will analyse your voice in about four minutes. No card.

Palette
Motion