Delete the ten.
Keep the decade.
Somewhere in your last 3,200 posts are a handful that could cost you a job, a round, or a friendship. xxxx.fish reads all of them the way someone hunting for ammunition would, then hands you the shortlist.
3,184 posts · Jan 2014 – Jun 2026
An example scan, not your data. Yours takes a few minutes.
- 3Nuclear
- 9Hot
- 34Medium
- 291Mild
- 2,847Clear
The same six posts
A filter finds words. We read the post.
Every other tool in this category matches banned words, or deletes by date. Run one over a real timeline and it flags the burrito joke, then waves through the thing that would actually end up in a screenshot. Switch the control and watch which words light up.
6 posts · one timeline · nothing cherry-picked
Nov 2019
this deadline is killing me. three more hours and I'm free
Flagged
Matched the word “killing”. No idea what the sentence means.
Feb 2022
board deck done. we're calling the numbers directional and nobody is going to ask
No match
Contains no banned words, so it is never shown to you.
Jun 2015
the new album is sick. no skips, start to finish
Flagged
Matched the word “sick”. No idea what the sentence means.
Apr 2018
if anyone has his employer's HR line, DM me. let's see how funny he thinks it is on monday
No match
Contains no banned words, so it is never shown to you.
Aug 2021
I would die for one of those breakfast burritos right now
Flagged
Matched the word “die”. No idea what the sentence means.
Sep 2017
she didn't get promoted because she's good. she got promoted because she's likeable in meetings
No match
Contains no banned words, so it is never shown to you.
Three flagged, three cleared. All three flags are harmless. All three posts that could cost you something were cleared. This is the tool most people are using.
Three flagged, three cleared. The other way around — 76, 88 and 91 — each with the sentence that earned it and a reason you can argue with. You decide what goes.
You pick the model that decides what counts as bad.
Four models read the same taxonomy and write the same kind of reason. They do not agree with each other, and that is the point—“career-ending” is a judgement call, so you should be able to choose whose judgement you are borrowing.
Claude
Best at reading tone, irony and intent.
anthropic/claude-haiku-4.5
Gemini
Fastest across very large timelines.
google/gemini-3.6-flash
ChatGPT
Strict, literal, flags more than it misses.
openai/gpt-5.6-luna
Grok
Native to X. Strictest on slurs and doxxing.
xai/grok-4.1-fast-reasoning
The same ten categories
Whichever model you pick scores every post against this list and names the categories that apply. Anything landing between 35 and 80 gets a second opinion from that provider’s stronger model, because cheap models are confident about the easy cases and wrong about exactly these.
- Slurs & hate
- Sexual
- Violence & threats
- Harassment
- Drugs & illegal
- Political extremism
- Conspiracy
- Doxxing
- Professional risk
- Just cringe
Why Grok is here
Grok is the only one of the four that can see X itself, so it is what fetches the thread a post sat in and how far the post actually travelled. That context goes to whichever model you chose, and your model scores the risk.
It is not the lenient option
On our calibration set Grok scored the slur and doxxing cases higher than Claude did. Being less censorious is about expressing values; this job is predicting what other people will seize on, and those are different things.
Your choice changes which model scores the posts. It does not change what we read, what we store, or how deletion works—and it never widens the scan beyond your own account.
Search in plain English. Review in the workspace.
Ask for a theme, a moment, or a kind of risk. The assistant finds the posts and opens a real result set—no separate AI inbox to decipher.
- Searches by meaning, not just exact words.
- Results become the same sortable workspace you already use.
- It can propose a selection; only you can approve a deletion.
Optional · 1 credit per turn · closed by default
Ask xxxx.fish
Your posts, in your workspace
You ask
Show professional risk over 70 from before 2025
Professional risk · before 2025
AI search · sorted by heat
- 78HotJul 2024
Leadership knew the numbers were soft and shipped anyway.
- 91NuclearMay 2023
Anyone who supports that policy deserves to be run out of town.
- 87NuclearOct 2022
If you know where the editor lives, send it over.
Close the tab. The work keeps going.
Scans and deletions are durable background jobs. You get a live status, a realistic estimate once the work is known, and a clear point at which a deletion can still be cancelled.
Scan
Measured after we start
We report the work actually found in your timeline instead of inventing an ETA up front.
Before deletion
1 hr to change your mind
Download the backup, confirm the selection, then keep the safety window—or explicitly start now.
While deleting
50 posts every 15 minutes
That is X's pace, not ours. The job saves progress and resumes without an open browser tab.
Example: 312 posts
Approximately 90 minutes (1 hr 30 min) of X-paced processing after the job becomes eligible. X pacing and rate limits can change the estimate.
Exactly what you’re handing over.
Sign-in is X’s own OAuth screen, so you approve this list there rather than taking our word for it. Here is every permission we ask for, and what each one actually does.
tweet.readRead the posts on your timeline so they can be scored — plus the thread around one, if you ask for Deep Context on it.
users.readRead your handle, display name and avatar, so the workspace knows whose history it is holding.
tweet.writeDelete a post or undo a retweet — and only the ones you have selected and confirmed.
offline.accessKeep a job running after you close the tab. A deletion sweep outlives any browser session.
Not your DMs
Direct messages need a scope we do not ask for, so we cannot read them.
Not your likes or follows
Those are separate permissions. We never request them, and X has stopped letting apps remove likes anyway.
Reversible on your side
Disconnect whenever you want. Erasing the account clears the cached posts, scores and searches we held.
Start free. Go deeper once. More is coming.
Start with 117 credits: enough for a 100-post scan with 17 left to try chat or deletion. Paid packs add credits for deeper reviews and follow-up actions. Watchdog monitoring is coming soon.
Your starting allowance
See what’s there first.
Free allowance
FreeMild
117 credits included
See what's there.
- Scan your 100 most recent posts
- Heat rating on every post
- Delete anything you find
One-time packs
Add credits for a deeper review.
Buy once · credits never expire
One-time purchase
Medium
400 credits included
Same price as the checkmark that got you into this.
- AI judgement across 400 posts
- Categories and reasoning on every flag
- Freeform semantic search
One-time purchase
Hot
1,500 credits included
A deep timeline sweep, properly read.
- Scan and judge up to 1,500 posts
- One narrative report across that scan
- Saved searches with one-click reruns
One-time purchase
Deepest timelineGhost
5,000 credits included
Opposition research on yourself.
- Scan up to 3,200 posts from X's timeline
- Deep forensics: thread context and live reach
- Timeline-wide pattern report
Monthly monitoring
Keep watch without buying a pack every time.
Monthly subscription
Coming soonWatchdog
300 credits each month
Automatic monitoring is in final testing.
- Every new post rated within minutes
- Daily sweep for old posts resurfacing
- 300 credits a month
Watchdog will be the only recurring plan. Existing subscribers can still manage billing from their account.
Metered actions
The meter is visible before the work starts.
Timeline AI judgement is included with the scan. Credits cover the X reads and the actions below, so there are no surprise overages.
Reading and removing posts
Charged per post, at the moment the work happens.
- Timeline AI judgement
- Included
- Read a post
- 1 per post
- Delete a post
- 1 per post
- Undo a retweet
- 2 per retweet
Optional extras
Nothing here runs until you ask for it by name.
- Assistant turn
- 1 per turn
- Deep context
- 5 per post
- Draft an explanation
- 1 per draft
- Timeline Report
- 50 per report
Credits never expire. No automatic overages.
The limits are part of the plan.
Deleted posts cannot be recovered. Every job starts with a backup export and waits one hour before the first deletion.
X no longer lets apps remove your likes. We will never imply otherwise.
The X API reaches your most recent 3,200 posts. Anything older needs an archive upload, which is not built yet.
Scanning is not deleting. Read your first hundred posts free, then decide what—if anything—goes.
By continuing you agree to the Terms and Privacy Policy.