FOR DEVELOPERS
Claude's text watermark hardly touches working code. What people notice in Claude Code is the visible attribution it writes into git, and that part is a setting you control.
$ claude "fix the retry logic in the upload queue"
Reading src/queue/upload.ts
Editing src/queue/upload.ts (+14 -6)
Running npm test
✓ 42 tests passed
Committing: Fix retry logic in upload queue
$ git log -1
MacBook Pro
ClaudeCleanLocal scanPASTED FROM CLAUDE CODE
SCANNING
Set attribution to false in settings.json and new commits stay clean.
ClaudeCleanLocal scanPASTED FROM CLAUDE CODE
SCANNING
Set attribution to false in settings.json and new commits stay clean.
Short answer
CODE VS PROSE
Anthropic answered this directly in its watermark announcement. The short version: exact code leaves no room for the mark, prose around it does.
But by definition, it will have a negligible effect on the actual code produced.
LITTLE ROOM TO MARK
Change a keyword or an operator and the program breaks, so the watermark has almost no equally good alternatives to choose between.
Illustration
def retry_delay(attempt: int) -> float:
return min(30.0, 0.5 * 2 ** attempt)ROOM TO MARK
The same idea could be phrased dozens of ways. Anthropic names comments inside code as a place the watermark can be used.
Illustration
# Back off exponentially so a flaky API
# gets breathing room, but never wait
# longer than thirty seconds per retry.WHERE THE MARK CAN ACT
Every token is a choice. Code leaves almost none.
A watermark nudges which of several equally good words gets picked. Watch the highlighted token on each side and the options it really has.
Illustration: in the code line every token has one correct value, so a watermark has almost nothing to choose between. In the comment each phrase has several equally good wordings, which is where the mark can act.
OpenAI makes the same point about its own textGrain mark: its help center says code is harder to watermark, and the EU transparency code does not require marks on code snippets or on outputs under 200 tokens.
EVERY SIGNAL, ONE TABLE
Seven things people call the Claude Code watermark, where each shows up, and the switch that controls it.
| Signal | What it is | Where it shows | How visible | How to control it |
|---|---|---|---|---|
| Statistical watermark in code | Pattern in token choices, keyed by Anthropic | Generated source code | Negligible: code has few equally good choices | None needed in practice |
| Statistical watermark in comments | Same pattern, more room to act | Comments, docstrings, READMEs, PR prose | Can register on long prose | Write or rewrite prose yourself |
| Co-Authored-By trailer | Co-Authored-By: <model> <noreply@anthropic.com> | Every commit Claude Code makes, by default | Visible to anyone reading git log | attribution.commit or attribution: false |
| PR description line | Generated with Claude Code, linking claude.com | Pull requests Claude Code opens | Visible on the PR page | attribution.pr set to an empty string |
| Claude-Session trailer | Link to the claude.ai session | Commits and PRs from cloud or Remote Control sessions | Visible, and the link points at your session | attribution.sessionUrl: false |
| C2PA credential | Signed metadata note | .png, .jpg and .svg files Claude produces | Readable by any C2PA-aware tool | Lost when metadata is stripped |
| Hidden characters | Zero-width spaces and similar | Not added by the watermark, per Anthropic | Only present if pasted in from elsewhere | grep for them, then delete |
Swipe to compare →
Sources: Claude Code settings reference and Anthropic's watermark announcement, checked October 9, 2026.
WHAT THE SETTING CONTROLS
What Claude Code adds to a repo, and what one setting removes
Flip attribution and watch which layers change. Only the git and PR text is a switch. The watermark and file credentials are not.
Orange rows are written by Claude Code and follow the attribution key. The rest stay as they are whatever the setting says. Sources: Claude Code settings reference and Anthropic's watermark announcement.
The trailer names the model
BEYOND THE COMMIT
Docs, comments and PR descriptions are prose, and prose is where invisible characters and AI phrasing pile up. Scan them in your browser before you publish.
SETTINGS.JSON
All of this lives in the attribution key. Put it in .claude/settings.json for one repo or ~/.claude/settings.json for every project.
WATCH THE SETTING LAND
One key in settings.json, then a clean commit
Pick the form that matches your Claude Code version. The editor types the exact config from the steps below and the terminal shows the next commit message.
EXPLORER
UPLOAD-SERVICE
.claude
{}settings.json
src
tests
{}package.json
MREADME.md
upload-service › .claude › settings.json
1
2
3
4
5
6
7
TERMINALPROBLEMSOUTPUT
$ claude "commit the retry fix"
Committed 8b1d0e4 on main
$ git log -1 --format=%B
Fix retry logic in upload queue
MacBook Pro
upload-service › .claude › settings.json
1
2
3
TERMINALPROBLEMSOUTPUT
$ claude "commit the retry fix"
Committed 8b1d0e4 on main
$ git log -1 --format=%B
Fix retry logic in upload queue
No Co-Authored-By, no PR line, no session link
On Claude Code v2.1.281 or later, this removes the commit trailer, the PR line and the session link.
.claude/settings.json
{
"attribution": false
}Versions before v2.1.281 reject the false value and skip the whole settings file. This shape works on both.
Works on any version
{
"attribution": {
"commit": "",
"pr": "",
"sessionUrl": false
}
}Some projects want disclosure in a set format. This writes the Linux kernel Assisted-by tag instead and clears the PR line.
Custom trailer
{
"attribution": {
"commit": "Assisted-by: LLM",
"pr": ""
}
}Per the settings reference, attribution rules in CLAUDE.md or memory take precedence over these lines unless the line comes from managed settings. The old includeCoAuthoredBy key, deprecated since v2.0.62, is ignored once commit or pr is set.
CLEAN UP HISTORY
The setting only affects new commits. For ones already made, rewrite the messages, ideally before anyone else has pulled them.
BEFORE AND AFTER THE REBASE
What the rebase does to your commit list
Only commits after origin/main are replayed. Each loses the Claude co-author and gets a new hash. Pushed commits keep theirs.
Commits on main, local
Every rewritten commit gets a new hash, so avoid this on shared branches.
The -i flag catches both Co-Authored-By and Co-authored-by spellings.
Terminal
git log -i --grep="Co-Authored-By: Claude" --oneline
git log --grep="^Claude-Session:" --onelineThis rewrites the latest message without the Claude lines and keeps everything else.
Terminal
git log -1 --format=%B \
| grep -viE "^(co-authored-by: claude|claude-session:)" \
| git commit --amend -F -Rebase replays each commit after origin/main and runs the same cleanup on it. Every rewritten commit gets a new hash, so avoid this on shared branches.
Terminal
git rebase origin/main --exec 'git log -1 --format=%B \
| grep -viE "^(co-authored-by: claude|claude-session:)" \
| git commit --amend -F -'Anthropic's watermark adds none, but pasted snippets can. The -P flag needs GNU grep, so on a Mac use ggrep from Homebrew, or git grep if your git has PCRE support.
Linux
grep -rnP "[\x{200B}-\x{200D}\x{FEFF}]" .Any git repo
git grep -nP "[\x{200B}-\x{200D}\x{FEFF}]"FOR THE PROSE IN YOUR REPO
Code is mostly exact, so the part worth checking is the writing around it. ClaudeClean treats any pasted text alike.
> source text
Drop in Markdown, a docstring or a PR description, or upload a .txt file. 2,000 characters work without an account.
scan report
3 zero-width1 soft hyphencleanable signals2Zero-width spaces, byte-order marks, soft hyphens and homoglyphs are flagged and removed in your browser. Nothing is stored.
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓The optional rewrite targets AI phrasing in documentation. It runs through OpenAI, is best effort, and is not meant for source code.
PRICING
The scan and hidden-character cleanup are free. Pro adds room for rewrites and files.
CLAUDECLEAN PRO
Cancel anytime.
Free: 10 credits a month with an account, or try 2,000 characters right here without one. Strong rewrites cost 2× credits.
OPEN SOURCE RULES
Some want a tag, some refuse AI code outright. Check the project before you strip or keep a trailer.
TRAILER ON, TRAILER OFF
Deleting the trailer moves no project into a different lane
The four policies quoted below, grouped by what they ask of AI-assisted code. Toggle the trailer and see what changes.
Allowed with a disclosure tag
Assisted-by: LLM tag, no AI Signed-off-by
Fix retry logic in upload queue
Co-Authored-By: Claude
Assisted-by: LLM
Needs approval first
LLM code presumed tainted, prior written approval from core
Fix retry logic in upload queue
Co-Authored-By: Claude
Not accepted
Declines AI-derived contributions
Fix retry logic in upload queue
Co-Authored-By: Claude
Forbidden since April 14, 2024
Fix retry logic in upload queue
Co-Authored-By: Claude
Grouping is ours, rules are each project's own. The kernel lane shows the tag its guidance asks for, which you can set through attribution.commit.
LINUX KERNEL
QEMU
GENTOO
NETBSD
Removing a trailer does not change a project's rules
WATCH
Anthropic's own walkthrough of the explore, plan, code and commit loop. The commit at the end is where attribution gets written.
YOUR CALL
There is no rule that you must keep it, and none that you must remove it. It depends on where the code goes.
SOURCES
Anthropic's own docs first, then the projects whose rules are quoted above.
Claude Code docs
Settings reference: Git and attribution
attribution, attribution.commit, .pr, .sessionUrl, the v2.1.281 false value and the deprecated includeCoAuthoredBy.
View sourceAnthropic
Claude text watermark
Why code carries less watermarking, comments as the exception, no hidden characters, C2PA on files.
View sourceLinux kernel docs
AI Coding Assistants
No AI Signed-off-by tags; the Assisted-by: LLM attribution format.
View sourceQEMU
Code provenance policy
Declines contributions that include or derive from AI-generated content.
View sourceGentoo Wiki
Council AI policy
Contributions made with natural language AI tools are forbidden, voted April 14, 2024.
View sourceNetBSD
Commit guidelines
Code generated by an LLM is presumed tainted and needs prior written approval by core.
View sourceFAQ