Technical reference: the MCP Server pages are the canonical documentation for endpoints, per-client setup, OAuth scopes, the twelve tools and their parameters, limits and troubleshooting. This page gives the general picture; go there for anything technical.
Before you start
Two things, and they take a moment to check.1
A Copyleaks account you can already scan with
Sign in to Copyleaks and check you can run a scan there on a paid balance. A subscription, a prepaid balance, or an account your organization pays for all work.
2
An AI agent that supports MCP
Copyleaks works with any agent that speaks MCP - it’s an open standard, not a per-agent integration. These are the ones we’ve tested and written setup steps for:
Using something else? The same connector address works - follow your agent’s own instructions for adding a remote MCP server.
If you’re using Claude, use Claude Code - not claude.ai
Claude Code isn’t only a terminal tool - it also runs as a desktop app, on the web, and as an IDE extension. All four forms have the full capability, including the desktop app.
Only use claude.ai or Claude Desktop if pasted text and questions about past scans are genuinely all you need - that much works well. For anything involving a file or an image, use Claude Code.
The six steps
Add Copyleaks to your agent
You don’t need to wait for Copyleaks to appear in any connector store. Add it yourself, today.The server is called
copyleaks_app, and its address is https://extensions.copyleaks.com/mcp.Claude Code - one line:Add the server
--scope user makes it available in every project on your machine. Then run /mcp inside a session to sign in.Claude Desktop or claude.ai - go to Settings → Customize → Connectors, press + → Add custom connector, paste the endpoint, then Add → Connect.Every other client - Gemini CLI, Codex CLI, Antigravity, Cursor, VS Code, or anything else - needs its own config file and its own field name for the address, and getting that field wrong is the usual reason a setup silently fails. Copy the exact block for your client from Client setup.Approve what it can do
Your browser opens, you sign in to Copyleaks, and you see exactly what you’re granting:
Approve, and you land back where you started - your connector settings, or your terminal.
Your agent can only do what you could do in Copyleaks. It can never buy credits, change your plan, delete anything, create a share link, download a report, hand you back a document you scanned, or open a scan someone else shared with you.
Start a new session
Do this before anything else - it’s the step people skip.Agents read their connector list when a session begins. If you added Copyleaks in the middle of a conversation, it isn’t there yet, and asking about it will look like the setup failed.
Using it in Claude Desktop - you add the connector on claude.ai; there’s no separate setup in the desktop app, because connectors follow your Claude account. After quitting and reopening: start a new chat, open the tools menu in the message box, check Copyleaks is listed and enabled, then just ask - “Check my Copyleaks credit balance.”If Copyleaks isn’t in that list, you’re either signed in to a different Claude account than the one you added it on, or the app didn’t fully quit.
Check it worked
Ask your agent:
“Check my Copyleaks credit balance.”If it comes back with a number, you’re connected. That’s the whole test.
Run your first scan
Paste a few paragraphs and ask:
“Check this for plagiarism and AI.”Your agent submits the scan, then checks back until it’s finished. Scans aren’t instant - while it runs you can ask how far along it is and get a percentage. You won’t get an estimate of when it’ll finish; Copyleaks doesn’t predict that.What comes back: your
Plagiarism score and your AI score, as percentages to one decimal place, plus which checks actually ran.Other ways to start a scan:A scan is text or image, never both. Asking whether a photo is AI-generated and asking whether the words in it are plagiarised are two different scans.
Go deeper on the result
This is the part that isn’t in the report. Keep asking:
“Which sources did I match?” “Show me the exact text that overlapped with the third one.” “Walk me through everything flagged, in the order I wrote it.” “Which of those are just my citations?” “Are those thirty actually thirty different sources, or the same article republished?”Nothing is withheld. Your agent starts narrow and widens on request - score, then sources, then the exact overlapping text on both sides. You pull the depth the question needs.Your agent uses the same words your report does:
Exact match · Minor changes · Paraphrased for match types; Web · Shared data hub · This Batch · Private cloud hub for where a match came from; AI content and the AI phrases inside it; Excluded text for what your Exclusions left out.To see the document itself, open the report. Every finished scan comes with a link - Open this scan in Copyleaks. Nothing is hidden behind it; it’s just where your marked-up document lives.Nine things worth knowing on day one
1. It uses your normal credits
1. It uses your normal credits
Same balance, same meter, whichever way you scanned. There’s nothing extra to buy and no separate agent balance. Your agent can’t tell you what a scan will cost, before or after - your usage meter is the record.
2. Results are what was flagged, not your document
2. Results are what was flagged, not your document
Matched passages, AI content, AI phrases, excluded text. The document itself stays in Copyleaks. That’s permanent, not a gap.
3. Links that need a sign-in don't work
3. Links that need a sign-in don't work
Copyleaks opens a link without signing in - so Google Docs, Drive, SharePoint, OneDrive, Notion, Dropbox and anything on a company network return a login page, not your document.
4. Large files need an agent with a shell
4. Large files need an agent with a shell
Small files go straight through everywhere. This is permanent on the two chat surfaces, not a temporary limitation.
5. Image scanning needs an agent with a shell
5. Image scanning needs an agent with a shell
The same line as fact 4. An image always travels through the upload step, so the Claude chat apps - claude.ai and Claude Desktop - can’t do image scans at all. Claude Code can, its desktop app included. Upload the image in Copyleaks instead; your agent can read the finished scan back.Images must be 512 × 512 or larger, 27 megapixels or smaller, and PNG · JPEG · BMP · WebP · HEIC · HEIF.
6. Long text goes as a file
6. Long text goes as a file
Pasted text is capped at 25,000 characters. A file has no character ceiling - so if the paste is refused, send the same content as a file.
7. 10 scans a minute, the same on every paid plan
7. 10 scans a minute, the same on every paid plan
Each file and each web address counts as one, so ten files is ten scans against the minute. Going over doesn’t fail anything - your agent starts what fits, queues the rest, and carries on by itself. There’s no faster tier to buy.
8. Your own scans only
8. Your own scans only
Anything a colleague shared with you stays in Copyleaks. And if your team wants this, each person connects their own account - there’s no shared connection and no team-level scan profiles.
9. Nothing it does deletes anything
9. Nothing it does deletes anything
Not scans, not results, not scan profiles. Deleting is done in Copyleaks, deliberately.
If something doesn’t work
Full list: Troubleshooting.
One known build gap is left, and it isn’t a failure: a score in a list of scans is shown as a whole number, while the scan itself carries one decimal. Where the two differ, the scan is the accurate one.
Where to go next
Tools
Everything you can ask for - the twelve tools and their parameters.
Limits and access
Credits, throughput, who can connect, and how to disconnect.
Client setup
Per-client config blocks and the field name each one expects.
How results work
Scores, matched sources, AI content and AI phrases.