MCP overview

Darwinium Portal MCP lets you talk to your Darwinium portal from Claude. Open your portal tab, ask Claude a question in plain English, and Claude reads what is on screen, runs the right queries, and gives you back an answer, a table, a new dashboard card, or a full report.

Claude only ever sees what you can already see. It uses the portal tab you are logged in to, with your permissions. There is nothing to set up on the Darwinium side.

Setting it up takes about five minutes, once. If you would rather someone technical did that part, send them the technical guide.

Contents

1. What you can do with it
2. A worked example: the weekly risk trends report
3. How it works
4. Getting set up
5. Getting good results
6. Keeping your data safe
7. If something is not working

1. What you can do with it

Once connected, you type a question into Claude the way you would ask a colleague. Here are the kinds of things it can do for you today.

Understand what is happening

• "Summarise what is on this page."
• "Which signals fired most in the last 7 days, and how does that compare with the week before?"
• "Show me how login events flow into decisions this week. Which steps end in a reject?"
• "Is the account email actually populated on our payment step? What are the most common values?"

Write and check queries

• "Write a query for transfers over $5,000 from a device first seen today, and tell me how many events it matches."
• "What attribute holds the ASN of the connection? Show me the exact syntax."
• "Open Investigations with the last 24 hours of rejected payments and run it."

Investigate an event or identifier

• "Open the event in the sidebar and explain why it scored the way it did. Which signals contributed most?"
• "On this identifier graph, which devices connect the most accounts?"
• "Has this device ever been labelled? What labels and by whom?"

Tune a policy

• "For the payment step, which signal has the biggest score but fires least often? Would lowering its weight change many decisions?"
• "What is the 99th percentile of this feature? I want a threshold that only catches the top 1%."
• "Draft a rule for cloned-app deposits followed by a transfer within an hour, using only feature names that exist in this workspace, and check it against real data."

Build dashboards

• "Create a board called 'Mule watch' with a stat tile for money out, a Top X of device signatures across accounts on the transfer step, and a Sankey from step to decision. Make the Sankey full width, then take a screenshot to check it rendered."
• "Add a Signal Trigger Rate card for the login journey with day-over-day change."

Produce something you can share

• "Write a weekly risk trends report for the last 7 days as a single web page I can send to the fraud team."

Claude works out which portal actions to use. Behind each answer it typically checks your real step and signal names, validates any query it writes, counts the matching events, and looks at a screenshot of anything it built. You just see the answer.

2. A worked example: the weekly risk trends report

This report was produced from a single prompt, against a sample-data tenant. The output is a standalone web page: the decision outcome split, money moved and blocked, the week's detections with day-over-day change, a ranked table of signal movers, and four case studies each with the exact Investigations query to reproduce it.

Open the full page: examples/weekly-risk-trends.html. It is sample data from the demobank tenant.

The prompt that produced it

"Build me a weekly risk trends report for the last 7 days as a single, self-contained web page I can send to the fraud team. Cover: the decision outcome split across all traffic, money moved versus money blocked, which signals rose or fell most day-over-day and what that means, and three or four case studies of the patterns behind the biggest moves. For each case study include the exact Investigations query to reproduce it and the attributes it keys on. Use our real step and signal names. Validate every query and quote the 30-day count."

What Claude did to get there

1. Looked up the real step, journey and signal names for the tenant, so nothing in the report is invented.
2. Ran the step-to-decision flow and the money aggregates for the outcome bar and the four headline tiles.
3. Pulled signal trigger rates with a previous-period comparison for the movers table.
4. For each case study, read the rule behind the signal, wrote a reproduction query, validated it, and quoted its 30-day count.
5. Checked that the attributes each query relies on are actually populated, so an empty result is never presented as "no fraud".
6. Wrote the page.

Every number came from your portal through your login, and every query in the report can be pasted straight back into Investigations.

3. How it works

Two small pieces are installed on your computer:

• A Claude add-on (the "MCP server") that lets Claude ask your browser to do things.
• A Chrome extension that receives those requests and carries them out inside your Darwinium portal tab, exactly as if you had clicked the buttons yourself.

Nothing new connects to Darwinium. The portal tab you are logged in to does all the work, with your permissions, so Claude cannot see anything you could not see yourself. The one thing that leaves your computer is your conversation with Claude, which goes to the AI service your organisation uses for Claude. See Keeping your data safe.

For the detailed architecture, see the [technical guide](/v1/docs/
nical-details#1-architecture)

4. Getting set up

You need Google Chrome, the Claude Desktop app, and your Darwinium portal login. If your team uses Claude Code instead of Claude Desktop, the steps are in the technical guide.

Step 1: Install the Claude add-on

1. Download the Darwinium Portal MCP file from the download page. It ends in .mcpb and works on Mac and Windows.
2. Double-click the file. Claude Desktop opens on its Extensions screen.
3. Click Install.

If double-clicking does nothing on a Mac, the file has been quarantined by macOS. The fix is in the technical guide, or ask IT.

Step 2: Install the Chrome extension

1. Open the Darwinium Portal MCP page on the Chrome Web Store. It only appears via this direct link.
2. Click Add to Chrome and accept the prompt.
3. Pin it so you can find it: click the puzzle-piece icon in Chrome's toolbar and pin Darwinium Portal MCP.

If your organisation blocks the Chrome Web Store, IT can install it another way. See the technical guide.

Step 3: Open your portal tab

In Chrome, go to your Darwinium portal and log in as usual. Leave that tab open and in front.

Step 4: Ask Claude for the pairing token

1. In Claude Desktop, start a brand-new chat.
2. Ask: "What is my Darwinium pairing token?"
3. Claude replies with a long string of letters and numbers. Copy it.

It has to be a new chat. Claude is only given the token when a conversation starts, so an older chat will not know it.

Step 5: Paste it into the extension

1. Go back to Chrome and click on your portal tab first.
2. Click the Darwinium Portal MCP icon in the toolbar.
3. Paste the token into the box and click Save & Connect.

The popup switches to Connected and shows your portal address. The portal's title bar also shows MCP Connected.

You are done. Go back to Claude and ask it something about the page you are looking at. You will not need to pair again unless you change computers or IT resets the connection.

5. Getting good results

• Have the right page open. Dashboard actions only work on the Dashboards page, timeline actions only on the Journey tab, and so on. If Claude says it cannot do something here, go to the right page, or ask Claude to take you there.
• Ask for the working. Claude validates queries and counts matches before showing them to you. If you want to see the counts or the queries it ran, ask.
• Claude asks before changing anything. Adding or removing labels, saving dashboard cards, deleting boards and editing policy files all change your portal for everyone. Claude will confirm with you first. Read the confirmation.
• After moving to a different page, start a new question. If results look wrong after you navigated, ask Claude to check what the page can do now.
• Claude Desktop and Claude Code cannot both be connected at once. If you use both, close one.

6. Keeping your data safe

• Claude sees only what you see. It works through your logged-in portal tab with your permissions. It has no Darwinium password or key of its own and cannot log in without you.
• Darwinium receives nothing new. The portal tab talks to Darwinium exactly as it always has. The add-on and the extension never contact Darwinium at all.
• Your conversation goes to the AI service your organisation uses for Claude. Anything Claude reads from the portal to answer you is part of that conversation. Treat it as you would pasting portal data into a chat, and do not ask about data you would not be comfortable sending.
• We recommend running Claude through Amazon Bedrock with Guardrails. This keeps your conversations with Claude inside your organisation's own AWS account, under your existing security controls, and lets your IT team automatically mask or block sensitive information such as customer names and account numbers in both directions. Amazon Bedrock Guardrails is AWS's service for exactly this: Amazon Bedrock Guardrails. The setup steps for IT are in the technical guide.
• You can disconnect at any time by clicking Disconnect in the extension popup, or by removing the extension from Chrome.

Darwinium's privacy policy: https://www.darwinium.com/privacy-policy.

7. If something is not working

Claude says the extension is not connected. Click the extension icon. If it says Disconnected, click your portal tab, then click Connect. If your computer just woke from sleep, give it 30 seconds.

Claude does not know what a pairing token is. Start a brand-new chat and ask again. If it still does not know, the add-on from Step 1 did not install. Check Settings → Extensions in Claude Desktop.

Claude says the tools are unavailable or the server disconnected. Only one Claude app can be connected at a time. Close any other Claude app you have open, then quit Claude Desktop fully (Claude → Quit, not just closing the window) and reopen it.

The extension says "Token mismatch". Click Re-pair in the popup, then repeat Steps 4 and 5.

Claude connects but says the page is not ready. Reload the portal tab and try again.

Anything else. Email support@darwinium.com with a description of what you were doing, or hand the technical guide to IT, which has a diagnostic they can run.

Links

• Technical guide (for IT and support): MCP Technical Details
• Chrome extension: https://chromewebstore.google.com/detail/darwinium-portal-mcp/jmhgcmlbdgepocibillmajdpmblocomd
• Claude add-on download: https://github.com/darwinium-com/portal-mcp/releases/latest
• Support: support@darwinium.com