Unit21 MCP, customizable webhooks, and everything else we shipped in August

This was a good month for anyone who wants their tools talking to each other. Unit21 MCP went live, so any AI assistant can connect directly to your Unit21 environment. Webhook payloads stopped being locked to a fixed set of fields. goAML expanded into two new countries, CTR filing closed a couple of long-standing gaps, and we're giving you an early look at what's coming next for detection. Here's what's new.

‍

Unit21 MCP: connect your AI assistant directly to Unit21

Unit21 now has a remote MCP (Model Context Protocol) server, which means Claude, ChatGPT, Cursor, or any other MCP client can connect straight to your Unit21 environment. There's nothing to install: you give your AI assistant the server URL for your environment, sign in to Unit21 the way you already do, and the assistant picks up from there.

‍

The important part is what the assistant can see once it's connected: only what you can already see. Access follows your existing Unit21 role and permissions, so an AI assistant connected this way can search Unit21 docs, look up your agent and org context, or run reporting dashboards you have access to, and nothing more. An org admin turns MCP on for the organization and grants agents MCP read access before anyone can connect, and disabling either one immediately cuts assistants back off.

‍

Once it's on, the value is obvious: analysts spend a lot of a case working in whatever tool they think in, not just in Unit21. Now that context can come to them, instead of the other way around.

‍

Webhook payloads you can actually customize

Webhook payloads used to be locked to a fixed set of fields, whatever your system needed or not. Now you can add the specific fields your systems actually use. It's a small change with a real payoff: alerts and case events can now trigger automation anywhere you want to route them, whether that's Salesforce, an email, a phone alert, or somewhere else entirely, without your team building a workaround to get the right data into the payload.

‍

goAML goes global: Malta and the Netherlands are live

goAML filing picked up two new jurisdictions this month. Malta is now a supported country for STR (Suspicious Transaction Report) filing to the FIAU, with Malta-specific validations built in, currency in euros, nationality filled in, at least one supporting document attached, so a report can't go out incomplete. The Netherlands got the same treatment for UTR (Unusual Transaction Report) filings to FIU-Nederland, including crypto and VASP transactions. Both are export-only: you build, validate, and export the filing in Unit21, then submit the generated file to the regulator yourself, on the same modernized goAML foundation we shipped in July.

‍

CTR filing closes two real gaps

Two changes land directly in the CTR filing workflow this month. First, analysts can now add entities and transactions to a CTR the same way they already could on a SAR, so the person who walks into the branch with cash, who is often not the account holder, no longer has to be typed in by hand every time. Second, whoever created a CTR now gets an email the moment its filing fails, whether that's a FinCEN acknowledgement reject, a warning, or a submission failure. Notifications are on by default. A failed CTR used to sit unnoticed until someone went looking for it; now the filer knows immediately and can act on it.

‍

Three more worth knowing about

Add an alert to any case. Adding an alert to a case used to mean picking from cases already associated with that alert. Now, with the right permission, you can search across all cases and link the alert to wherever it actually belongs.

‍

AI Raw Data View. Now GA for all AI Paying and POC customers. It gives you a view into all the data flagged on an AI snapshot, so when a customer or auditor needs proof of exactly what context an AI agent had access to, that answer is surfaced directly instead of requiring a manual dig through the underlying articles.

‍

Preview an AI agent task against multiple alerts at once. Task previews for an online research task can now run against up to five alerts, cases, or filings at once, with results shown side by side, and errors on an individual item surface in the preview panel instead of failing silently. Configure a task once, see how it performs across a representative sample, instead of testing item by item.

‍

Coming soon: AI for Detection, Rule Writer Agent

Most fraud and AML rules start as a sentence someone describes to an analyst, long before they're fields, operators, and thresholds. Rule Writer Agent is what we're building to remove that translation step. Describe the pattern you're trying to catch in plain language, and it builds the deployable rule underneath it, variables, thresholds, and trigger conditions included.

‍

It runs a guided, multi-turn flow rather than a single prompt-in, rule-out shot: a guardrail check screens the request, the agent evaluates it against your actual data model, and if something's ambiguous, a threshold left vague, a population undefined, it asks rather than guessing. Every rule it builds comes with a plain-language description next to the compiled logic, so the next person who opens it doesn't have to reverse-engineer what it does.

‍

It's a companion to Rule Recommendations, not a replacement: Rule Writer Agent is how you build a new rule from an idea, Rule Recommendations is how you keep the rules you already have performing well. More on this soon.

‍

Want to see what's live today in action? Reach out to your CSM or get a demo to see Unit21 MCP, customizable webhooks, and the rest of August's releases firsthand.

‍