AI LEAD SCORING FOR TWENTY CRM

Lead Scoring Agent for Twenty CRM

A free, open-source Twenty app that scores every person and opportunity from 0 to 100 and says why in a sentence or two. Open the view sorted by score and work the best leads first.

Twenty CRM AppMIT LicenseCloud and Self-HostedNothing to Configure

Overview

Twenty has no built-in way to tell a rep which new lead deserves a call first. New people and opportunities arrive as flat rows, and the rep works them in the order they landed or opens each one to guess. That gets slower as volume grows, and the hot lead waits behind the cold ones.

The Lead Scoring Agent adds four fields to People and Opportunities: the score, a short reason, when it was scored and a status. Two ready-made views, People by Lead Score and Opportunities by Lead Score, list the highest scores at the top. The reason sits next to the number so a rep can judge whether to trust it.

I built it because the Twenty SDK documentation uses lead scoring as its example agent but ships nothing you can install. The code is on GitHub and the package is on npm.

At a Glance
What you need to run it

Twenty server 2.49 or later, Cloud or self-hosted

A default AI model set in Settings, AI

Free and MIT licensed

One AI call per scored record, billed to your Twenty AI credits or your own provider

Records with nothing to judge are skipped and cost nothing

TypeScriptTwenty SDKVitest

How It Works

Scored in Seconds

A lead created by hand or through the API is scored inline, so the number is there by the time the rep looks. An edit that changes what matters triggers a re-score: job title, company or email for a person, and amount, stage or close date for a deal.

Imports Go Through a Queue

A CSV import of thousands of rows would otherwise fire thousands of AI calls at once. Imported records are marked queued instead, and a one-minute job scores 20 per object per run. Records that existed before install are queued the same way.

Credits Run Out Gracefully

If the workspace is out of AI credits, records stay queued and are scored when credits return, instead of failing. A genuine error marks the record failed with the reason in the summary field, and the status page can re-queue it.

What the Score Weighs

Fit (seniority, title, company), intent (stage, amount, recent activity) and how complete the record is. The agent is told to use only the data it was given and to score conservatively when key fields are missing.

A Status Page in the Sidebar

The Lead Scoring Agent page shows how many People and Opportunities are scored, queued, skipped or failed, with a button to queue anything not yet scored.

Least-Privilege Access

The app reads People, Companies, Opportunities and timeline activity, and updates only People and Opportunities. It has no delete permission, holds no API keys and calls no outside service.

Install

From the Twenty UI

Open Settings, Applications, Marketplace, find Lead Scoring Agent and click Install. If it is not listed, open the filter and turn on Show unlisted apps.

Make sure the workspace has a default AI model under Settings, AI. Nothing else needs configuring.

From the Twenty CLI
bash
npx twenty@latest auth login --url https://<your-workspace>.twenty.com

npx twenty@latest app install \
  --universal-identifier 94f2e86e-62e5-4695-b4e8-50059cd7db30

Self-hosted servers pick up npm-published apps with the catalog sync. See the README for the full details.

Design Decisions

Spot Imports From the Record, Not by Counting

My first plan was to count how many records were created in the last few seconds and call it a bulk import above some threshold. It adds a query to every event and misjudges small imports. Every Twenty record carries a created-by source, and importing sets it to IMPORT, so the handler routes on that field instead.

One AI Call per Record

Structured agent output has to be a flat schema, so several records cannot share one call. Cost is roughly one call per scorable record however you trigger it. The queue limits how fast calls happen, and the skip rule and the field-scoped triggers limit how many.

The Agent Never Re-Triggers Itself

Update triggers watch business fields only. The score, summary, time and status fields are not on the list, so writing a result cannot start another scoring run.

Test Against Real Records

Two bugs only showed up on a real workspace: existing records were never scored because events only fire for new data, and recent activity was silently missing after Twenty renamed a timeline field. Both are fixed and covered by tests, and the changelog lists them.

Privacy

The only thing that leaves your workspace is the scoring prompt, sent through Twenty's AI service to your default model. For a person that is name, job title, primary email, LinkedIn URL and the company's name, domain and city. For an opportunity it is name, amount, stage, close date and the company's name and domain. Both also send the event names of the 10 most recent timeline entries, such as noteLinked, and never their content.

Email bodies, notes and attachments are not sent. The score, summary, time and status live on the record in your workspace, and the app keeps no other copy. Uninstalling removes the four fields and their values.

Frequently Asked Questions

Related Projects

CRM Integration
Method CRM MCP Server
Production-ready MCP server with 20 tools for Method CRM
AI Automation
Trending Repo Scout
AI-scored GitHub trending digest with a live trend dashboard
AI Automation
Reddit Comment Engagement Agent
Compliance-first Reddit engagement with AI quality scoring and human approval

Notes

Lead Scoring Agent is an independent project. Twenty is a trademark of its owners, and this app is not affiliated with or endorsed by them.

Bug reports and questions go to the GitHub issue tracker. If you want this set up or adapted for your own Twenty workspace, get in touch.

Try It on Your Workspace

Install the app, let it score your existing records, then open People by Lead Score