Twenty: An Open-Source CRM Designed for AI Agents

CRM as Code

Twenty shifts CRM management from a GUI-based configuration to a developer-centric workflow. Users define objects and fields using TypeScript files within a project scaffolded via create-twenty-app. This approach allows teams to version their entire CRM schema in Git, manage changes through pull requests, and perform rollbacks using standard CI/CD pipelines. The underlying architecture is a TypeScript monorepo using Nx, Nest.js, Postgres, and Redis, which provides a familiar stack for full-stack developers to extend or audit.

AI-First Architecture

The platform integrates a native Model Context Protocol (MCP) server that exposes CRM objects as tools for LLMs. This enables agents like Claude to execute complex operations, such as creating companies, logging deals, or querying pipeline status, through plain-English prompts. Developers can configure read and write permissions per object to constrain agent access. Version 2.6 introduced tool annotations to improve agent reliability and an OpenAI integration that allows the CRM to function directly within ChatGPT. The system also features an auto-syncing catalog for AI models, ensuring the platform supports new models without requiring manual updates.

Self-Hosting and Enterprise Trade-offs

Twenty is AGPL-licensed and deployable via Docker Compose, allowing teams to maintain complete data sovereignty without seat caps or vendor lock-in. While it offers a modern, performant alternative to Salesforce or HubSpot for small-to-medium teams, it lacks the complex governance, deep approval chains, and multi-departmental permission structures required by large enterprise organizations. Self-hosting requires the team to manage their own infrastructure, including database maintenance and version upgrades.