Twenty, or a CRM your coding agent writes as code you own?

Where Twenty wins

As of the research date this page carries, August 4, 2026, Twenty is the largest project in the researched set by GitHub stars — 54.3k — with YC S23 backing, roughly $5.5M raised, and weekly releases. It runs as a NestJS/React/PostgreSQL platform you can self-host with Docker, or as a cloud product; the per-seat cloud pricing in the research file came from third-party teardowns rather than a primary page, so treat the figures there as indicative and do not budget from this page. The headline fact is simpler than any of that: it is a CRM a technical team can run and use, and this framework is not.

Its coding-agent story is real rather than a slide. The research file records an MIT-licensed SDK, a create-twenty-app command, a CLAUDE.md checked into the repository, AI agents inside the product, and native MCP confirmed for cloud workspaces — with self-host parity explicitly unverified. Measured on getting an agent productive against a running CRM this afternoon, that is ahead of where this project stands: the bootstrap here runs from a checkout rather than from the registry, where the reserved `npm create` name still installs nothing, so ownership means copying source and upgrading means merging (L-08), and the MCP server here is stdio and local only, with no hosted or authenticated endpoint (C-18).

And the plain asymmetry. Twenty 2.0 added an apps platform, in-product AI agents and git-backed workspace config, and its customization surface spans a settings UI, APIs and an Apps SDK covering entities, serverless logic functions and UI widgets. This framework has no authentication, tenancy or RBAC at all (L-01), stores everything in SQLite (L-02), and is not a product you sign up for (L-07). If what you want is a CRM, that comparison is not close, and pretending otherwise here would forfeit your trust in everything else on this site.

Where this is different

There is one axis, and it is where the extension lives. In the research file's own characterisation, Twenty's customization is deployed into their runtime: the agent writes extensions against their abstractions. Here the agent writes the application. A customer-authored domain package attaches and detaches with the kernel's fingerprint unchanged, and reaches another package only through a capability it declares (C-13). A module manifest becomes a migration, a service, a REST resource, an SDK method and Admin screens with no page code (C-01). One command reads checked-in source and reports what an application actually is — packages, capabilities, resources, actions, policies, providers — as deterministic JSON, then lists its own blind spots (C-14).

Licensing shape, with the counterweight stated in the same breath. The file records Twenty as an AGPL core plus @license Enterprise files covering SSO and advanced RBAC, alongside an MIT SDK, and describes the resulting lock-in as the platform runtime plus AGPL share-alike for derivatives plus those gated files. This repository is MIT with one edition and nothing gated (ADR-023). The counterweight matters more than the contrast: their advanced RBAC sits behind an enterprise licence, and here it does not exist in any edition at all (L-01). Not gated is not the same as included. The file also states that its licence descriptions are factual summaries rather than legal advice, and share-alike questions belong with counsel, not with a comparison page.

The agent surface differs in kind rather than degree. Code generation and destructive operations are dry-run unless an explicit apply flag is passed (C-18), and the first thing a coding agent reads is a deterministic report of checked-in source rather than the state of a running workspace (C-14). That distinction only matters if your review process is a human reading a diff — which is the assumption this whole framework is built on, and is a preference rather than a proof of superiority.

How to choose

Choose Twenty if you want a CRM to use and extend rather than a substrate to build one from; if a hosted option, a community and weekly releases matter; if a platform runtime is a feature rather than a constraint; and if AGPL share-alike on derivatives plus enterprise-gated SSO and RBAC are acceptable for what you are shipping.

Choose this framework in a narrow case: the deliverable is one bespoke commercial process — approvals, discount policy, obligations, handovers — that must exist as source in your repository, reviewed as a diff, with the approval boundary asserted by a test rather than configured in a UI. That case also requires you to supply authentication, hosting, PostgreSQL, import and export yourself, because none of them is here.

If your actual requirement is SSO or role-based access control, the honest answer is that neither column is this one: the research file puts Twenty's behind an enterprise licence, and this project does not have it in any form.

The research file names Twenty as the strongest competitive threat to this project and says the convergence is in narrative more than in product. Read that as a reason to check its current state yourself before deciding, not as a verdict either way.

What this comparison does not cover

No hands-on evaluation. Nothing from Twenty was installed, configured, benchmarked or run in producing this page. Every fact traces to docs/strategy/COMPETITOR_MAP.md, researched on August 4, 2026, and this environment has no network access to re-verify any of it.

No feature matrix. There is no statement here about Twenty's CRM features, data model, UI, performance, support or roadmap, because that file does not carry them and inventing them would cost more credibility than the comparison is worth.

The hedges the source carries are repeated rather than dropped: cloud pricing is third-party sourced, self-host MCP parity is unverified, and self-host upgrade fragility is recorded there as a documented issue in their tracker (#14705) rather than something reproduced here.

No build-rate or speed comparison in either direction. The benchmark has not been run, and any number quoted for this project is not ours (L-03).

The source's own standing caveat, which is the most important sentence on this page: Twenty could extend its MIT SDK layer downward into genuine code ownership, and that would move this comparison. A dated page is only as good as its date — when that research is refreshed, this entry must be re-read against it.