Skip to content
Data

MIT edges out AGPL-3.0 across 457 self-hosted apps, and 53 carry an open-core, source-available, project-specific, proprietary or dual license

Nov 12, 2026 · 8 min read

data as of · Nov 2026
Data & methodn=457
Datalatest
n
457 apps
measured
Nov 2026
method
How we test
dataset
latestdoi:10.5281/zenodo.22983606
license
CC BY 4.0
Cite this
SelfHost Atlas (2026). MIT edges out AGPL-3.0 across 457 self-hosted apps, and 53 carry an open-core, source-available, project-specific, proprietary or dual license. https://selfhostatlas.com/blog/self-hosted-app-licenses. Data: doi:10.5281/zenodo.22983606

The licenses of the 457 apps in our catalog as we record them: 117 MIT, 112 AGPL-3.0, 68 Apache-2.0, 50 GPL-3.0, and a long tail of 54 other strings. 202 are copyleft, 202 permissive and 53 something else, and the measured apps in the something-else group idle at a median 607 MB against 125 MB for the permissive ones.

apps under MIT vs AGPL-3.0, of the 457 in the 2026-11-13 export
117 vs 112
apps in the data
457
data as of
Nov 2026

Correction, 2026-11-14. The figures in this post were updated on 2026-11-14 from dataset release v1.11.0. An earlier version, on the 2026-10-16 export of 256 apps, said AGPL-3.0 led MIT (72 to 66) and copyleft led permissive (119 to 108); with 201 more apps in the catalog, MIT leads AGPL-3.0 117 to 112 and the copyleft and permissive buckets are level at 202 each.

Every app page on this site shows a license. Taken together, the 457 strings say something about the self-hosted software we have chosen to cover (a curated catalog, not a census): MIT narrowly ahead of AGPL-3.0, copyleft and permissive level on our grouping, an AGPL lead over the GPL, and a long tail of licenses that are neither.

The four that cover three quarters

License Apps
MIT 117
AGPL-3.0 112
Apache-2.0 68
GPL-3.0 50
Everything else (54 strings) 110
Figure 1 · source: catalog.json, exported 2026-11-13 · n=457

Four licenses cover 347 of the 457 apps (76%), and the top two alone cover 50%. The AGPL lead over the ordinary GPL is clear: counting every string that starts with "AGPL-3.0" (the plain one, the or-later variant and six hybrids) gives 119 apps, against 78 for every GPL and LGPL string together.

Copyleft, permissive, or neither

We sorted the 58 distinct strings into three buckets. The first two are the standard licenses on their own; the third is everything that is not a plain standard license.

  • Copyleft (202 apps): AGPL-3.0, AGPL-3.0-or-later, GPL-3.0, GPL-3.0-or-later, GPL-2.0, GPL-2.0-or-later, LGPL-3.0, LGPL-2.1-or-later, MPL-2.0, EUPL-1.2, EPL-2.0 and OSL-3.0, counting the file-level copyleft of MPL and EPL in.
  • Permissive (202 apps): MIT, Apache-2.0, BSD-3-Clause, BSD-2-Clause, Zlib and the PostgreSQL license.
  • Neither (53 apps): open-core strings, source-available licenses, project-specific licenses, three proprietary servers and two dual-licensed apps.

The two buckets are level, and small choices tip them: count the two dual-licensed apps, which both offer a permissive option, as permissive and permissive leads 204 to 202; count the nine MPL-2.0 and EPL-2.0 apps as permissive as well and it leads 213 to 193.

Figure 2 · source: catalog.json, exported 2026-11-13 · n=457

The 53 that are neither

This is the group worth reading one by one, because the string on the page is the whole story: these are the licenses where "open source" needs a footnote.

Open core (19). Nine apps record "MIT core with proprietary enterprise modules": Chatwoot, Infisical, Langfuse, Lightdash, LiteLLM, Medusa, PostHog, Stirling-PDF and Strapi. Metabase is the AGPL-3.0 version of the same arrangement, and Weaviate the BSD-3-Clause one. Flowise and Sim record an Apache-2.0 core with commercially licensed enterprise modules; Appsmith and Kestra an Apache-2.0 community edition with the enterprise edition as a separate image; Dittofeed and Novu an MIT codebase with a closed or commercially licensed enterprise part; ToolJet an AGPL-3.0 core with enterprise code in separate repositories; and Budibase GPL-3.0 overall with its pro packages under BSL-1.1.

Source-available and similar (14). Akaunting, Outline and HashiCorp Vault under BSL-1.1; n8n and NocoDB under the Sustainable Use License; Graylog under SSPL-1.0; Invoice Ninja under Elastic-2.0; Sentry under FSL-1.1-Apache-2.0; Bugsink under PolyForm Shield 1.0.0; Directus under MSCL-1.0. Four more are limited to non-commercial use: Drone (the Drone Non-Commercial License on the Enterprise Edition image), GameVault (CC-BY-NC-SA-4.0), RecipeSage (whose self-host images are for personal, non-commercial use) and SparkyFitness.

Project-specific or hybrid (15). AppFlowy (AGPL-3.0 client apps, a commercial license for the self-hosted server), AFFiNE (MIT client, AFFiNE EE License server, Community Edition parts MPL-2.0), Tandoor (AGPL-3.0 with Commons Clause), Dify (a modified Apache-2.0), Open WebUI (a modified BSD-3-Clause with a branding clause), NocoBase (Apache-2.0 plus supplementary terms), DataEase (GPL-3.0 with additional conditions), ESPHome (MIT for the Python part, GPL-3.0 for the C++ runtime), EtherCalc (CC0-1.0 with legacy files under other licenses), MiroTalk P2P (AGPL-3.0 with paid licenses for closed-source use), and Joplin Server, LobeHub, nopCommerce, PLANKA and Vtiger, each under a license of its own name.

Proprietary (3). Plex and Emby, both media servers, and Nexus Repository, whose Docker community edition is under proprietary terms.

Dual-licensed (2). Adminer (Apache-2.0 or GPL-2.0) and Eclipse Mosquitto (EPL-2.0 or BSD-3-Clause). Both offer a permissive option; they sit here only because their string is not one plain license.

The 53 are spread over 32 categories. Five supply 16 of them: low-code internal tools (four), and notes, business intelligence, AI agent platforms and AI chat interfaces (three each). Twenty-six of the 53 are listed in our catalog as an alternative to a specific SaaS product, a similar share to the catalog as a whole (196 of 457).

Where copyleft and permissive concentrate

For the 70 categories with at least four apps (449 apps between them), the share that is copyleft runs from all of a category to none of it. A selection, ordered by that share:

Category Apps Copyleft Permissive Neither Most common
Social network (Fediverse) 6 6 0 0 AGPL-3.0 (6)
Media automation 8 7 1 0 GPL-3.0 (5)
CRM 8 6 1 1 AGPL-3.0 (5)
Web analytics 8 6 1 1 AGPL-3.0 (4)
Project management 12 8 3 1 AGPL-3.0 (4)
Password manager 6 4 2 0 AGPL-3.0 (3)
Photo backup & gallery 6 4 2 0 AGPL-3.0 (3)
Media server 6 3 1 2 GPL-3.0 (2), Proprietary (2)
File sync & storage 6 3 3 0 Apache-2.0 (3)
Notes 8 3 2 3 AGPL-3.0 (3)
SSO & Identity 11 4 7 0 Apache-2.0 (4)
Monitoring 14 4 10 0 Apache-2.0 (5)
Code hosting 7 1 6 0 MIT (5)
AI chat interfaces 10 0 7 3 MIT (5)
Backend & BaaS 7 0 5 2 MIT (3)
Platform-as-a-Service 5 0 5 0 Apache-2.0 (3)

One reading of the table: the categories at the top (fediverse servers, media tools, password managers, photo libraries) are mostly copyleft, and AGPL-3.0 in particular, while the categories at the bottom (code hosting, backends, deployment platforms) are mostly permissive. The other 54 categories with four or more apps are not shown; most of them sit between these rows.

License bucket and idle RAM

For the 418 apps with a measured idle RAM, the medians by bucket are 125 MB for permissive (177 apps), 177.5 MB for copyleft (190) and 607 MB for the neither bucket (51). We have not tested why, and we would not read the license as the cause. Our stack-tag analysis found that idle RAM rises with the number of data stores an app ships, and the plain reading is that the neither bucket holds more of those multi-service products. Docker-readiness barely differs: 195 of 202 permissive, 200 of 202 copyleft and 51 of 53 neither have an official image or Compose file.

What these strings are

The license is the string we record for each app, kept as the project states it: an SPDX-style identifier where one fits, and the project's own name for the license where it does not. It describes the self-hostable edition we list, so an open-core product shows the core's license and the note about enterprise modules. The buckets above are ours: "copyleft", "permissive", "open core" and "source-available" are our shorthand for the groups, not fields in the data, and we have not reviewed each license's terms or had the strings checked by a lawyer. Our catalog is a curated selection rather than a census of self-hosted software, and a license can change between our check and your install, so read the project's LICENSE file before relying on any row here.

About the numbers

This post uses catalog.json from our open dataset, exported on 2026-11-13 and published as release v1.11.0 (457 records). The idle-RAM medians use the 418 measured figures in install-verification.json from the same export (measured between 2026-08-28 and 2026-11-12). The dataset is CC BY 4.0 and cited as DOI 10.5281/zenodo.22983606.

Apps in this post

Related

Data
378 of 418measured apps idle inside a 2 GB server's 1,126 MB app budget (2026-11-13 export)

A 2 GB VPS has the idle RAM for 378 of the 418 self-hosted apps we measured; most AI agent platforms, no-code databases and BI tools do not fit

Applying our 1,126 MB app budget for a 2 GB server to 418 measured idle figures, category by category: 40 of the 65 categories with four or more measured apps fit entirely, while four of six AI agent platforms and three of five no-code databases and BI tools do not, and the 2 GB plans in our catalog listed at $6 to $18 a month on 2026-09-27.

Nov 16, 2026 · 11 min read · data as of Nov 2026n=418
Data
23 MBmedian idle RAM of the 26 measured apps tagged Rust (2026-11-13 export)

Self-hosted apps tagged Rust or Go idle at medians of 23 and 28 MB; within almost every language, each extra data store raises idle RAM

Idle RAM for 418 self-hosted apps grouped by the language, framework and data-store tags in our catalog: Rust-tagged apps idle at a median 23 MB, Go-tagged at 28 MB, Java-tagged at 817 MB, and apps with no database tag at 104 MB against 7,970.5 MB for the four with three data-store tags.

Nov 9, 2026 · 8 min read · data as of Nov 2026n=418
Data
1,310 MBmedian Docker disk footprint right after install, of 418 apps (2026-11-13 export)

The median self-hosted app takes 1.3 GB of disk right after install; 36 apps at 5 GB or more hold 37% of the total

Docker disk footprint for 418 self-hosted apps, each installed on a server pruned of earlier apps: median 1,310 MB, from 16 MB (rustypaste) to 27,324 MB (PostHog). Disk tracks idle RAM at the rank level (rho 0.82), but 32 apps that idle under 256 MB still take more than 2 GB of disk.

Nov 5, 2026 · 11 min read · data as of Nov 2026n=418

One email when we publish new measurements.

No digests, no roundups. Just the new numbers and what changed.

unsubscribing deletes your address

Analytics cookies only (Google Analytics, PostHog). No ad networks, no cross-site tracking. See our privacy policy.