Open-source AI chat apps in 2026: 51 mapped, 12 compared, and which fits what

October 19, 2026

If you’d like the ChatGPT or Claude experience on your own terms, with your own keys, your own models or your own server, there’s an open-source app for it. There are a lot of them, though, and they aren’t interchangeable. On 18 October 2026 I listed 51 that have around 500 GitHub stars or more, then read the documentation of twelve against the uses people usually have in mind. Both spreadsheets are yours to download: the 51 apps, with their stars and dates, and 64 selection criteria with a blank scoring sheet.

In brief

1. Open WebUI leads on stars and on features. It holds 156,000 of the 971,500 stars across the 51 apps, and of the twelve I read, it ticks the most of the 27 features I checked.
2. A quarter of the list has gone quiet. Thirteen of the 51 had no push in the last 12 months, GPT4All and its 77,400 stars among them, and seven of the thirteen were created in March 2023.
3. The basics are everywhere now. All twelve show a model’s reasoning, and ten can use MCP servers and chat with your files. The differences are in team features, offline use, licences and telemetry.
4. Among servers you host, Open WebUI fits most uses; among apps you install, DeepChat does. Jan and PocketPal AI are the picks for private, offline use, and LibreChat ties Open WebUI for teams.
5. Read the licence and the telemetry setting before you settle. Three of the twelve add terms to a standard licence, and six collect usage data until you switch it off.

Most popular apps are still active, and the quiet ones are mostly from 2023

Of the 17 apps with 10,000 stars or more, 15 had a push in the last 12 months. The 13 that went quiet are mostly the ChatGPT front ends of 2023: seven of the twelve apps created in March 2023 are stale.

Each dot is an app. Up is more GitHub stars, on a log scale from 500 stars to 200,000 stars; right is longer since the last push, from pushed this month to four years without a push, stretched near the left so this month’s pushes don’t pile up. The middle lines sit at 10,000 stars and at the workbook’s 12-month threshold for stale, which splits the apps into popular and active, popular but quiet, smaller and active, and smaller and quiet. The named dots are the twelve I compare below and the four quiet apps that still have 5,000 stars or more.

Desktop is the most common home

Thirty-three of the 51 apps have a desktop build, 21 a web or self-hosted version and 16 a mobile app. Most stay on one kind of platform; only Khoj and AIaW cover all three.

How I compared twelve of them

I compared twelve apps on the factors behind eight everyday uses, all taken from the criteria workbook. The twelve are the ten its scoring sheet names, plus PocketPal AI and Google AI Edge Gallery, which run models on the phone.

Every weight stays at the workbook’s default of 3, so each factor counts the same: a Yes is one point and a Part half a point.

The twelve, side by side

The twelve split into servers you host in a browser and apps you install, and the split shapes almost everything else.

The four apps that run as a server are the ones built for teams, and the private, offline end belongs to Jan and PocketPal AI. Across runs from for one person to for a team, by how many of the four team features an app has; up runs from cloud first to private and offline, by how many of the four privacy ones. Part counts half, and both are counted off the tables below. That gives four groups: personal and private, team and private, personal and cloud first, and team and cloud first, where LobeHub sits alone.

Most of the twelve now share the same core: every one shows a model’s reasoning, and the two apps that run models on the phone are the only ones without full MCP support or documented chat with files. Single sign-on and sending one prompt to several models side by side are the rarest.

Feature by feature

Each cell is Yes, Part, – for no, or ? where the project’s documentation doesn’t say. Telemetry is Off, Opt-in, or On by default with a switch to turn it off.

Chat, coding and knowledge

Tools and agents

Models and privacy

Teams and everyday use

Best for…

For most uses there’s a clear pick among the servers you host and another among the apps you install. Ties go to the app with more of the 27 features overall.

Method and caveats

• The 51 apps and their figures come from the apps workbook: open-source desktop, mobile and self-hosted web apps that can replace the ChatGPT and Claude apps, with about 500 GitHub stars or more and their source on GitHub. Archived repositories, wrappers around the ChatGPT website, agent frameworks, inference runtimes and closed-source apps are left out, and its Info sheet names them. Stars are rounded, read from GitHub and the OSS Insight API on 18 October 2026; creation and last push are months, and an app is stale when its last push is more than 12 months old.
• The features of the twelve come from each project’s own README, documentation and release notes, and from its source code where the documentation was silent, all read on 18 October 2026. I didn’t install or run any of them. Documentation lags the code: DeepChat’s README still lists artifacts, which were removed in September 2026.
• Chatbox’s repository is its Community Edition, kept in step with a private Pro edition. The tables count what the repository documents, not the paid hosted models, search and cloud MCP servers, and nothing says the official builds come from the same source.
• Jan, Chatbox and Witsy have no user accounts, so I counted single sign-on and admin controls as no for them.
• The criteria workbook names a licence addendum for Cherry Studio. It was removed in September 2025, and its licence today is the plain AGPL-3.0.
• The picks are counts, not verdicts. Your weights will differ from the default, and the criteria workbook’s scoring sheet is there for that. Stars measure attention, not quality.