How the 1,000 most-starred open-source repositories are built (2026)

Key findings

• AI agents are now drawn as users. 33.3% of the repositories created in 2025 or later show an AI agent or coding assistant as one of the outside actors that use the software, next to people. Among repositories created before 2020 it is 2%.
• LLMs moved into the architecture. A component for an LLM or AI model provider appears in 41.8% of repositories created 2023 to 2024 and 2.4% of those created before 2020. MCP shows up in 31.8% of the 2025-or-later repositories.
• The command line beats the browser. 29.6% of all diagrams include a command-line interface and 14.3% a web interface. In Rust it is 56.2% CLI, in Go 47.7%.
• Half keep state, few use queues. 52.5% draw a data store (a database, cache, file or object storage). Only 7.4% draw a queue or event bus.
• Monorepos are a TypeScript and Rust habit. 17.1% of diagrams link files in workspace folders (packages/, apps/, crates/, libs/, services/). For TypeScript it is 54.3%, for Rust 43.8%, for Python 6.4% and for Go 5.8%, where 54.7% use the cmd/ and internal/ layout instead.
• Plugin systems are common. 22.9% name a plugin or extension system, 34.4% of JavaScript repositories.

AI shows up in the architecture

Grouped by the year GitHub says each repository was created, the newer the project, the more often its diagram has a component for an LLM or model provider, speaks the Model Context Protocol, or draws an AI agent among the actors that use it. Projects from 2023 and 2024 lead on LLM components; projects from 2025 on are the first where agents are a common kind of user. Examples: morluto/rea, mattpocock/skills, obra/superpowers and multica-ai/andrej-karpathy-skills.

What the top repositories are made of

Across all 1,000: 52.5% draw a data store and 7.4% a queue or event bus; 87.6% draw at least one outside actor, most often a “user” or a “developer”. Repositories with a job or event queue include Significant-Gravitas/AutoGPT, langgenius/dify, langflow-ai/langflow and google-gemini/gemini-cli.

By language

Each language with at least 40 repositories in the sample. Highest share in each column: CLI Rust (56.2%); Web UI TypeScript (20.8%); Desktop Rust (26%); Plugins JavaScript (34.4%); Data store Go (75.6%); Workspace TypeScript (54.3%); LLM Python (29.5%).

How repositories are laid out

17.1% of diagrams link components to files in workspace folders, the sign of a monorepo split into packages. packages/ is by far the most common, then apps/ and crates/. Examples: freeCodeCamp/freeCodeCamp, deepseek-ai/deepseek-harness, react/react and anomalyco/opencode. In Go, 54.7% of diagrams link into cmd/ or internal/; in Rust, 37% link into crates/.

The files diagrams point to most, by language

The number of repositories whose diagram links a component to a file with that name (READMEs left out).

Python

__init__.py 53, main.py 48, app.py 32, server.py 32, base.py 31

TypeScript

index.ts 68, server.ts 30, main.ts 20, index.tsx 18, storage.ts 18

JavaScript

index.js 27, main.js 9, server.js 6, app.js 5, package.json 5

Go

main.go 42, server.go 35, manager.go 24, client.go 21, engine.go 16

Rust

mod.rs 38, main.rs 37, lib.rs 34, server.rs 17, config.rs 13

C++

main.cpp 11, CMakeLists.txt 4, server.cpp 4, __init__.py 3, env.h 3

The ten most-starred repositories

1. codecrafters-io/build-your-own-x 561,192 stars · Markdown · 34 components
   Guide catalog · Systems topics · Software topics · Graphics topics · AI topics
2. sindresorhus/awesome 534,108 stars · no primary language · 7 components
   Resource catalog · Contributor guidance · List definition
3. public-apis/public-apis 497,404 stars · Python · 11 components
   API catalog · Catalog validation
4. freeCodeCamp/freeCodeCamp 458,787 stars · TypeScript · 22 components
   Learning experience · Curriculum content · Platform services · API and persistence
5. morluto/rea 445,142 stars · TypeScript · 23 components
   Interfaces and orchestration · Analysis domains · Analysis providers · Runtime and evidence
6. EbookFoundation/free-programming-books 401,341 stars · Python · 10 components
   Catalog entry points · Learning resources · Supplementary resources · Reader access
7. openclaw/openclaw 394,231 stars · TypeScript · 24 components
   User clients · Gateway and sessions · Agent runtime · Channels and plugins · Configuration and memory
8. donnemartin/system-design-primer 377,390 stars · Python · 18 components
   Learning resources · Interview practice · System design examples · Object-oriented examples
9. nilbuild/developer-roadmap 371,595 stars · TypeScript · 15 components
   Learning Content · Content Synchronization · Sync Support
10. jwasham/coding-interview-university 364,128 stars · no primary language · 17 components
    Canonical guide · Localization · Reference assets · Automation · External learning

Six of the ten are lists, books or courses, which GitDiagram diagrams as the structure of their content. They stay in the sample because they are what the most-starred repositories are.

Method

We read GitDiagram’s index of stored public diagrams on 2026-11-09 and took the 1,600 entries with the most stars. Of those, 532 had a diagram from before GitDiagram stored a component graph (made between December 2024 and March 2026), 41 were renamed repositories stored under two names, 1 was a fork and 1 no longer exists on GitHub. Stars, primary language and creation year come from the GitHub API on 2026-11-09; the sample is the 1,000 most-starred of the rest. Only public repositories and public diagrams were read. The median diagram has 21 components, 21 connections and 4 groups; 87.9% of components link to a real file or folder.

A trait counts when this regular expression matches a component or group name, lowercased:

• Has an LLM or AI-model component: /\b(llms?|language models?|model providers?|ai providers?|llm providers?|openai|anthropic|ollama|gemini|inference (api|providers?|backends?))\b/
• Speaks MCP: /\bmcp\b|model context protocol/
• Has a command-line interface: /\bcli\b|command[- ]line|\bterminal (ui|interface|app)\b|\btui\b/
• Has a web interface: /\bweb (ui|app|application|interface|client|frontend|dashboard|console)\b|\bfrontend\b|\bdashboard\b/
• Has a desktop app: /\bdesktop\b|\belectron\b|\btauri\b/
• Has a plugin or extension system: /\bplugins?\b|\bextensions?\b|\badd-?ons?\b/
• AI agent as a user: an outside actor matching /\bagents?\b|\bmcp clients?\b|\bassistants?\b|\bcopilot\b|\bclaude code\b|\bcodex\b/

The scripts that fetched and counted the data are in GitDiagram’s repository. Anyone can check a single row by opening that repository’s diagram.

Limits

• The sample is the most-starred repositories that someone has diagrammed on GitDiagram, not every top repository on GitHub. It starts at 561,192 stars and ends at 25,275.
• Diagrams are written by an AI model from the file tree, the README and up to 12 source files, then checked against the real tree. They show the main parts, not every file, and can miss or misname a part.
• Diagram size is set by GitDiagram, not by the repository: a diagram has at most 34 components, so the median of 21 says more about the generator than about the code.
• Traits are found by matching words in component and group names (the rules are below and in the dataset). A component named differently is missed; a name that matches by accident is counted.
• Lists, books and courses are included, because they are among the most-starred repositories (56 have no primary language).
• Diagrams were made between 2026-03-28 and 2026-11-09; a repository may have changed since its diagram was made.

Download and cite

One row per repository, with its rank, stars, language, creation year, counts and traits: CSV or JSON (with the rules). Also as Markdown. Free to reuse under CC BY 4.0. Cite as: GitDiagram, “How the 1,000 most-starred open-source repositories are built (2026)”, 2026-11-09, https://gitdiagram.com/open-source-architecture-2026.

To see any repository’s own diagram, replace “hub” with “diagram” in its GitHub URL, or read how to visualize a codebase.