State of Claude Code 2026: What 2,500 Public Repos Reveal

We analyzed 2,500 public repositories that use Claude Code. 85% ship a CLAUDE.md file, but only 25% define a single custom subagent. Most developers use Claude Code as a smarter assistant that follows written instructions. Far fewer have crossed into the agent system it was built for: subagents, skills, hooks, and MCP. That gap is the most interesting thing in the data.

Here is what 2,500 real .claude/ setups look like in 2026, and what the numbers say about how people actually work with Claude Code.

The headline: a CLAUDE.md, and usually nothing more

Of the public repositories we sampled, 84.9% had a CLAUDE.md file and 62.1% had a .claude/ directory of some kind. But the deeper features drop off fast. Only 24.6% defined a custom subagent, and just 13.3% used hooks.

A CLAUDE.md is a plain-text instruction file. It tells Claude Code your conventions, your stack, and your house rules. It is the easy first step, and almost everyone takes it. The agent system is the harder, more powerful step, and most people have not taken it yet.

That is the core finding: the Claude Code ecosystem in 2026 is wide and shallow. Lots of people have it installed and pointed at their repo. A minority have built a system on top of it.

The adoption ladder

Feature adoption falls off in a clear order. Here is the share of all 2,500 sampled repositories using each capability.

The drop from "has a CLAUDE.md" (85%) to "uses any agent feature" (54%) to "defines a subagent" (25%) is the shape of the whole ecosystem. Each rung up the ladder loses roughly a third of the field.

If you only look at the repos that bothered to create a .claude/ directory, the picture is more committed: 45% use skills, 41% use slash commands, 40% define subagents, and 21% use hooks. The people who set up structured config are far more likely to go all the way. The split is between developers who treat Claude Code as autocomplete and developers who treat it as a platform.

CLAUDE.md files are getting big

A CLAUDE.md is not a one-liner anymore. The median file in our sample was 6.2 KB, which is roughly 100 to 150 lines of instructions. A third of them (33.8%) were over 10 KB.

The long tail is striking. The top 10% of CLAUDE.md files exceeded 24 KB, the top 1% passed 62 KB, and the single largest we found was 341 KB. That is one instruction file the length of a short book, which says something about how much context a serious Claude Code user ends up encoding.

This is the quiet trend underneath the headline. People who stick with Claude Code keep feeding the CLAUDE.md until it becomes the institutional memory of the project. For more on doing that well, see CLAUDE.md mastery and what Claude Skills are.

The power users go deep

Among the 614 repositories that defined at least one subagent, the median was 6 agents. The top 10% ran 25 or more. These are not toy setups. They are teams of specialized agents, each scoped to a job.

Skills go even further. Among repositories using skills, the median was 8 and the long tail ran into the hundreds. A small group of builders is clearly treating Claude Code as the runtime for an entire automated workflow, not a coding helper.

This matches what we see in practice. Once you cross into subagents, you stop asking "can it write this function" and start asking "can it own this whole part of the build." If you want to see how that works, read what subagents are and the best ones in 2026 and the agent fundamentals guide.

Who is actually using Claude Code

The ecosystem skews toward web and AI work. TypeScript leads at 30.2% of sampled repos, Python follows at 22.7%, and together they are more than half the field.

It is not just hobby projects, either. 78.2% of the repositories were pushed to in the last 90 days, so this is a live ecosystem, not a graveyard of abandoned experiments. Most repos are small and personal (the median star count was 0), but the sample also includes serious names shipping .claude/ config: PostHog, NVIDIA's TensorRT-LLM, Automattic's Calypso, AutoGPT, and PrefectHQ's FastMCP among them. Real teams are committing Claude Code setups to their main branches.

What the gap means

Read the data one way and it is a story about adoption. Read it another way and it is a story about opportunity.

85% of these projects have written down instructions for Claude Code. Only 25% have built an agent on top of it. The plumbing that turns Claude Code from an assistant into a build system, the subagents that own a layer, the skills that encode a workflow, the hooks that enforce quality, the production stack underneath, is exactly where most people stop. Not because it is not worth it. Because wiring it from scratch is work.

That is the gap the Build This Now Code Kit is built for. It is a $29 one-time harness that drops into a Next.js and Supabase project and turns Claude Code into a full build team out of the box: planning agents, a build pipeline, adversarial evaluators, quality gates, and auth, payments, and a database already wired in. Instead of being one of the 75% who stop at a CLAUDE.md, you start at the far end of the adoption ladder. No subscription, at buildthisnow.com.

If you would rather build the harness yourself, that is genuinely doable, and the data shows a real community doing it. Start with what Claude Code is, then work through subagents, skills, and hooks.

Methodology

We used the GitHub code search API to find public repositories containing a CLAUDE.md file or files under a .claude/ directory (agents, skills, commands, settings.json, and project .mcp.json). To get past the per-query result cap, we partitioned the CLAUDE.md search by file size and combined it with the .claude/ path searches. After deduplication, that returned 8,298 unique repositories.

We then classified a sample of 2,500 of those repositories. For each one, we fetched repository metadata (language, stars, last push) and the full recursive file tree, then counted subagents (.claude/agents/*.md), skills (.claude/skills/*/SKILL.md), slash commands (.claude/commands/*.md), hooks (files under .claude/hooks/), MCP config (.mcp.json), and CLAUDE.md size in bytes.

Limitations

The data reflects public, GitHub-indexed repositories as of June 2026. It misses private repositories, which is where a lot of serious commercial Claude Code work lives, so true adoption of the deeper features is almost certainly higher than these public numbers. We classified the file tree of the default branch only, and 10 very large repositories had truncated trees that undercount their files. MCP adoption is measured by the presence of a project .mcp.json file, so MCP configured inside settings.json is not counted, which makes 17.0% a floor rather than a ceiling. Star counts are a rough proxy for seriousness, not a precise one. We report what the public corpus shows and nothing we did not measure.

FAQ

How many developers actually use Claude Code subagents? In our sample of 2,500 public repositories, 24.6% defined at least one custom subagent. Among repos that had any .claude/ config directory, that rose to 39.5%. Roughly one in four Claude Code projects has crossed from plain instructions into the agent system.

What percentage of Claude Code projects have a CLAUDE.md file? 85% of the public Claude Code repositories we analyzed had a CLAUDE.md, making it by far the most common configuration. The median was 6.2 KB and 34% were over 10 KB.

How big is a typical CLAUDE.md file? The median was 6.2 KB, roughly 100 to 150 lines. The top 1% exceeded 62 KB and the largest single file we found was 341 KB.

Which languages dominate the Claude Code ecosystem? TypeScript (30%) and Python (23%) together are more than half of all sampled repositories, followed by JavaScript, HTML, Shell, and Rust.

How was this study conducted? We found 8,298 public repositories with Claude Code config via the GitHub code search API, then classified a 2,500-repo sample by fetching each file tree and counting agent features and CLAUDE.md size. See the full methodology above.