Anthropic's Claude Code crosses 127K stars: the Python rewrite hiding in plain sight

The headline number on Claude Code is 127K stars, but the more interesting number is 79.7%. That's how much of the repo is now Python. A few months ago, the same repo was Shell-dominant, with Python and TypeScript splitting the rest.

Nobody at Anthropic announced a rewrite. The language stats just shifted, commit by commit, until Python ran the show. That kind of quiet architectural move usually says more about a project's direction than any release post.

What Happened

anthropics/claude-code hit 127K stars and 20.8K forks with 54 contributors. The most recent CHANGELOG and feed.xml updates landed 12 hours before I'm writing this, which is the typical cadence. The repo went GA with v1.0.0 in May 2025, so most of that traction came in roughly 12 months.

What I'd flag from the repo state:

• The codebase is 79.7% Python, 13.7% Shell, 5.0% TypeScript, with PowerShell and Dockerfile making up the rest. That's a Python-first project now. Anyone evaluating Claude Code as a "Node tool" is working from outdated information.
• 5K+ open issues and 544 open PRs. The user base has scaled past what the maintainer team can triage in real time.
• Recent CI work migrated Claude authentication in workflows to workload identity federation, replacing static API keys (#61584). That's the kind of internal hygiene most projects don't ship publicly. The fact that it's visible suggests that Anthropic treats the repo as a reference implementation.
• The examples/ directory now ships MDM deployment templates. Translation: enterprise IT teams are deploying Claude Code at scale, and Anthropic is meeting them where they are.
• Commits co-authored by @claude show up regularly. Anthropic is dogfooding the tool on its own codebase, which is the most credible signal you can ship.

Key Features

• Terminal-native agent: runs in your shell, reads the codebase, and responds to natural language for edits, explanations, and git operations. The bet is that most developers want AI assistance in the terminal, not the IDE.
• GitHub @claude mentions: tag @claude on an issue or PR, and the agent participates directly. This is the integration most teams underestimate, because it pulls AI review into the place where collaboration already happens.
• Plugin system: the plugins/ directory ships with extensible commands and agents, including a code-review plugin that posts inline PR comments and a learning-output-style plugin for interactive decision-making.
• Devcontainer support: ships with a .devcontainer config that pre-loads the VS Code extension and configures marketplace firewall rules. Useful for teams standardizing remote dev environments.
• MDM deployment templates: enterprise IT teams get example configs for managed device deployment. A small detail that signals where the customer base is moving.
• Native installers across platforms: curl and PowerShell scripts, Homebrew, WinGet. The npm path is still listed but explicitly deprecated.

Why It Matters

A few patterns I'm seeing more broadly that line up with this:

• The IDE is losing ground as the default surface. Cursor and Copilot are still huge, but the developers spending the most time on AI workflows are increasingly in the terminal. Claude Code's traction is one data point in that shift.
• The Python pivot says something about the audience. Shell scripts work for power users. Python works for data engineers, ML teams, and the long tail of developers who already have Python toolchains. Choosing Python as the dominant implementation language is a bet on which audience matters next.
• GitHub is becoming an agent runtime. Between @claude mentions, MDM deployment templates, and workload identity federation, the integration surface has moved well past "AI app you open." The agent now lives inside CI, code review, and managed device fleets.

The collapse of "ask the AI" and "apply the change" into a single step is where the productivity gain actually lies. Everything else is interface polish.

Example Use Case

A backend team maintains a Django monorepo and needs to refactor a 400-line view into smaller service modules. A developer opens their terminal in the project root, runs claude, and types: "Split views/orders.py into separate service files for validation, pricing, and fulfillment. Keep the existing tests passing."

Claude Code reads the file, proposes a split with updated imports, creates the new modules, and runs the existing test suite to verify nothing breaks. The developer reviews the diff, then asks Claude Code to "commit this with a message explaining the refactor," and the tool handles the git workflow.

Total context switches: zero. This is the workflow I'd demo to anyone still copy-pasting between a chat window and their editor.

Competitive Context

A few things stand out when you put Claude Code next to the obvious alternatives:

• Scope is narrower than Cursor or Copilot. The README is explicit about Claude Code being a terminal-based agent for coding workflows. Cursor wants to own the IDE; Claude Code wants to own the shell. Those are different bets.
• 20.8K forks is the number to watch. Stars are easy to accumulate. Forks mean people are building on the codebase, customizing plugins, or adapting commands for their team. Forks separate popular repos from foundational ones.
• The plugin architecture creates a moat Copilot doesn't have. Once developers invest in writing custom commands and agents per repo, they lock in. Copilot doesn't give teams that kind of extensibility surface.

The competitive question isn't whether Claude Code is better than Cursor. The question is whether developers want AI in the IDE or AI in the terminal. The fork count and contributor activity suggest a real audience is voting for the terminal.

My Take

The Python language stat is the detail worth sitting with. Repos don't shift from Shell-dominant to 79.7% Python without an architectural decision behind the move. Either Anthropic rewrote the core, brought in Python-heavy contributors, or both. None of those scenarios show up in the changelog, but all of them show up in the language bar.

The MDM deployment templates and workload identity federation point to enterprise deployment as a serious priority. Anthropic isn't just chasing individual developer seats. They're building for whatever your security team approves for managed device fleets. That changes who Claude Code competes with.

The open question is whether the plugin ecosystem develops outside Anthropic. If teams ship their own plugins for internal tooling and workflows, Claude Code becomes a platform. If it stays official-plugins-only, it will remain a great CLI. Those are different futures, and the next six months should tell us which one is happening.