Best GitHub Repos for Claude Code in 2026: Skills, Memory, Context and Cost

The GitHub repositories Claude Code users actually install in 2026, grouped by job: skills libraries, memory, code context, token tracking and curated indexes — with live star counts.

Short version: pick by job, not by stars. You need at most one tool each for skills, memory, code context and cost tracking, plus a curated index to browse. Star counts below are from the GitHub API on 18 September 2026. Popularity is not a security review, so read install scripts before running them.

Most-starred GitHub repositories used with Claude Code, September 2026

Skills and behaviour

**anthropics/skills** (about 177,000 stars): Anthropic's official Agent Skills repository. The reference for how skills are written.
**multica-ai/andrej-karpathy-skills** (about 214,000): a single CLAUDE.md distilled from Andrej Karpathy's coding guidance.
**affaan-m/ECC** (about 262,000): describes itself as an agent harness performance system of skills and instincts. Large; read what it installs before adopting it wholesale.

Memory across sessions

**thedotmack/claude-mem** (about 94,000): captures what the agent learned so the next session does not start cold.
**OthmanAdi/planning-with-files** (about 27,000): file-based plans that survive context resets on long tasks.

Code context

**colbymchenry/codegraph** (about 71,000): a pre-indexed code graph served over MCP, so the agent queries structure instead of grepping.
**Graphify-Labs/graphify** (about 119,000): turns a codebase with its docs, schemas and configs into a navigable knowledge graph.
**yamadashy/repomix** (about 28,000): packs a repository into one file for pasting into a chat.

Cost and tokens

**JuliusBrussee/caveman** (about 106,000): a skill and proxy built on saying less to spend fewer tokens.
**getagentseal/codeburn** (about 11,000): a free local tracker of token usage and cost across coding tools.

Which Claude Code repo for which job

Indexes to browse

**hesreallyhim/awesome-claude-code** (about 54,000): commands, hooks, skills and workflows.
**wshobson/agents** (about 40,000): a plugin marketplace of specialised agents for Claude Code and other harnesses.

How to add them without bloat

Name the problem first.

- Lost context, large codebase, or cost: pick the matching section.

Install one tool.

- Read its README and install script before running it.

Measure.

- Compare /context before and after.

Remove what does not earn its place.

- Fixed context you are not using is a cost on every request.

For the skills side specifically, see our list of the best Claude Code skills.

Frequently asked questions

What should I install first for Claude Code?+

Start with a short CLAUDE.md for your project, then Anthropic's official anthropics/skills repository to learn how skills are structured, then one tool for the problem you actually have. If sessions lose context between days, a memory tool such as claude-mem. If your codebase is large, a code-context tool such as CodeGraph. If costs are the worry, a tracker such as codeburn first.

Do these repos work with Codex and Cursor too?+

Many do. Skills written in the open SKILL.md format work across Claude Code, Codex, Gemini CLI, Cursor and others. MCP-based tools, such as CodeGraph, work with any MCP client and often auto-configure several agents during install. Claude Code-specific pieces, such as hooks and plugins, are the part that does not transfer directly.

Are highly starred Claude Code repos safe?+

Stars measure popularity, not safety. Several of these projects install scripts, hooks or background services that run on your machine with your permissions. Read the README and the install script before running it, prefer repositories with visible maintainers and recent commits, and try new tools in a throwaway project before adding them to a repository with production credentials.

How do I avoid bloating Claude Code with too many tools?+

Add one tool per real problem and remove what you stop using. Every MCP server adds tool definitions to every request, every always-loaded instruction file adds tokens, and overlapping tools confuse the model about which to use. Check the effect with /context after adding something; if the fixed context jumps and the benefit is unclear, take it out.

Where can I find more Claude Code resources?+

The curated indexes are the best starting point: hesreallyhim/awesome-claude-code collects commands, hooks, skills and workflows, and ComposioHQ/awesome-claude-skills and VoltAgent/awesome-agent-skills collect skills. The StackPicks Agent Skills directory lists the most-installed skills with a note on what each does, links each one to its source, and marks recent additions as new.