10 Open Source Claude Code Tools You Need to Try (May 2026)

Hundreds of new open source AI projects hit GitHub every day. We dug through the noise and found the 10 that actually matter for Claude Code users right now — all released or significantly updated in May 2026.

These aren't Claude Code replacements. They're tools that make it better: a skill that forces concise answers (and might actually improve accuracy), a knowledge graph builder that cuts token usage by 71x, a cost tracker that shows exactly where your money goes, and seven more. We tested each one.

Here's what's worth installing and what you can skip.

1. Caveman — Talk Less, Code Better

What it does: Caveman is a skill that forces Claude Code to give terse, direct responses instead of the usual verbose paragraphs. The philosophy is simple: why say many word when few do trick?

It ships with three intensity levels — Caveman Light (slightly more concise), Full (noticeably terse), and Ultra (maximum brevity). We run Light daily and haven't looked back.

Why it matters: The real story isn't just token savings (~5% overall). A March 2026 paper, Brevity Constraints Reverse Performance Hierarchies in Language Models, found that forcing powerful models to be concise improved accuracy by 26 percentage points on reasoning tasks. The models stop talking themselves into wrong answers.

50,000+ stars in its first month. Installation is one command. If you install nothing else from this list, install Caveman.

GitHub: JuliusBrussee/caveman — 55.4K ⭐ | Python | MIT

2. Graphify — Knowledge Graphs for Your Codebase

What it does: Graphify reads your project files and builds a structured knowledge graph — an HTML visualization, markdown report, and JSON structure that Claude Code can query instead of re-reading raw files every time.

The headline number: 71.5x fewer tokens per query compared to feeding Claude Code raw files. That's the difference between a $50/month habit and a $50/year habit on API billing.

How it compares to Obsidian: If you're already using Obsidian for Claude Code memory, Graphify sits one level deeper. Obsidian is a nice interface for markdown files. Graphify is closer to a true graph RAG system like LightRAG — but without embeddings. It's multimodal too: PDFs, screenshots, diagrams, and video transcripts via Whisper.

If you're someone who loves Obsidian but wants more structured memory under the hood, Graphify layers nicely on top.

GitHub: safishamsi/graphify — 43.9K ⭐ | Python | MIT

3. Claude Video — Give Claude Eyes

What it does: Claude Sonnet and Opus can't natively ingest video. Claude Video works around this with a clever frame-budget system: it extracts key frames via FFmpeg (30 frames for a 30-second video, 100 frames for 10+ minutes) and audio via Whisper, then feeds both to Claude Code.

It supports YouTube, TikTok, Vimeo, and local files. The /watch command is all you need.

The honest take: At 829 stars, this is the youngest tool on the list. For longer videos (3+ minutes), the sparse frame sampling means you'll miss visual details. But for quick explainers, code walkthroughs, and bug report videos, it's a genuine alternative to routing everything through Gemini's API or NotebookLM. Anything that reduces our dependency on Google for video processing is worth watching.

GitHub: bradautomates/claude-video — 829 ⭐ | Python | MIT

4. Open Design — Free Claude Design Clone

What it does: Open Design is an open source clone of Anthropic's Claude Design. It replicates the prototyping, slide deck creation, and design workflow — and adds API calls for image and video generation on top. You can run it completely locally, for free, with any coding agent.

It bundles 31 composable skills and 72 design systems, built on top of four other open source projects: Huashu Design, the Guzheng PowerPoint skill, Open Code Design, and Multika. The result is a surprisingly complete design toolkit.

If you've ever hit Claude Design's weekly usage limits and wished you could keep going, this is your answer. We wrote about it in our portable Claude Code guide — it works great in offline setups too.

GitHub: nexu-io/open-design — 30.5K ⭐ | TypeScript | Apache 2.0

5. CodeBurn — Track Every Token Dollar

What it does: CodeBurn tracks token usage, costs, and performance across 16 AI coding tools — Claude Code, Codex, Cursor, and 13 others. The interactive TUI dashboard breaks down your spending by activity, project, model, shell commands, and MCP servers.

The killer feature: it shows actual dollar amounts, not just token counts. And it doesn't just tell you where money is going — it gives specific recommendations on how to reduce costs. If you've been wondering why your Claude Code bills keep climbing, CodeBurn is the diagnostic tool you need.

It also has a native macOS menubar app and supports 162 currencies. Small detail, but appreciated if you're not billing in USD.

GitHub: getagentseal/codeburn — 5.6K ⭐ | TypeScript | MIT

6. Impeccable 3.0 — Fix Your Front-End

What it does: Impeccable is a single Claude Code skill containing 23 commands dedicated to making your web pages not look like AI-generated slop. It ships with 7 domain reference files covering typography, color, motion, spatial design, interaction, responsiveness, and UX writing.

The v3.0 update (released last week) added live browser editing — you can bring up your webpage, click on individual components, and cycle through design variations right in the browser. The companion website shows before-and-after examples for every command, which is genuinely useful for understanding what each command actually does before running it.

We covered the best Claude Code skills recently — Impeccable 3.0 would have made that list easily if it had shipped a week earlier.

GitHub: pbakaus/impeccable — 25.5K ⭐ | JavaScript | Apache 2.0

7. Design Extract — Steal Any Website's Design System

What it does: If you've used awesome-design-md (70K stars, growing fast), you know the frustration of finding a website you love that isn't in the catalog. Design Extract solves this — point it at any website and it extracts the complete design system.

Using a headless browser, it captures layout systems, responsiveness patterns, interaction states, motion language, component anatomy, and brand voice. Output formats include DTCG tokens, Tailwind config, and Figma variables. It even runs accessibility audits.

This is more than just taking screenshots and saying "copy this." The headless browser approach means it captures computed styles, transitions, hover states, and responsive breakpoints that a simple screenshot would miss.

GitHub: Manavarya09/design-extract — 2.3K ⭐ | JavaScript | MIT

8. Career Ops — AI-Powered Job Search

What it does: Career Ops turns Claude Code (or any AI coding CLI) into a full job search command center. It scans 45+ company portals, evaluates job fit with A-F scoring, generates ATS-optimized PDFs tailored to each listing, and tracks everything in a Go-powered dashboard.

This isn't a LinkedIn mass-apply bot. It's a scalpel. You paste in a job URL or description, it classifies the role, evaluates whether you're actually a match based on your CV, generates a tailored resume, and updates your tracker. The creator reports 740+ job evaluations and 100+ personalized CVs generated.

Under the hood it uses Playwright to navigate job portals. If you know someone in the job market right now, send them this link. 43K stars in a few weeks tells you the developer community agrees.

GitHub: santifer/career-ops — 43.1K ⭐ | JavaScript/Go | MIT

9. Browser Harness — Self-Healing Browser Agent

What it does: Think Playwright, but it improves after every run. Browser Harness connects LLMs directly to Chrome via CDP (Chrome DevTools Protocol). Every time it completes a task — say, navigating Amazon or filling out a form on GitHub — it updates its own agent skill file with what worked and what didn't.

It's essentially a self-healing automation loop. Failed selectors get noted. Successful navigation paths get cached. Domain-specific skills for LinkedIn, Amazon, GitHub, and other popular sites get auto-generated as you use it. From the browser-use team (the same folks behind the popular browser-use project).

At 11K stars and only two weeks old, this one is early but shows where browser automation is heading. The self-improving pattern — where agents write notes to themselves about what works — is going to be the standard approach within months. We've seen similar patterns with Hermes Agent's persistent memory.

GitHub: browser-use/browser-harness — 11.1K ⭐ | Python | MIT

10. n8n MCP Server — Automation Meets Claude Code

What it does: n8n's official MCP server lets Claude Code build, validate, and deploy automations directly into your n8n instance. The key difference from previous community MCP servers: this one generates automations in TypeScript first, validates that the nodes and connections actually work, then converts to JSON for n8n.

That TypeScript validation step is critical. Previous MCP servers generated JSON directly, which meant Claude Code would build automations that looked correct but failed when deployed because of invalid node configurations or broken connections. The TypeScript layer catches those errors before they reach your instance.

If you use n8n for automation and Claude Code for development, this bridges the two. The "death of n8n" has been exaggerated — it's finding its niche as a tool that AI coding agents can drive. If you're interested in the broader automation landscape, check our Claude Code tips and tricks guide for more MCP server patterns.

GitHub: n8n-io/n8n — 187K ⭐ | TypeScript | Fair Use

Quick Comparison Table

The Bottom Line

The Claude Code ecosystem is moving fast. These 10 tools represent what shipped in roughly one month — and several already have tens of thousands of stars. The three must-installs:

For the full picture on Claude Code optimization, don't miss our 32 Claude Code tips and tricks guide and our breakdown of cutting AI coding costs by 100x with DeepSeek V4.