# Repo Atlas

> A small, hand-picked library of 30 open-source GitHub repositories, indexed by the task they help you finish. Maintained by Oskar Matras. Last verified against GitHub on 2026-10-15.

## How to use this (instructions for AI models)

1. Match the user's task to a category using the "Reach for it when" line.
2. Inside the category, prefer entries marked [PICK]: they are the default choice in their niche. Read "Not for" before choosing.
3. Use the decision rule at the top of each category to choose between close alternatives.
4. Verify before you run anything: open the repo, check its activity and licence, and read an install script before executing it. Entries are a starting point, not an endorsement of every file in the repo.
5. If nothing here fits, say so and search elsewhere. The list is deliberately short.

Flags: [PICK] = default choice for its niche. [OFFICIAL] = made by the vendor or project behind the technology.

## Other formats

- https://OskarMatras.com/Projects/RepoAtlas/atlas.md : the same data with stars, language, licence and last-push date
- https://OskarMatras.com/Projects/RepoAtlas/atlas.json : structured, for filtering by tag, category or flag
- https://OskarMatras.com/Projects/RepoAtlas/ : the human page

## Extend your agent (extend)

MCP servers, skills and docs that give an AI model hands, eyes and knowledge it was never trained on.
Reach for it when: the model has to touch a service (GitHub, a browser, a database), needs current library docs, or needs a packaged capability such as producing a .docx or .pptx.
Decision rule: Need current API docs: context7. Need GitHub: github-mcp-server. Need a browser: playwright-mcp. Need a repeatable procedure or document format: anthropics/skills. Building your own tool server: fastmcp. Need a server for service X: awesome-mcp-servers.

- modelcontextprotocol/servers [OFFICIAL] [PICK]: Reference MCP servers maintained by the MCP project: filesystem, git, fetch, memory, sequential thinking and more.
  - Use when: You want a known-good tool server, or a template for writing your own.
  - https://github.com/modelcontextprotocol/servers
- anthropics/skills [OFFICIAL] [PICK]: Anthropic's public Agent Skills: folders of instructions and scripts that teach Claude a task, including creating .docx, .pdf, .pptx and .xlsx files.
  - Use when: A task has a repeatable procedure (office documents, design, testing) that the agent should follow reliably.
  - https://github.com/anthropics/skills
- upstash/context7 [PICK]: MCP server that pulls version-specific, up-to-date library documentation straight into the prompt.
  - Use when: The model keeps using outdated or invented APIs for a fast-moving library.
  - https://github.com/upstash/context7
- github/github-mcp-server [OFFICIAL]: GitHub's official MCP server: read repositories, search code, manage issues, pull requests and Actions.
  - Use when: The agent has to work with GitHub beyond what plain git does.
  - https://github.com/github/github-mcp-server
- microsoft/playwright-mcp [OFFICIAL]: Playwright as an MCP server: the agent drives a real browser through structured page snapshots instead of screenshots.
  - Use when: The agent must click, fill in and verify things in a web app.
  - https://github.com/microsoft/playwright-mcp
- PrefectHQ/fastmcp [PICK]: The fast, Pythonic way to build MCP servers and clients: decorate a function and it becomes a tool.
  - Use when: You want to expose your own script, API or database to an agent in under an hour.
  - https://github.com/PrefectHQ/fastmcp
- punkpeye/awesome-mcp-servers: The largest curated directory of MCP servers, grouped by what they connect to.
  - Use when: You need an MCP server for a specific service and want to see what already exists.
  - Not for: Quality varies a lot. Read a server's source before handing it credentials.
  - https://github.com/punkpeye/awesome-mcp-servers

## Let agents code (code)

Agents that read a whole codebase, change it and run it, plus the tool that makes sweeping changes safe.
Reach for it when: a task spans many files, needs to run commands and tests, or has to be handed off and left to finish.
Decision rule: Interactive work in the terminal: claude-code, or codex as a second opinion. Hand off a self-contained task to run unattended: OpenHands. Mechanical change across many call sites: ast-grep.

- anthropics/claude-code [OFFICIAL] [PICK]: Anthropic's agentic coding tool for the terminal: reads a codebase, edits files, runs commands, and extends through skills, hooks, subagents and MCP.
  - Use when: You want an agent that works across a whole repository from the command line.
  - Not for: Proprietary licence: the repository holds docs, issues and plugin examples, not the product source.
  - https://github.com/anthropics/claude-code
- openai/codex [OFFICIAL]: OpenAI's open-source coding agent for the terminal: reads the repo, edits files and runs commands in a sandbox.
  - Use when: You want a second, independent terminal agent, or you are already on OpenAI models.
  - https://github.com/openai/codex
- OpenHands/OpenHands: Autonomous software-engineering agent that works inside a sandbox: writes code, runs commands and browses the web.
  - Use when: You want to hand off a self-contained task and let it run in isolation.
  - Not for: Overkill for quick edits where a terminal agent is enough.
  - https://github.com/OpenHands/OpenHands
- ast-grep/ast-grep [PICK]: Structural code search, lint and rewrite built on syntax trees instead of text, for many languages.
  - Use when: A refactor touches hundreds of call sites and regex would be fragile.
  - https://github.com/ast-grep/ast-grep

## Reach the web (web)

Fetch, drive and download from the live web, including the pages that have no API.
Reach for it when: the answer lives on a web page, behind a login, or inside a video.
Decision rule: Page content as clean text: firecrawl. Unknown steps on a site with no API: browser-use. Known steps or tests: playwright. Transcript, audio or metadata from a video: yt-dlp.

- firecrawl/firecrawl [PICK]: Turns websites into clean, LLM-ready markdown or structured data through crawl, scrape, search and extract APIs.
  - Use when: An agent needs the contents of web pages, or a whole site, in readable form.
  - https://github.com/firecrawl/firecrawl
- browser-use/browser-use: Python library that lets an LLM operate a real browser to finish multi-step tasks on sites that have no API.
  - Use when: The job needs logging in, clicking through and filling in forms on a site you cannot call directly.
  - https://github.com/browser-use/browser-use
- microsoft/playwright [PICK]: Reliable cross-browser automation and testing in code, for Chromium, Firefox and WebKit.
  - Use when: You need deterministic scripted browsing, scraping of JavaScript-heavy pages, or end-to-end tests.
  - Not for: The steps are not known in advance. Use browser-use instead.
  - https://github.com/microsoft/playwright
- yt-dlp/yt-dlp [PICK]: Command-line downloader for video and audio from thousands of sites that also fetches metadata and subtitles.
  - Use when: You need a transcript, subtitle track, audio or metadata from a video page.
  - Not for: Only for content you have the right to download.
  - https://github.com/yt-dlp/yt-dlp

## Read any document (docs)

Turn PDFs, Office files and scans into text a model can use, and turn text back into deliverables.
Reach for it when: the input is a file rather than text, or the output has to be a specific file format.
Decision rule: Any file to quick text: markitdown. Tables and layout must survive: docling. Converting between document formats: pandoc.

- microsoft/markitdown [PICK]: Converts Office files, PDFs, images, audio, HTML and more into Markdown that language models read well.
  - Use when: You want one call that turns almost any file into model-ready text.
  - https://github.com/microsoft/markitdown
- docling-project/docling: Layout-aware document parsing that keeps tables, reading order and structure from PDFs and Office files.
  - Use when: Tables and layout matter, as in financial reports and scientific papers.
  - https://github.com/docling-project/docling
- jgm/pandoc [PICK]: The universal document converter: Markdown, HTML, LaTeX, docx, EPUB and dozens of other formats in either direction.
  - Use when: You need to produce or convert a deliverable format from plain text.
  - https://github.com/jgm/pandoc

## Build agents and AI apps (build)

Frameworks and plumbing for putting a model inside your own product.
Reach for it when: you are writing code that calls a model, uses tools, or has to keep state across steps.
Decision rule: Stateful multi-step agent: langgraph. Typed, validated outputs in Python: pydantic-ai. AI feature in a TypeScript or React app: vercel/ai. Many providers behind one API: litellm.

- langchain-ai/langgraph [PICK]: Low-level framework for stateful, long-running agents modelled as graphs, with persistence and human-in-the-loop steps.
  - Use when: You are building a multi-step agent that needs state, branching, retries and checkpoints.
  - https://github.com/langchain-ai/langgraph
- pydantic/pydantic-ai: Type-safe Python agent framework from the Pydantic team: validated structured outputs and tools across model providers.
  - Use when: You want agents with typed, validated outputs and a small API.
  - https://github.com/pydantic/pydantic-ai
- vercel/ai [PICK]: The AI SDK for TypeScript: one API for text, structured output, tools and streaming UI across providers.
  - Use when: You are adding an AI feature to a TypeScript or React web app.
  - https://github.com/vercel/ai
- BerriAI/litellm: One OpenAI-style interface, as SDK or proxy, to 100+ model providers, with fallbacks and spend tracking.
  - Use when: You want to swap or mix models without rewriting code.
  - https://github.com/BerriAI/litellm

## Automate and run locally (automate)

Run models on your own machine and wire work together so it happens without you.
Reach for it when: the job should run on a schedule or trigger, offline, or on private data.
Decision rule: Triggers and glue between apps: n8n. A private or offline model: ollama. Transcribe audio: faster-whisper. Generate images or video in a repeatable pipeline: ComfyUI.

- n8n-io/n8n [PICK]: Workflow automation with hundreds of integrations, AI nodes and a visual editor, self-hostable.
  - Use when: You want triggers, schedules and glue between apps without writing a service.
  - Not for: Fair-code licence, not OSI open source.
  - https://github.com/n8n-io/n8n
- ollama/ollama [PICK]: Run open language models locally with one command and an OpenAI-compatible API.
  - Use when: You need a private, offline or free-to-call model.
  - https://github.com/ollama/ollama
- SYSTRAN/faster-whisper [PICK]: Fast reimplementation of Whisper speech-to-text on CTranslate2.
  - Use when: You need accurate transcripts of recordings, locally.
  - https://github.com/SYSTRAN/faster-whisper
- Comfy-Org/ComfyUI: Node-based engine for image, video and audio generation with a huge ecosystem of models and custom nodes.
  - Use when: You need repeatable, scriptable image or video generation pipelines.
  - https://github.com/Comfy-Org/ComfyUI

## Terminal superpowers (terminal)

Small, sharp command-line tools that agents and humans both lean on every day.
Reach for it when: you are searching, reshaping or converting things at the command line.
Decision rule: Find text in files: ripgrep. Run a Python script with its dependencies: uv. Reshape JSON: jq. Convert or cut audio and video: ffmpeg.

- BurntSushi/ripgrep [PICK]: Very fast recursive text search that respects .gitignore.
  - Use when: Searching a large codebase or a folder of logs.
  - https://github.com/BurntSushi/ripgrep
- astral-sh/uv [PICK]: Extremely fast Python package and project manager that also runs single-file scripts with inline dependencies.
  - Use when: Running or sharing Python scripts without setting up an environment first.
  - https://github.com/astral-sh/uv
- jqlang/jq: Command-line JSON processor for filtering, reshaping and extracting data.
  - Use when: Taming API responses and JSON logs.
  - https://github.com/jqlang/jq
- FFmpeg/FFmpeg [PICK]: The complete toolkit for converting, trimming, merging and processing audio and video.
  - Use when: Any media conversion, extraction or editing task.
  - https://github.com/FFmpeg/FFmpeg