Understand any codebase in minutes. Deliver context to your team or to Claude Code, Co-Work, and MCP clients.
Onboard turns repositories into institutional memory for humans and AI agents.
| ORGANIZATION | REPO | ONBOARD INSIGHT | CONTRIBUTORS | LAST COMMIT |
|---|---|---|---|---|
| anthropics | anthropic-sdk-python | Auto-generated from OpenAPI spec but 40% of client code is hand-written streaming logic with custom retry backoff | 85 | 3 hours ago |
| react | Fiber reconciler rewritten 3x; 80% of complexity lives in 14 scheduler files most contributors never touch | 1,600+ | 2 days ago | |
| vercel | next.js | 90% of build complexity in 12 compiler transforms; README documents App Router but Pages Router still handles 60% of traffic patterns | 3,200+ | 1 day ago |
| django | django | ORM query compiler has undocumented edge cases only found in 2014 mailing list threads; 3 core maintainers own all migration logic | 2,400+ | 1 day ago |
| fastapi | fastapi | Entire framework is a thin wrapper around Starlette; real magic is in the dependency injection system (1 file, 600 lines) | 680+ | 5 days ago |
| langchain-ai | langchain | Abstraction layers change faster than docs; 5 different chain APIs coexist, only LCEL is actively maintained | 2,100+ | 4 hours ago |
| astral-sh | uv | Reimplemented pip's resolver in Rust with a custom pubgrub fork; single contributor wrote the entire virtual environment layer | 780+ | 3 hours ago |
| huggingface | transformers | Each model is copy-pasted (not inherited) by design; changing the base attention class means patching 200+ files manually | 2,700+ | 2 hours ago |
| qdrant | qdrant | Custom HNSW diverges from the paper in 3 ways; Raft consensus bolted on after v1.0 with a hand-rolled WAL | 190+ | 1 day ago |
| anthropics | claude-code | Tool permission system is the real architecture; 70% of code is sandboxing and safety rails, not LLM orchestration | 45 | 6 hours ago |
| tailwindlabs | tailwindcss | v4 is a ground-up rewrite with a Rust-based engine; plugin API completely different, most v3 plugins will break | 400+ | 8 hours ago |
| microsoft | playwright | Browser protocol bindings are auto-generated from Chromium DevTools Protocol; test runner shares zero code with the core engine | 500+ | 12 hours ago |
| vercel | ai | Unified streaming interface across 12 LLM providers; useChat hook abstracts 300 lines of SSE parsing into a single React call | 860+ | 1 day ago |
| shadcn-ui | ui | Not a component library: a code generation CLI. Components are copy-pasted into your project, not imported. Radix primitives under the hood | 1,400+ | 3 hours ago |
| better-auth | better-auth | Framework-agnostic auth with 28 strategies; session management is a custom token rotator, not JWTs. Plugin system bolted on in v0.4 | 280+ | 2 days ago |
| astral-sh | ruff | Reimplemented 800+ Flake8/isort/pyupgrade rules in Rust; the parser alone is faster than CPython importing the AST module | 950+ | 5 hours ago |
| anthropics | anthropic-sdk-typescript | Generated from same OpenAPI spec as Python SDK; streaming uses AsyncIterator pattern but tool_use events require manual reassembly from deltas | 62 | 6 hours ago |
| tanstack | table | Headless by design: zero DOM code in core. Column definitions are type-level contracts; the generic system prevents 90% of runtime errors | 420+ | 3 days ago |
| pallets | flask | Application and request context stack is a thread-local proxy chain that breaks in async. Blueprints share state in ways the docs don't cover | 700+ | 2 days ago |
| tokio-rs | tokio | Work-stealing scheduler is the core innovation; io_uring support exists but is behind a feature flag with 40+ known edge cases on older kernels | 900+ | 1 day ago |
| pydantic | pydantic | v2 core is written in Rust (pydantic-core); the Python layer is a thin wrapper. model_validator and field_validator have subtle ordering differences | 580+ | 8 hours ago |
| supabase | supabase | A monorepo of 14 independent services glued by Kong API gateway config; the "magic" is PostgREST auto-generating REST from your Postgres schema | 1,200+ | 4 hours ago |
| tailwindlabs | headlessui | Accessibility-first component primitives with zero styling; separate React and Vue implementations that share no code but mirror the same API | 240+ | 5 days ago |
| openai | openai-python | Stainless-generated client with manual patches for streaming; beta namespace holds features that graduated 6 months ago but were never moved | 350+ | 1 day ago |
| microsoft | TypeScript | The checker is a single 40K-line file; type narrowing alone is 3K lines of control flow analysis. No plugin system exists by design | 960+ | 12 hours ago |
| langchain-ai | langgraph | State machine for LLM agents; checkpointing serializes full graph state to SQLite/Postgres. The "graph" is actually a Pregel-inspired message passing system | 320+ | 2 hours ago |
| rust-lang | rust | The borrow checker is 5% of the codebase; 60% is codegen, diagnostics, and the trait solver. MIR optimizations happen in 12 separate passes | 5,100+ | 3 hours ago |
| ollama | ollama | Go server wrapping llama.cpp with a Docker-like model registry; quantization happens at pull time, not inference. Metal/CUDA backends are separate C++ codepaths | 520+ | 4 hours ago |
| biomejs | biome | Rust-based linter and formatter replacing ESLint+Prettier; the parser handles JSX, TS, JSON, CSS in one unified AST. 97% Prettier-compatible by design | 380+ | 6 hours ago |
| prisma | prisma | Schema-first ORM with a Rust query engine compiled to WASM; migrations are SQL files generated from schema diffs. The client is fully type-safe codegen | 900+ | 1 day ago |
| drizzle-team | drizzle-orm | TypeScript ORM where queries ARE the schema; no codegen step. SQL output is predictable because the query builder maps 1:1 to SQL syntax | 340+ | 8 hours ago |
| pola-rs | polars | Apache Arrow columnar engine in Rust with lazy evaluation; query optimizer rewrites your chain before execution. Python API is a thin PyO3 binding | 450+ | 5 hours ago |
| spring-projects | spring-boot | Auto-configuration scans 180+ condition annotations at startup; the "magic" is a META-INF/spring.factories file that wires everything. GraalVM native breaks half of them | 1,100+ | 1 day ago |
| oven-sh | bun | Zig runtime replacing Node; the bundler, test runner, and package manager share a single AST. JavaScriptCore instead of V8, with custom native bindings in C++ | 620+ | 3 hours ago |
| vitejs | vite | Dev server uses native ESM (no bundling); production uses Rollup. These are fundamentally different codepaths that sometimes produce different behavior | 900+ | 6 hours ago |
| schemathesis | schemathesis | Property-based API testing from OpenAPI specs; generates edge cases your manual tests miss. Stateful testing replays sequences of API calls to find ordering bugs | 120+ | 5 days ago |
| duckdb | duckdb | In-process OLAP database in C++; reads Parquet, CSV, JSON directly without import. The vectorized execution engine processes data in 2048-row batches | 370+ | 4 hours ago |
| openai | openai-agents-python | Lightweight agent framework with handoff pattern; agents transfer control via tool calls. Tracing is built-in but the guardrails system runs as a parallel LLM call | 180+ | 1 day ago |
| mrdoob | three.js | WebGL abstraction where every frame rebuilds the render list; the material system compiles GLSL shaders on the fly. WebGPU renderer is a parallel rewrite sharing zero code | 1,800+ | 12 hours ago |
| vitest-dev | vitest | Jest-compatible test runner powered by Vite's transform pipeline; in-source testing lets you write tests next to code. Worker threads isolate each test file | 440+ | 2 days ago |
| BerriAI | litellm | OpenAI-compatible proxy for 100+ LLM providers; the provider mapping is a 4K-line dictionary. Spend tracking and rate limiting bolted on via middleware | 500+ | 2 hours ago |
| expressjs | express | The entire framework is middleware chaining on top of Node http module; v5 was in beta for 10 years. Router is a linear scan, not a trie | 340+ | 3 days ago |
| vercel | turborepo | Rust-based monorepo build system; task hashing determines what to rebuild. Remote cache is a content-addressable store that skips 80% of CI on average | 460+ | 1 day ago |
| anthropics | skills | Declarative agent skill definitions: each skill is a TOML manifest + prompt template. Claude Code discovers and loads them at runtime from the registry | 290+ | 1 hour ago |
| twbs | bootstrap | Sass variables drive the entire design token system; the grid is now CSS Grid but falls back to flexbox. JS plugins use a custom Data API that mirrors jQuery patterns | 1,300+ | 3 days ago |
| sveltejs | svelte | Svelte 5 runes are a compiler-driven reactivity system replacing stores; the compiler outputs vanilla JS with no runtime. Transitions are CSS animations generated at compile time | 780+ | 4 hours ago |
| honojs | hono | Runs on every JS runtime (Cloudflare, Deno, Bun, Node) with the same API. The router uses a RegExpRouter that compiles all routes into a single regex for O(1) matching | 400+ | 6 hours ago |
| colinhacks | zod | Runtime validation that infers TypeScript types; the type gymnastics in the generic chain are 2K lines of conditional types. Used by tRPC, Astro, and React Hook Form | 600+ | 2 days ago |
| denoland | deno | Rust runtime embedding V8 with built-in TypeScript; permission system is capability-based. The LSP, formatter, linter, and test runner are all in the single binary | 840+ | 3 hours ago |
| withastro | astro | Island architecture ships zero JS by default; components from React, Vue, and Svelte coexist in one page. Content Collections validate frontmatter with Zod at build time | 740+ | 5 hours ago |
| trpc | trpc | End-to-end type safety with no codegen; the client infers server types via TypeScript generics. v11 replaced the middleware chain with a pipe-based composition model | 350+ | 1 day ago |
| grafana | grafana | Go backend with React frontend in a single repo; 60+ data source plugins share a common SDK. The dashboard JSON model is the real API, not the UI | 3,800+ | 2 hours ago |
| tinygrad | tinygrad | Entire ML framework in under 5K lines of Python; the lazy evaluation engine fuses operations into GPU kernels. Targets AMD, NVIDIA, Apple Silicon, and custom accelerators | 280+ | 1 hour ago |
| kubernetes | kubernetes | The API server is a generic CRUD framework; all resources (Pods, Services) are just CRDs with special controllers. Scheduler runs 13 scoring plugins in parallel per pod | 6,900+ | 1 hour ago |
| Effect-TS | effect | Algebraic effect system for TypeScript bringing ZIO patterns to JS; errors are typed in the channel. The fiber runtime handles structured concurrency without async/await | 290+ | 8 hours ago |
| hetznercloud | cli | Go CLI for Hetzner Cloud using cobra; supports JSON output, SSH key management, and server provisioning. Context system handles multiple projects with separate API tokens | 120+ | 4 days ago |
| remix-run | remix | Full-stack React framework built on web standards; loaders and actions run on the server, forms work without JS. Now a Vite plugin after merging with React Router v7 | 900+ | 1 day ago |
| spring-projects | spring-ai | Spring ecosystem for AI: unified API across OpenAI, Anthropic, Ollama with Spring Boot auto-config. Vector store abstraction supports 12 backends including pgvector | 310+ | 6 hours ago |