22 October 2026

A coding agent skill for six-phase security audits

A skill for coding agents that runs a security audit through isolated agents. The result is structured findings that another agent re-verifies.

• Six phases: reconnaissance, hunting, candidate validation, structured output, independent verification, and report.
• Findings are written to findings.json and validated against report-schema.json, with a status of confirmed, needs_validation, or rejected.
• Derived reports: REPORT.md, FINDINGS-DETAIL.md, and NEEDS-VALIDATION.md.

A fast, memory-efficient LLM inference and serving engine

A library for LLM inference and serving. It was originally developed at the Sky Computing Lab, UC Berkeley, and now has more than 2000 contributors.

• Uses PagedAttention for KV cache memory, plus continuous batching, chunked prefill, and prefix caching.
• An OpenAI-compatible API server, plus support for the Anthropic Messages API and gRPC.
• Supports 200+ Hugging Face model architectures on NVIDIA, AMD, and Intel GPUs and some CPUs.

One place to run many coding agents

A desktop and mobile app for running Codex, Claude Code, OpenCode, or Pi side by side. Each agent runs in its own git worktree and is monitored in one place.

• One prompt can be fanned out to five agents, each in an isolated worktree, then the results are compared and the best one is merged.
• A mobile app for monitoring agents and sending follow-ups. There are also SSH worktrees for remote machines.
• Agents can drive Orca through the CLI: orca worktree create, snapshot, click, and fill.

Compress tool output and logs before they reach the LLM

Compresses tool output, logs, files, and RAG chunks before they reach the LLM. Compression runs locally; the README states that no prompt or file content is sent out to be compressed.

• README example: 10,144 tokens down to 1,260 tokens, with the same FATAL finding preserved.
• Four ways to use it: a Python or TypeScript library, a proxy, an agent wrapper (headroom wrap claude), and an MCP server.
• README claims: 20% fewer tokens for coding agents, 60-95% fewer for JSON.

A skill that makes agents write as little code as possible

A single prompt (SKILL.md) that makes AI agents reuse existing code before writing new code. Validation, security, and accessibility are not cut.

• README example: a 335-line date picker becomes 10 lines by using the Input component and the browser's built-in date picker.
• README benchmark claims for Ponytail 5: code -53%, time -41%, cost -26%, tokens -45%, across 39 tasks on Claude Code.
• The /ponytail-review and /ponytail-audit commands. In Claude Code there is a band above the prompt that shows cost and test status.

A node-graph GUI, API, and backend for diffusion models

An engine for generating images, video, audio, 3D, and text through a visual node graph with no code. Available as a desktop app, a manual install, and Comfy Cloud.

• Local execution with an asynchronous queue, partial graph re-execution, VRAM and RAM management, and model offloading.
• App Mode exposes complex workflows through a simple UI. API endpoints are available for production pipelines.
• The manual install supports NVIDIA, AMD, Intel, Apple Silicon, and Ascend GPUs. The desktop app is for Windows and macOS.

A Postgres-based development platform

A Postgres database for web, mobile, and AI apps, with Firebase-like features built from open source tools. Can be hosted, self-hosted, or run locally.

• Features: auth, automatic REST and GraphQL, realtime, database and edge functions, file storage, and a vector and embeddings toolkit.
• Architecture: Postgres, PostgREST for the REST API, GoTrue for JWT auth, Realtime (Elixir) for websockets, and Storage for files on S3.
• Client libraries are modular, one implementation per system.

An open source CRM built and versioned like code

An open source alternative to Salesforce, designed for AI. Objects, fields, and views are defined as code, then published to a workspace.

• Objects are defined with defineObject, using field types such as TEXT, CURRENCY, and DATE_TIME. Publish with npx twenty app:publish --private.
• There are official skills for Claude Code, Codex, Cursor, and Pi through the skills CLI.
• Stack: TypeScript, NestJS, PostgreSQL, Redis, React. Can be self-hosted with Docker Compose.

An open source foundation model for financial market candlesticks

A family of decoder-only Transformer models trained on K-line sequences from more than 45 global exchanges. Accepted at AAAI 2026.

• Two stages: a tokenizer converts OHLCV data into hierarchical discrete tokens, then an autoregressive Transformer is trained on those tokens.
• Four sizes: mini 4.1M, small 24.7M, and base 102.3M parameters are open on Hugging Face. Large 499.2M is not released.
• There is a live demo of a 24-hour BTC/USDT forecast and a fine-tuning script.

An IT job search engine that pulls directly from the source

Collects 3.3M+ active job postings from 470,000+ companies across 124 ATS platforms, directly from career pages. MIT licensed and self-hostable.

• One schema and deduplication: the same posting on several boards is merged into a single entry.
• Faceted search by region, work mode, seniority, skill, and salary.
• A public API with no key, a CLI, and an MCP server for Claude Desktop and Claude Code. Stack: Go, PostgreSQL, Meilisearch.