This August, 34 projects captured the attention of developers, with several recurring favorites alongside new standout tools. This roundup highlights the most significant projects from the past month.

Deep Seek Harness

Deep Seek Harness is an agent runtime environment where core components such as the model adapter, tool registry, session log, and agent loop are designed as plugins. This modularity allows for flexible composition of these elements through profiles, enabling applications for browser UIs, headless jobs, SDK access, or automation. Patch layers provide a mechanism to modify configurations without altering the privileged core. Session management utilizes an append-only event log, consolidating replay and persistence functionalities.

AnyDoc

AnyDoc converts various document formats, including Office files, OpenDocument, EPUB, CSV, RTF, and text-based PDFs, into a consistent Markdown flavor. The core engine, written in Rust, automatically detects file formats from their content and preserves structural elements like tables, footnotes, equations, and speaker notes. This converter is accessible via a Command Line Interface (CLI), Node.js, Python, and WebAssembly. The local conversion capability of the browser build makes it particularly suitable for handling private files.

Comp CRM

Comp CRM redefines the conventional AI CRM design by assigning ownership of the work queue to the agent, while the interface displays discovered information and its rationale. The system proactively researches contacts on a schedule, records strong evidence directly, and routes weaker matches to human operators for review, thereby minimizing guesswork. Durable sessions ensure continuity across redeployments, and the shell sandbox operates without network access or database credentials.

ARC Task Gen

ARC Task Gen generates novel grid puzzles that align with the measurable characteristics of the public ARC AI 1 evaluation set, aiming to avoid the contamination of existing benchmarks. The system samples grid size, colors, and example counts from real tasks, producing one puzzle per model call. It then filters out semantically similar duplicates and malformed outputs. The generated results adhere to standard ARC JSON format, ensuring compatibility with existing evaluation harnesses.

Kimi K3 in C

Kimi K3 in C enables the execution of the Moonshot's 2.78 trillion parameter model through a C99 inference engine without requiring a GPU, specialized hardware (blast), or an ML framework. It maintains a portion of the dense model trunk in memory while streaming the remainder from disk, and caches routed experts within the available RAM. The developers reported a peak memory usage of 8.24 GB and a processing time of 26.5 seconds per token. The checkpoint and packed trunk still necessitate approximately 1.7 terabytes of fast storage.

OpenGym

OpenGym is a self-hosted application for tracking workouts and bodyweight exercises, storing training history in plain JSON files locally. It facilitates the creation of weekly routines, guides individual sessions with timers and historical weight data, and visualizes progress and personal records through charts. The web application functions offline, synchronizes across devices, and utilizes passkeys for authentication.

ThreeUI

ThreeUI is an open-source catalog of interactive 3JS components that can be previewed, customized, and integrated into React projects. Each component page features a live renderer, control elements, a variant selector, and the complete community source code. The NPM package supports both top-level and per-component imports. This repository exclusively contains the free community catalog.

IP as logo

IP as logo is an agent skill designed to generate simple, rounded mascot concepts from product briefs, offering an alternative to vague prompts for image models. It proposes three distinct directions and subsequently generates six square candidates characterized by limited shapes, three semantic colors, and specific lower-corner compositions. The repository provides instructions for use but does not include a generator; a supported image model is still required.

Praxis

Praxis transforms an existing runnable project into a persistent research loop. Parallel agents test competing implementations, and a task-owned evaluator converts results into structured evidence. Subsequent generations leverage this evidence to guide future experimentation. Runs can be monitored, paused, resumed, and replayed, with provenance attached to reported improvements.

Anti-slop

Anti-slop is an opinionated Oxlint plugin for TypeScript and JavaScript teams that enforces strictness in Continuous Integration (CI) environments by flagging unsafe shortcuts. Its rules prohibit chained assertions, unknown widening, module mocking, and assertions lacking nearby safety explanations. An accompanying agent skill automates the integration of the plugin and merges the lint configuration into the repository.

Open bot

Open bot assigns each AI coworker its own container, browser profile, file system, and granted tools. All actions are routed through a policy gateway prior to execution. Users can monitor the screen, take manual control for tasks like logins or two-factor authentication, and review permitted, refused, or failed actions post-execution. The system is framework-independent, allowing any AGUI agent to connect.

Kumara

Kumara is a cross-platform team chat application where human users and AI agents can share conversations, a task board, and a calendar. Agents maintain personas in memory, claim tasks, and coordinate through server-side freshness checks to prevent conflicts between stale replies and newer decisions. Users have the option to run agent brains in managed cloud pods or connect local cloud code and Codex instances without sharing provider keys with the server.

Fushi

Fushi is a self-contained terminal coding agent that routes work across multiple Large Language Model (LLM) providers, avoiding vendor lock-in. Its think-act-verify loop supports file editing, shell command execution, interaction with MCP servers, launching sub-agents, and session preservation through in-memory checkpoints. The agent is distributed as a single Go binary with integrated permission controls and command classification.

Anatomy Ally A

Anatomy Ally A transforms a web browser into an interactive anatomy study tool. Users can rotate and zoom 3D organs, isolate layers, open cross-sections, compare specimens, and access lessons or quizzes without any installation. The library covers major body systems with searchable organ notes, clinical context, and support for multiple languages. It is built using 3.js and Next.js.

My context

My context is a local-first desktop memory layer for AI tools, designed around user work history rather than isolated chat sessions. It ingests messages, documents, and meeting records into an SQLite database, establishing links between people, projects, and evidence within a personal context graph. Search queries provide source references, and the system's digital self can draft replies, requiring user approval before sending.

Unlazy

Unlazy is an agent skill that ensures task completion is validated by actual passing checks. Before commencing substantial work, it translates requirements into an acceptance ledger with runnable commands and expected outputs. Agents can divide work into scoped sub-tasks, claim non-overlapping paths, and re-verify returned results prior to reporting completion. It is compatible with Claude code and CodeX.

FX

FX is a compact coding agent developed in Zig, intended to function more like a Unix command than a terminal-based IDE. The repository lists a 7.8-byte binary that supports interactive sessions, one-shot JSON output, MCP tools, and sub-agents. It can be run natively, connected to editors via ACP, or embedded as WebAssembly within larger applications.

H3.C

H3.C enables native execution of Minimax H3 video generation on Apple silicon using C and Metal, bypassing the need for hosted services. It handles prompt-to-video generation with audio, supports first and last frame conditioning, and accepts ordered image, video, and audio references. An interactive mode keeps the model primed between prompts, while SSD streaming offers a trade-off between speed and reduced resident storage.

SolAdvisor

SolAdvisor divides AI coding tasks among four specialized lanes rather than assigning all work to a single agent. Saul manages requirements and final acceptance, Luna handles routine tasks, and Terra addresses higher-risk changes. A separate Saul reviewer then examines the diff with read-only access, ensuring the code's author does not also grade its own work. It operates as a plugin within Codex and explicitly reports any fallback to a weaker model.

Claudish to English

Claudish to English translates Claude code messages into plain English directly within the terminal, without altering the original transcript. It inserts a plain English block below each response or can replace it entirely. By default, it runs locally using a Llama model, ensuring data privacy. If a translation fails due to timeouts or missing keys, the original text is displayed.

Walgit

Walgit is a Git server that eliminates the need for a database and leader election. A single binary sits in front of S3 or GCS, with a write-ahead log in object storage serving as the source of truth. Each local repository functions as a cache. Any instance can accept a push, as there is no election protocol to determine leadership. Clones are served as static files from the bucket or a Content Delivery Network (CDN), minimizing server data transfer.

Morphicons

Morphicons morphs any icon into any other across various libraries such as Lucide, Tabler, and Heroicons, without requiring manual declaration of part mappings. It employs Procrustes analysis to automatically determine optimal rotation and interpolates in polar coordinates, preventing shape distortion during transitions. Animations can be interrupted mid-flight while maintaining their velocity.

Phone harness

Phone harness is a system built entirely using macOS's native functionalities. iPhone mirroring displays the phone's screen as a desktop window. The harness then captures screenshots, utilizes Apple's Vision OCR to identify on-screen text, and simulates clicks at detected coordinates. A subsequent screenshot verifies the action's outcome. This process requires no jailbreaking or injected applications.

Terminal code

Terminal code integrates the full VS Code experience directly into the terminal, not merely a look-alike interface. It achieves this by stacking two existing components: Coder's code server, which runs VS Code in a browser, and a custom-built browser that renders the VS Code interface within the terminal using the Kitty graphics protocol.

Mango disk

Mango disk is a cross-platform disk cleaner that prioritizes displaying space consumption before initiating deletions. It visualizes large folders using a treemap, identifies exact duplicate files by content, and recognizes rebuildable caches from browsers, developer tools, containers, and local AI models. Cleanup operations require user review and confirmation, with all actions being recorded.

Joy AI video edit

Joy AI video edit allows for video editing during the recording process, eliminating the need to wait for clips to finish or manage offline render queues. Instructions are processed as frames stream in, utilizing a 16 billion parameter diffusion transformer that operates causally on each frame without future peeking. It supports background swaps, subject changes, and style transfer, achieving a reported speed of 30.19 frames per second at 720x1280 resolution.

Imaginary 3D

Imaginary 3D generates CAD models for hardware enclosures, moving beyond static meshes. Users describe the product, provide reference images and dimensions, and the system constructs covers, hinges, latches, and the complete assembly, incorporating clearances and print tolerances. The agent operates on actual geometry, not text interpretations. It measures models, checks for collisions along hinge motion paths, and commits the design only after passing all checks.

Zephyr

Zephyr is an open-source virtualization platform designed as an alternative to VMware. It encompasses compute, virtual networking with security groups, and storage with snapshot and backup capabilities, built on an asynchronous, stateless architecture. Backed by ZStack, a company with established production deployments, Zephyr provides specific migration tools and guides for transitioning from VMware.

Molly

Molly is a browser engine developed for AI agents, prioritizing functionality over visual rendering. Written in Rust, it executes real JavaScript through V8 and provides a Document Object Model (DOM), but bypasses layout and rendering unless a screenshot is requested. The development team reports a peak memory usage of 73 megabytes, compared to Chrome headless's 7773 megabytes, and a Connected Device Protocol (CDP) ready time of 34.85 milliseconds, versus Chromium's 169.37 milliseconds. It includes an MCP server for direct agent control.

Koji

Koji is a five-chapter interactive experience depicting a night walk through a Kyoto mountain temple, rendered live in 3js. The entire application is contained within a single static HTML file, requiring no build steps or package managers. Serving the file initiates the experience, where scrolling moves a WebGL camera through a procedural temple, torii gate, and drifting leaves layered over pre-rendered artwork, with dynamic fog and bloom effects across five sections.

Virtual Mac on iPad

Virtual Mac on iPad enables running macOS on an iPad Pro equipped with an M1 or M2 chip, overcoming iPadOS limitations that prevent Xcode installation. It utilizes hardware virtualization and reconstructs Apple's frameworks on jailbroken devices using Dopamine or TrollStore X. Upon booting, users gain a full macOS session with support for Magic Keyboard or touchscreen input, and functional Xcode.

Gemma translator

Gemma translator transforms a Raspberry Pi 5 into an offline conversational translator for two individuals speaking across a table, requiring no internet connection after initial setup. Speech input is processed by Gemma 4 for translation, and the output is delivered as speech. The interface is divided into two lanes, with a keyboard that flips for the active speaker. The entire translation pipeline operates locally.

Frontier Agent

Frontier Agent is a terminal application designed for extended agent tasks, rather than brief chat interactions. It supports running React for a single focused agent or an agent team, which distributes work across parallel sub-agents on a live task board managed by a coordinator. All operations are confined to a sandboxed file system with read-only inputs, a designated working directory, and separate outputs to prevent file conflicts.

ScrollCraft

ScrollCraft is a Claude code skill for developing scroll-driven websites that deviate from common AI-generated aesthetics. It enforces a refuse list that excludes elements like gradient text and feature card grids, and mandates one of eight predefined page grammars. A fingerprint gate assesses each new site against the user's previous work, requiring differences in at least four dimensions.