Releases Koder, a local/offline AI coding and computer use harness with browser UI, supporting llama.cpp, MCP, visual models, and multi-task planning.
Warning: Incoming self-promotion of 11-weeks worth (1300 commits) of AI vibe-coding. I'll take the down-votes if they come - and I'll understand since it's pouring in with projects nowadays, but I thought I'd share this anyway. I'm releasing "koder" - my coding and computer use harness (agent+tooling) publicly today. It's meant for local/offline use on Linux, but since it's OpenAI compatible it's BYOM (bring-your-own-model) so you can use a cloud based ones as well. Focus has not been on being another universal one-size-fits-all solution, but rather being good for my specific scenario: Linux, llama.cpp, Qwen 3.6 27b Q8 - and it's absolutely rock solid for me. I still use Codex a lot (Koder is written using Codex!), but I find myself moving more and more stuff to the local side, which is the ultimate goal. So your results might differ with other models (for me Gemma 4 does not work well with this). Since 'koder' is a generic agent, you can throw all kinds of tasks at it. I've dobe reverse engineering using Ghidra and decompilers, I've coded tools, it even does OpenSCAD with visual feedback (code -> render loop), it does online research - it does more or less anything I throw at it. On the feature side it's fairly complete: skills, MCP, supports visual models, thinking on/off, caveman thinking compression, rich visualization towards user, milestone/task planning -> multi chat orchestration, embedded file browser, lane based milestone/task editor. As most of my other tooling this is written in Go, so grab a single binary and get going. It might be rough around the edges here and there, but it definitely at the "works very well for me" state. Why yet another agent? Because I wanted to deep dive, and I wanted things "my way". Maybe you'll like it too. https://github.com/lkarlslund/koder
OpenAI demonstrates the 'Computer Use' feature in Codex, allowing the AI to directly interact with local GUI applications on macOS using an accessibility framework and the fast Spark model for non-blocking, high-speed automation.
The author compares several agentic coding harnesses (Codex CLI, Claude Code, Gemini CLI, OpenCode, Pi) and finds Pi the leanest and best for local models, praising its simplicity and compatibility with Qwen 27B-MXFP8.
KingCoding is a mobile app that brings Claude Code and OpenAI Codex AI coding assistants to your pocket. It appears to be a developer tool launched on Product Hunt.
Juggler is an open-source GUI coding agent that provides a visual workbench for interacting with LLMs to edit code, featuring inspectable tool calls, branching threads, and editable context. It runs locally or remotely and supports multiple models.
The author built TinyHarness, a low-memory-footprint AI harness compatible with Ollama, Llama.cpp, and vllm, aiming to compete with tools like pi and opencode.