open-source take on Claude in Chrome. side-panel agent that reads, clicks and fills forms across your tabs bring your own model: Anthropic, OpenAI, Gemini or local Ollama no account, no backend, keys stay in Chrome storage it's beta, so keep it off anything irreversible github.com/DumDuky/myna-brow…

hidden claude code flag: --advisor claude --model sonnet --advisor opus sonnet handles the routine work, opus gets pulled in for planning, ambiguous failures and completion checks experimental, and it's not listed in claude --help code.claude.com/docs/en/advi…

a rhythm runner where the song builds the level 4 tiny neural nets (~22 MB total) listen to your mp3 in the tab: beats, notes, vocals, pitch. the game gets built from that same song, same world every time, and nothing gets uploaded play: struna.ai code: github.com/Felliks/Struna

my claude code sessions all look the same until one silently waits on a permission prompt agentvalley parks them in the mac notch: one row per agent and sub-agent, which model it's on, and a red badge when one needs you each agent gets its own pixel programmer. free + open source, mac with a notch only agentvalley.app

a 4B model that writes whole Three.js scenes from one prompt the Q4 file is 2.7 GB, small enough to run locally these scenes are from the model card, one self-contained HTML file each still an early preview, dev says full weights + dataset come later ollama run hf.co/ccore/Qwen3.5-4B-3D-Co… huggingface.co/ccore/Qwen3.5…

too many subagents in your agent? anthropic says try folding them back into one loop skills for the how-to, tools for hard limits. one customer saw roughly 2x faster runtime subagents still win for parallel work and context bloat claude.com/resources/article…

how much compute does it take to beat warcraft 3's hardest built-in AI? one desktop. a 3090 and 16 cpu cores ~50 copies of the real game at 15x speed, three weeks of self-play, 840k games. it now wins 89% against insane alphastar used 32 TPUs per agent for 44 days robin kahlow's write-up + code: warlock.ai/blog/warcraftsim-…

for redesigning an existing site, opus has been closer to what i want than codex. way less back and forth but i don't buy that one model wins everything what's the one task where you'd still pick the other one?

tiny mac menu bar app that shows how much of your claude code, codex, gemini and grok limits you've burned 5-hour and weekly bars, plus when each one resets. it asks the official CLIs, never touches your tokens github.com/dancik/limitbar

Scrolling the terminal to work out what your coding agent just did gets old agentplay replays each session as a map and a timeline: what it read, what it edited, what failed, what shipped works with Claude Code, Codex, Cursor, Copilot, Gemini CLI. local, one command: uvx agentplay github.com/Entelligence-AI/a…

the maker kept screenshotting handwritten calculus into claude and reading long answers back. wanted a teacher looking over the shoulder instead now you write by hand, press one key, and it circles the line where you slipped and gives a hint, not the answer. box an integral and you get a 3blue1brown-style animation works with claude code, codex or gemini. linux only for now github.com/AbdullahAlhariri/…

went to sleep feeling productive. agent was running all night it asked one question at 1:12 am and has been waiting for me since

claude code: /btw ask a side question without it touching the conversation. it answers from what the session already knows, even while claude is mid-task. can't read files or run anything, so it's for "what was that file called again", not for digging.

no shell is running here. a 360M model just predicts the next screen after every keypress, and it still remembers the file it wrote a few commands ago. vim, tmux and snake work too. type pythoon and it happily opens python. a real shell wouldn't. simedw.com/projects/worldtui…

kitten tts 2 is out and it clones a voice from about 5 seconds of audio • 1.7B model, ~506 MiB download • runs on a plain CPU, several times realtime on apple silicon • 47 built-in voices plus cloning the original kitten was a 25 MB model. the new one is a different animal. github.com/KittenML/KittenTT…

178 motion graphics from Claude, Kimi K3 and GPT-6 Astra, each next to the exact prompt behind it open one you like, copy the prompt, swap in your subject credited to the original creators. save it for your next motion idea motionpromptgallery.pages.de…

coding agents can write a whole launch film in code and still can't hear a frame of it. earshot gives them ears. it describes a sound in words, finds the best moment in a clip, checks audio against the picture cuts. CPU only, CLI or MCP. github.com/Leonxlnx/earshot

mlx-lm: 87 tokens/s with 4 requests at once, still 87 with 8. Locals: 160 → 271 local server for coding agents on Apple Silicon. the gain shows up with several agents at once, one request is about the same works with opencode, cline, aider, claude code github.com/Generated-Compute…

New coding model dropped this week and it's already sitting at $0 on openrouter flint by recursive. 40B MoE with only ~2.5B active, 262K context, handles tool calls plug it into cline, kilo code or any openai-compatible agent and throw a bug at it weights are on hugging face too if you'd rather run it yourself catch: the free endpoint is rate limited, and the benchmark scores are their own openrouter.ai/recursive/flin…

Mario 64's courtyard plaque is a 32x32 texture the text on it is about 3 pixels tall people have argued over what it says for 30 years one dev simulated how a 1996 artist would've made it (font, shrink, blur) and searched for the text that best fits the pixels best match: "Eternal Star / We hope so" not "L is real 2401" code + full report are open, you can try to beat the score github.com/supermochi64/sm64…