Tools Bench.

Product launches and open-source repos with enough signal to earn a second look.

Last Brew Time: Sep 2, 2026, 11:13 AM PT

Insight

The loudest AI momentum in this run isn't new agents, it's retooling the agents developers already have

01.SKILL WRAPPERS FOR EXISTING CODING AGENTS

02.COPYCAT DEMOS RIDING A SINGLE OFFICIAL MODEL DROP

03.QUIETER AGENT-FACING INFRASTRUCTURE FROM CREDIBLE BACKERS

Featured

GitHub50.6K

ChromeDevTools/chrome-devtools-mcp

Chrome DevTools for coding agents

x
Announcing our public preview of Chrome DevTools MCP!

Market Signal

Why It Has Market Pull

This is Google's own official MCP server connecting coding agents to a live Chrome instance, and it has become one of the most widely adopted MCP tools for browser debugging and automation in agent workflows. Backed by the Chrome DevTools team itself and showcased at Google I/O Connect Berlin, it has strong developer credibility and real hands-on praise on Hacker News.

  • Verified GitHub stats: 50,591 stars, 3,553 forks, pushed within the last day.
  • Only 87 open issues against 50k+ stars, a healthy ratio suggesting active triage rather than an abandoned spike.
  • Demoed live at Google I/O Connect Berlin and the GDE Europe Summit driving an agent to find and fix a Chrome performance issue end to end.
  • Officially maintained by the ChromeDevTools GitHub org, not an unaffiliated community wrapper.

feedbacks

What People Are Saying

  • "I've only used it a bit, but it's working well so far."HN comment

  • "Chrome Devtools is similarly an extremely high value MCP for me."HN comment

  • "now has an (experimental) CLI as well"HN comment

  • "identify a long task blocking the main thread, trace it to a specific function, and propose a fix"conference demo writeup

  • "gives your AI coding assistant access to the full power of Chrome DevTools for reliable automation, in-depth debugging, and performance analysis"MCP directory listing

  • "Announcing our public preview of Chrome DevTools MCP!"X post

GitHub18.4K

firecrawl/pdf-inspector

Fast Rust library for PDF inspection, classification, and text extraction. Intelligently detects scanned vs text-based PDFs to enable smart routing decisions.

Market Signal

Why It Has Market Pull

A fast Rust library from Firecrawl, the YC-backed web-scraping API company, for classifying and extracting text from PDFs while intelligently routing scanned pages to OCR. This is a real, actively maintained open-source project already powering Firecrawl's own production document-parsing stack, with genuine external contributors building on top of it, making it worth a closer look.

  • 18,434 GitHub stars and 1,244 forks, with commits pushed the same day it was checked.
  • 204 open issues and dozens of merged pull requests from both the Firecrawl team and outside contributors.
  • Backed by Firecrawl, which closed an oversubscribed $14.5M Series A (total funding $16.2M) with YC and Nexus Venture Partners participating, on a platform serving 350,000+ developers.
  • Benchmarked by Firecrawl against 200 real PDFs, beating pymupdf4llm, markitdown, and opendataloader on reading order and table structure.
  • An external developer built and shared PyO3 Python bindings specifically to use it inside a RAG pipeline.

feedbacks

What People Are Saying

  • "I needed to use pdf-inspector in a Python data processing pipeline for a RAG system, so I built PyO3 bindings around the library"GitHub issue

  • "the README and binding docs do not explain how to interpret these confidence values or what thresholds are appropriate"GitHub issue

  • "For right-to-left scripts, extractText returns each line's characters in visual order rather than logical order"GitHub issue

  • "pdf-inspector mangles Shannon's A Mathematical Theory of Communication — headings, footnotes, tables"GitHub PR

  • "outperforms pymupdf4llm, markitdown, and opendataloader on overall score, reading order, and table structure"Firecrawl blog post

  • "Two issues found while profiling this crate against real and adversarial PDFs: a decompression-bomb OOM"GitHub PR

Product Hunt498

Kilo Code for JetBrains

Kilo Code for JetBrains is a fully native, open-source coding agent for IntelliJ IDEA, WebStorm, PyCharm, GoLand, Rider, PhpStorm, CLion, RubyMine — any JetBrains IDE. Designed for both local and remote dev, with parallel agents in isolated worktrees, GitHub PRs and diffs inline, and 500+ models.

Market Signal

Why It Has Market Pull

Kilo Code is a fast-growing, open-source AI coding agent that just extended full native support to every major JetBrains IDE, backed by seed funding and hundreds of thousands of installs across VS Code and JetBrains. Its JetBrains launch was one of the strongest AI dev-tool launches this cycle, with deep, specific praise from real developers.

  • 670 upvotes and 273 comments on the JetBrains launch page; a prior launch ranked #1 Product of the Day.
  • $8 million seed funding raised.
  • Grown from 0 to 750,000+ combined downloads across VS Code and JetBrains.
  • 5.0/5 rating across 46 Product Hunt reviews from named developers.
  • Open-source, supports 500+ models at provider rates with no markup, across VS Code, JetBrains, CLI, and cloud agents.

feedbacks

What People Are Saying

  • "Pretty much has all the features I need. Many models to choose from."Product Hunt review

  • "Kilo Code's orchestrator mode splits it into subtasks, runs whatever model you point it at with no markup on the tokens."Product Hunt review

  • "The step chart and token use visualization, the worktrees implementation is chef's kiss."Product Hunt review

  • "The auth flow with Linux and Cursor and Brave was rough around the edges."Product Hunt review

  • "Kilo Code has been a game-changer, built two practical tools, received formal recognition and bonus."Product Hunt review

  • "No ability to use it within Zed, I hope that happens soon."Product Hunt review

HF Spaces337 likes

MiniMax H3 Turbo LoRA

Video generation with a synchronized soundtrack MiniMax-H3 — unquantized, split across two Spaces Joint video and soundtrack out of a single denoising pass, at bfloat16 with no quantization anywhere. This Space is the denoising half: the 61.73 GiB transformer and the two autoencoders. The 62.14 GiB Qwen3-VL conditioner runs in qwen3vl-conditioner, which this Space calls over the gradio API for every request. The weights are the public MiniMaxAI/MiniMax-H3 diffusers checkpoint. MiniMax-H3 is 195.9 GiB in bfloat16 and a ZeroGPU Space is evicted at 150 GB of storage. An unquantized single Space is therefore impossible, which is why quantized demos of it run NVFP4 or float8 weights. Cut the Mini...

x
6-step MiniMax H3 Turbo Generation with Maestro!

Market Signal

Why It Has Market Pull

This is the official Hugging Face demo from MiniMax itself for H3, the open-weight omni-modal model that generates synchronized video-plus-audio in a single pass and became the first open-weight model to top the Artificial Analysis video-editing leaderboard. With hundreds of likes, active feature requests from real users, and coverage from major tech press, this is a legitimate, well-covered launch worth tracking closely.

  • 337+ likes on the official MiniMaxAI Space.
  • First open-weight model to rank #1 on the Artificial Analysis video-editing leaderboard.
  • Covered by Forbes and MarkTechPost.
  • 33-billion-parameter model generating up to 15-second, 2K, 24fps video with native 32kHz stereo audio in one pass.
  • 9 active community discussion threads, including feature requests for a video/audio API gateway.

feedbacks

What People Are Saying

  • "jy-video-gateway"Hugging Face discussion

  • "Audio input model"Hugging Face discussion

  • "An Omni-Modal Video Model That Generates 15-Second 2K Clips With Native Stereo Audio"MarkTechPost

  • "placed first for video editing on the Artificial Analysis leaderboard, the first open-weight model to top an AI video ranking"industry coverage roundup

  • "MiniMax's H3 video model shipped as open weights, but its license excludes several major markets"tech coverage

  • "6-step MiniMax H3 Turbo Generation with Maestro!"X post

HF Spaces335 likes

Microduck Sandbox

Real trained RL policies for the Microduck robot, running fully in the browser: MuJoCo compiled to WebAssembly steps the physics, onnxruntime-web runs the policy network at 50 Hz. No server, no backend. Two locomotion variants of the same robot are included: legs (walking, the default) and rollers (the wheeled skating variant). Press M (or hold D-pad up ~1 s on a gamepad) to switch; the roller model, meshes and policies are lazy-loaded on the first switch. | Mode | Checkpoint | What it does | |--------|-----------|--------------| | Run (legs) | BESTalphawalking.onnx | Velocity-tracking locomotion (arrows / WASD to steer) | | Sit | BESTalphasitstand.onnx | Sits down on its hull, stands back u...

github
It rolls. Put wheels on, hold D-pad up, and it loads the other brain.

Market Signal

Why It Has Market Pull

Microduck Sandbox is an official Pollen Robotics demo that runs the real reinforcement-learning locomotion policies for its $399 open-source Microduck biped fully in-browser, no install required. It shows genuine hands-on builder engagement rather than passive traffic, backed by a credible robotics company with a real hardware product behind it.

  • 338 likes on Hugging Face.
  • 6 active community discussions, including a user-built 4-duck racing mod and a keyboard-control bug fix.
  • Built by Pollen Robotics, which co-launched the $399 open-source Microduck robot with Hugging Face in August 2026.
  • Runs real trained ONNX policies client-side via MuJoCo-WASM plus onnxruntime-web at 50Hz.
  • Full sim-to-real training stack open-sourced separately under Apache-2.0.

feedbacks

What People Are Saying

  • "Modified version of Microduck Simulator for competitive onnx skating policy. Environment features a living room scene with simple waypoint circuit and 4 ducks racing around it."Hugging Face discussion

  • "Fixes title page kick controls. Currently listed as A and E but actually uses Q and E"Hugging Face discussion

  • "Show layout-aware keyboard labels"Hugging Face discussion

  • "Add click-to-walk waypoint source."Hugging Face discussion

  • "boot failed SyntaxError"Hugging Face discussion

  • "Hugging Face Unveils Microduck: A $399 Open-Source 25cm Biped You Train with Reinforcement Learning"press coverage

Sources

GitHub

affaan-m/ECC

246.1K

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

mattpocock/skills

245.0K

Skills for Real Engineers. Straight from my .agents directory.

DietrichGebert/ponytail

121.1K

Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.

JuliusBrussee/caveman

102.5K

🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman

Product Hunt

Computable GPU Index (CGI)

402

Computable GPU Index (CGI) is a USD price per GPU-hour, computed from the published on-demand rental rates of a fixed panel of providers. The methodology is mathematically robust, and anyone can verify and reproduce every print.

Gauth AI Course

253

Gauth’s AI Course turns any subject into an interactive course, starting with 200+ AI math courses covering every US high school topic, from Algebra I through AP Calculus. The AI Course includes visuals, mindmaps, active-recall quizzes built in and enables you to pause any lesson to ask the AI Tutor a question. You can also generate your own course in seconds, to create personalized lessons at your own pace and share it with others.

TrustedRouter

147

One OpenAI-compatible API for hundreds of models on attested infrastructure, end to end encryption, with Confidential routes, ZDR routes, provider failover, BYOK, and no prompt or output logs.

EAS Observe

134

Observe measures how fast your Expo app starts and how fast each screen becomes usable, on real user devices. Install the library, wrap your root layout, and startup metrics arrive with nothing to instrument. Every native build and every EAS Update gets its own marker on the chart, so you can see which release moved the line. Full device context on every session: model, OS, country, network type, thermal state, frame data. Free plan covers 100K events a month.

Nodeterm

126

A node-based free open source terminal manager for macOS, Linux and Windows. Real terminals and coding agents as live nodes on an infinite canvas — watch Claude, Codex and Gemini work, wire them together, and never lose a session to a reboot.

Murmell

106

Our first Product Hunt launch brought many new users to Murmell and showed us that the infrastructure was not ready to scale. So we rebuilt it from scratch. We also created unified agent memory: close your laptop, freeze a canvas, and return to the exact same workspace, terminals, files and conversations. Your agents retain their context and continue where they stopped, while teammates can join the same live environment and collaborate in real time.

YC Launch

Hacker News

Aura – a Rust agent that investigates and fixes production incidents

16 pts

We run a SaaS that handles petabytes of data. Our SRE team experimented with using claude, openclaw, langchain, etc. within our incident response workflows. We struggled with overflowing context, lethal trifecta vectors, hallucinations, and burned a lot of frontier tokens mostly on easy work. Approval fatigue was a challenge, and we drew a hard line at relaxing permissions in production. Long story short, we built and open-sourced AURA, a Rust-based harness specifically designed for the type of... (16 points, 2 comments).

OpenInstinct – open-source, self-hostable Instinct clone

5 pts

We love Instinct, but have been increasingly worried about the data footprint we are handing over to them, and what they might do with that data. So we built an oss self-hostable version. It includes a vault that can store cards, logins, and personal information, so it can execute complex tasks on your behalf, like: "Get me two tickets to the odyssey on saturday at my nearest theatre" "Find me the best golf grip trainer and order it for me" "Read my email and find opportunities to save money by... (5 points, 0 comments).

Flawd is mutation testing for the AI era

4 pts

Hey, all! Flawd is a mutation testing tool that can target five languages (Python, JS, TS, Go, Rust) and it runs locally as a single binary on your machine. Importantly (for many), your code never leaves your machine and all processing takes place directly on your dev box or CI runner. Mutation testing is a process in which faults (known as mutants) are intentionally injected into your codebase, tests are run and any faults which are not detected by your tests are known as survivors (or survivin... (4 points, 5 comments).

Kit. Claude Code but Concise

10 pts

Kit is a coding agent runtime. It gives the model one tool for building and running programs. It provides a terminal client, and Agent Client Protocol (ACP) server, and A2A endpoint and a subagent orchestrator in one static binary. (10 points, 1 comments).

ShevtoneAudio Orchestrator – Turning MIDI into Full Orchestration

6 pts

Hi HN, I'm Bor Shev, a composer and developer. Over the past two years I've been developing ShevtoneAudio Orchestrator. The idea is simple: instead of generating a finished piece of music and replacing the composer, Orchestrator takes the composer's own MIDI and develops it into a full orchestration. It analyzes the musical material — harmony, melody, rhythm, dynamics, structure and orchestral density — and creates an arrangement across strings, brass, percussion and other sections. The importan... (6 points, 4 comments).

HF Spaces

Qwen-Image-Edit-2511-LoRAs-Fast

2.7K likes

Demo of the Collection of Qwen Image Edit LoRAs Qwen-Image-Edit-2511-LoRAs-Fast is a Hugging Face Space tagged with gradio, mcp-server, region:us. It has 2704 likes on Hugging Face.

SenseNova-U1.5-8B-MoT

104 likes

Unified text-to-image and image editing model Text-to-image and image editing demo for sensenova/SenseNova-U1.5-8B-MoT, a natively unified multimodal model (18B params, bf16) built on the NEO-unify architecture. Leave the image upload empty for text-to-image generation, or upload one or more images and write an edit instruction for image editing. Advanced options exposes denoising steps, guidance scale, timestep shift, image guidance (editing) and the seed. The step count defaults to 28 rather than the model card's 50: a fixed-seed A/B found 28 keeps composition, prompt adherence and text rendering intact — losing only some micro-texture in landscape and skin, and nothing measurable when edi...

Breeze TTS 2

43 likes

Bilingual TTS with voice design, cloning, and direction Open-weight bilingual (English/Chinese) text-to-speech model with: Voice Design — Create a voice from a natural-language description. Voice Clone — Clone a speaker from reference audio and its exact transcript. Voice Direction — Clone a reference voice while steering tone, emotion, and pace. Supports inline vocal events: (laugh), (sigh), (cough), (clears throat) in English; [笑], [叹气], [咳嗽], [清嗓子] in Chinese. Weights: BreezeBlue/Breeze-TTS-2 Code: breezeblue-ai/breeze-tts (Apache-2.0) License: Model weights and outputs are for research and non-commercial use only. Breeze TTS 2 is a Hugging Face Space tagged with gradio, mcp-server, regio...

H3 Acceleration Arena

41 likes

Blind A/B ranking of MiniMax-H3 acceleration variants Human-judged ranking of ~26 MiniMax-H3 acceleration variants over a 200-prompt corpus, from blind pairwise votes on pre-generated clips, with confidence intervals, cost and slice breakdowns. The design and its reasoning are in arena/DESIGN.md; the app's own notes are in arena/README.md. This Space is private and must stay private until deliberately flipped. It streams ~3,700 clips out of the private dataset multimodalart/h3-pre-gen-arena. See Going public below. hfoauth: true above creates the OAuth app and injects OAUTHCLIENTID, OAUTHCLIENTSECRET, OAUTHSCOPES and OPENIDPROVIDERURL. arena/space_auth.py implements the flow by hand (this is...

Wan2.2 14B Fast Preview

1.7K likes

generate a video from an image with a text prompt Wan2.2 14B Fast Preview is a Hugging Face Space tagged with gradio, mcp-server, region:us. It has 1683 likes on Hugging Face.

MiniMax-H3 Ultra Fast

193 likes

Ultra-fast local NVFP4 video + synchronized audio generation MiniMax-H3 Ultra Fast — local conditioner + pruned NVFP4 on Blackwell Joint video and synchronized sound from MiniMax-H3, rebuilt for a single 96 GB Blackwell ZeroGPU worker. | layer | optimization | |---|---| | Weights | 12.5 GB pruned NVFP4 transformer: 20.1B effective parameters instead of 33.1B/61.7 GiB BF16. | | Compute | Native CUDA 13 NVFP4 tensor-core GEMMs through comfy-kitchen; higher-precision norms, embeddings and output heads. | | Residency | Transformer, conditioner and both VAEs remain GPU-resident during generation—no layerwise CPU offload. | | Conditioner | Local 15.7 GB Qwen3-VL NVFP4-AWQ checkpoint containing onl...