New AI Tools Weekly #17: AI tools audit what agents install, write, and reach (2026)

This issue covers AI tools that launched or gained traction between September 14 and September 21, 2026. The loudest signal in that window is not a new model or a new agent. It is a set of releases that decide what an agent may install, what it may publish, and which parts of your machine it may touch.

GitHub Trending supplied the momentum signal, through weekly star counts across all languages, Python and TypeScript. 1 Product Hunt supplied the launch signal, and its weekly leaderboard for the week of September 14–20 was readable this week, so the ranks below are the board's own positions rather than a same-day snapshot. 2 That board closes on September 20, one day before this issue's window ends; the two launches that arrived after it closed are labelled as launch-day entries from the September 21 snapshot instead. A supplementary public search on X returned event promotion and general commentary, but no unverified launch worth carrying over a first-party repository or product page, so nothing from that source is included.

The 18 tools below sit in seven themes. Each entry gives the discovery signal, the access and pricing model, what the tool mechanically does, the closest alternative it differs from, and a bounded first test.

At a glance

Auditing the agent skill supply chain — agent-skills (Tech Leads Club), Claude-Red
Output discipline: the anti-slop stack — Humanizer, No AI Slop, i-have-adhd
Repository-scale review and specification — Open Code Review, OpenSpec, Hyrax AI
Local-first voice, music and video — Voicebox, YuE2, HyperFrames
Agents in a real session, with a record — BrowserSkill, Atlas
Driving agent fleets from somewhere other than the desk — Kilo Code mobile, Superset Mobile, Weave Router 2.0
The commercial layer under AI products — tiun., CREEM 2.0

Auditing the agent skill supply chain

An agent skill is an installable package. It arrives from a public marketplace, it lands in a directory your agent reads on every turn, and until recently almost nothing inspected it first. Two releases this week approach that from opposite ends: a registry that vets what it publishes, and a red-team library that shows what an unvetted skill directory can look like.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
agent-skills (Tech Leads Club)	GitHub Trending weekly: +1,161 stars (6.6k total). 3	MIT licence. Free, with no paid tier, published on npm as `@tech-leads-club/agent-skills` and usable either through `npx` or a global install. 3	A curated skill registry and CLI for AI coding agents. The catalogue ships no binaries, every skill is scanned with Snyk Agent Scan before publication, prompts are human-curated, and the CLI adds sanitisation, path isolation, symlink guards, an atomic lockfile and an operation audit log. The README cites a Snyk scan finding critical vulnerabilities in more than 13% of marketplace skills as the reason the layer exists. 3	Open marketplaces publish whatever a contributor uploads and rely on the user to judge it afterwards. This registry makes review a gate before publication, and its CLI keeps a local audit trail of what your agent installed, which is the record you need when a skill later misbehaves. 3	Run `npx @tech-leads-club/agent-skills` and install one skill into a scratch project. Then run `agent-skills audit` and confirm the install is logged. If you cannot find what your agent installed, the audit trail has not earned its place yet.
Claude-Red	GitHub Trending weekly: +2,649 stars (6.6k total). 4	MIT licence. Free; the install is a clone into the agent's skills directory, with no service or account behind it. 4	A curated library of 78 offensive-security skills across 23 categories. Each skill is a structured `SKILL.md` file that primes the model with expert-level methodology for one attack surface, from SQL injection to shellcode and EDR evasion. The agent auto-loads a matching skill when a conversation triggers it, so mentioning SQL injection pulls in the relevant playbook. 4	Generic red-team prompt collections leave the agent to improvise method. This library encodes the method per attack surface and lets the agent select it by trigger, which is the same mechanism a malicious skill would use — the clearest available demonstration of why a skills directory is an attack surface. 4	Clone it into a throwaway container with no credentials mounted, then ask the agent to plan a SQL injection test against a deliberately vulnerable local target. You are testing whether the skill fires and how much surrounding context it pulls in — not whether the exploit works.

Take the registry when you want a vetting layer between your agent and public skill marketplaces. Take Claude-Red when you are the one doing the offensive testing, and treat its existence as the argument for the vetting layer.

Output discipline: the anti-slop stack

Three repositories attacked the same problem this week from three different angles: model-written prose has a register readers have learned to distrust, and coding agents explain before they answer. None of these is a model. All three are Markdown skills that change how an agent you already run behaves, and all three spread through the same skills CLI.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
Humanizer (blader)	GitHub Trending weekly: +3,045 stars (50.8k total). 5	MIT licence. Free; `npx skills add blader/humanizer --global`, after which the skill answers to `/humanizer`. It also installs as a Claude Code plugin or a Claude Desktop skill. 5	Rewrites AI-sounding text until it reads like a person wrote it, without changing what the text says. Because the whole skill is Markdown, it works with any agent that supports skills. 5	Detectors score a passage and leave the editing to you. Humanizer performs the edit, which makes the interesting part of the test the diff rather than the score. 5	Paste a paragraph you know a model wrote into `/humanizer` and read the diff line by line. The test fails if any factual claim, number or qualifier changed.
No AI Slop (petergyang)	GitHub Trending weekly: +1,819 stars (10.9k total). 6	MIT licence. Free; `npx skills add petergyang/no-ai-slop --skill no-ai-slop --global --yes`. 6	Removes more than 20 named AI writing patterns — the "It's not X. It's Y." construction, "What nobody tells you is…", "The future isn't coming. It's already here." — while trying not to flatten the writer's own vocabulary, cadence and humour. 6	Humanizer targets the general texture of model prose. This skill targets a written list of named constructions, which makes its work auditable: you can check each listed pattern off rather than judging the result by feel. 6	Run it over something you wrote yourself, not over model output. The skill has to remove the listed patterns without stripping the quirks you put there on purpose, and that is the failure worth watching for.
i-have-adhd (ayghri)	GitHub Trending weekly: +5,249 stars (49.4k total). 7	MIT licence. Free; installed as a plugin or skill for a Claude-style agent, following the repository's install instructions. No diagnosis is needed to use it. 7	Stops a coding assistant from burying its answer. The skill enforces action-first output: the command to run, the file and line to change, numbered steps, and no closing pleasantries. The repository's own before-and-after shows a three-paragraph explanation of an outdated token library collapsing into one install command, a numbered three-step edit, and a single line telling you what to paste back if a test fails. 7	The usual brevity prompt asks the agent to sound terser. This skill changes the order of the answer instead, so the instruction you act on arrives before the reasoning that justifies it. 7	Give the agent a task you have already solved by hand. Read only the first line of the reply and check whether it alone was enough to act on.

Humanizer and No AI Slop both edit text you already have, and they overlap enough that running both on one draft is waste. Pick the pattern list when you want a checklist and the general rewriter when you want the texture changed. i-have-adhd is the odd one out: it changes the shape of an answer rather than the writing in it, so it belongs on the agent you work with, not on the document you ship.

Repository-scale review and specification

Agent-written code arrives faster than human review can absorb it. Three tools this week sit at three points along that line: one reviews a diff, one audits and repairs a whole repository, and one makes the agent agree on a specification before it writes anything at all.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
Open Code Review (Alibaba)	GitHub Trending weekly: +15,504 stars (38.7k total) — first place on the all-language weekly board. 8	Apache-2.0. Free and self-hosted: `npm install -g @alibaba-group/open-code-review`, then point it at an OpenAI- or Anthropic-compatible model endpoint. Requires Git 2.41 or newer. 8	A code-review CLI with a hybrid design: a deterministic pipeline decides which files to review and how to group them, then an LLM agent reviews each bundle. Related files are bundled into one review unit and each bundle runs as a sub-agent with isolated context, which keeps very large changesets tractable. Separate positioning and reflection modules check that every comment points at the right line. The `ocr scan` command reviews whole files where there is no diff to read. 8	General-purpose agents reviewing code cut corners on big changesets and drift on line numbers, because the method lives in natural-language instructions. Alibaba's own benchmark — 50 repositories, 200 real pull requests, 10 languages, 1,505 annotated defects — reports higher precision and F1 than a general-purpose agent on the same model using roughly one ninth of the tokens, with lower recall as a deliberate trade toward fewer false alarms. Those figures are the vendor's, not an independent result. 8	Run it on a pull request you have already reviewed by hand, using the same model you would have given a general-purpose agent. The interesting number is how many comments you have to dismiss, not how many it produced.
OpenSpec (Fission)	GitHub Trending weekly: +1,593 stars (69.7k total). 9	MIT licence. Free; `npm install -g @fission-ai/openspec@latest`, which needs Node.js 20.19 or newer. 9	Adds a lightweight specification layer so a human and an agent agree on what is being built before code exists. A new beta feature, Stores, keeps those specs in a repository of their own, so one plan can span several code repositories and a platform team can own specs that product teams read. 9	Requirements that live only in a chat history are unreviewable and vanish with the session. OpenSpec makes the agreed plan a file that travels through git, which is a different claim from prompting the agent to plan more carefully. 9	Start a spec for a small feature on a project that already has one, then read the resulting change file and check whether it caught an error case you would have had to fix after the code was written.
Hyrax AI	Product Hunt launch snapshot, September 21: marked Launching today. 10	Free Options. The free plan is a mini audit, five findings and five fixes a month; paid tiers sit above it. 10	Connects to GitHub and audits the entire codebase across six engineering domains — security, correctness, performance, maintainability, architecture and operations. It builds context around the project's architecture and conventions, writes a failing test to prove a reported problem is real, then writes the fix and runs the repository's own tests, build and lint before opening a pull request for a human to review and merge. 10	Most AI review tools comment on the pull request in front of them and stop there. This one works on the repository behind the pull request and hands back a branch rather than an opinion. 10	Point it at a repository you own but do not depend on this week, and read the first proposed pull request before merging anything. The claim to check is the failing test: if the test does not actually reproduce the reported problem, the rest of the pipeline is decoration.

Open Code Review is the tool for teams that want review running inside CI on every changeset. Hyrax AI is for teams that want the backlog of existing repository debt worked down. OpenSpec is categorically different — it moves the agreement earlier, and it only pays off if the specification is something a human actually reads.

Local-first voice, music and video

The media cluster moved onto your own hardware this week. Two of these three run entirely on the machine and the third renders locally with no per-render fee. The trade-off is stated plainly in each repository: the music model wants a 24 GB NVIDIA GPU, and the voice studio wants an Apple Silicon or Intel machine, Windows, or Docker.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
Voicebox (jamiepine)	GitHub Trending weekly: +2,346 stars (55.3k total). 11	MIT licence. Free, with downloads for macOS (Apple Silicon and Intel), Windows and Docker from voicebox.sh. 11	A local-first voice studio. It clones a voice from a few seconds of reference audio, generates speech in 23 languages across seven TTS engines, dictates into any text field from a global hotkey, and gives an MCP-aware agent a voice of your choosing. A bundled local model refines output, and per-profile personas keep characters apart. Post-processing covers pitch, reverb, delay, chorus and compression, and the Chatterbox Turbo engine accepts paralinguistic tags such as `[laugh]` and `[sigh]`. 11	The two cloud incumbents each cover half of the loop — one sells speech out, the other sells dictation in. This one covers both in a single application and keeps models, reference audio and captures on your machine. 11	Clone a voice from a 10-second clip you own the rights to, then dictate one sentence into an unrelated app with the hotkey. Watch the first-run model download size before committing a laptop to it.
YuE2 (multimodal-art-projection)	GitHub Trending weekly: +2,427 stars (9.9k total). 12	Open source with a split licence. Code, agent skill and documentation are Apache-2.0; the model weights are CC BY-NC 4.0 with an added creator permission — free for personal, creator and musician use including monetising output, free for non-commercial academic use, and companies must negotiate a commercial licence. 12	Frontier-scale music generation. An autoregressive/non-autoregressive mixture-of-transformers predicts a symbolic score and semantic music tokens, flow matching generates acoustic latents, and a VAE decodes them into 48 kHz stereo. Creation, cover and edit modes differ only in where the score comes from — the model, a transcribed recording, or an edited composition — and the repository ships a `yue2-music` skill so an agent can drive it. 12	Most music generators are audio-only black boxes. Keeping the score as an intermediate artefact makes the composition editable and reusable, which is what allows the zero-shot cover and the agentic editing path at all. 12	Install it on a machine with 24 GB of VRAM and generate one short instrumental. Then check whether the retained score file alone is enough to reproduce the same piece, which is the whole promise of a symbolic stage.
HyperFrames (HeyGen)	GitHub Trending weekly: +2,546 stars (52k total). 13	Apache-2.0, with no per-render fees and no commercial-use threshold. Free; `npx hyperframes init my-video`, `npx hyperframes preview`, `npx hyperframes render`. 13	Renders video from plain HTML. A composition is an HTML file with data attributes; animation comes from GSAP, CSS, Lottie, Three.js, Anime.js or a custom runtime; and `index.html` plays in a browser with no build step. The same input produces the same frames, which is what makes it usable in CI. 13	Both this and Remotion render through headless Chrome and FFmpeg, and the repository says so; the difference is the authoring model. Remotion asks for React components, HyperFrames asks for HTML that an agent can already write, with a non-interactive CLI by default. 13	Build a ten-second composition, render it twice, and compare the frame checksums. If the output is not byte-stable, automatic rendering in a pipeline will produce surprises.

Voicebox and HyperFrames are the two that fit an ordinary workstation; YuE2 is the one that needs a GPU budget approved first. All three keep the intermediate representation — refined audio, HTML composition, symbolic score — rather than only the finished file, which is what makes each of them editable after the first render.

Letting an agent into a real session, and keeping the record

Giving an agent access to the live web usually means one of two bad options: hand it credentials it should not have, or teach a headless browser to fight through a login wall. Two tools this week take a different route — borrow the session you already have, and keep a written record of what came of it.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
BrowserSkill (Tencent)	GitHub Trending weekly: +4,077 stars (6.2k total). 14	MIT licence. Free. The agent-install path is a single instruction file the agent follows; manually you run the repository's `install.sh` with your shell to get the `bsk` CLI, then load the extension from the Chrome Web Store. Chrome and Edge are supported today and Firefox is planned. 14	Connects a shell-capable agent — the repository lists Cursor, Claude Code, Codex, OpenClaw, CodeBuddy and others — to your already signed-in browser. Browser work runs in a separate visible Agent Window so you keep using your own browser, and an agent that needs a tab you already have open must borrow that tab explicitly and hand it back when the task ends. 14	Headless frameworks such as Playwright start from no session and need separate test accounts. This one starts from a real login and turns tab access into an explicit, returnable loan, which is the part that makes it auditable after the fact. 14	Either install it, or — cheaper — read `AGENT_INSTALL.md` first and check what the agent is being told to do to your machine. Then have it read a page behind a login you are already signed into and confirm the rest of your browsing is undisturbed.
Atlas (pacifio)	GitHub Trending weekly: +1,556 stars (4.7k total). 15	Apache-2.0. Free desktop build: a `.dmg` for macOS 13 or newer, an `.msi` for Windows 10 or newer. Linux builds from the same Tauri codebase but is untested, and the project is self-hostable. 15	Source control for coding agents. Atlas records which agent session produced which commit, keeps the reasoning alongside the change, and lets you query changes across agents in one place instead of switching tools. Notes are Markdown, canvases are JSON, sessions are JSONL, and the editor itself is a file on disk, so nothing is locked in. A transcript import backfills existing Claude Code history. 15	Version control records the diff; the prompt, the abandoned attempts and the tool calls are gone the moment the scrollback clears. Atlas keeps that second record and attributes each commit to the session that made it. 15	Run two different agents against one repository, then ask Atlas which session produced which commit. If it cannot answer that, the attribution is not working and nothing else in the tool matters.

BrowserSkill widens what an agent can reach, and Atlas records what it did when it got there. Teams that adopt the first should look hard at the second, because "the agent did something in a logged-in session" is the hardest incident to reconstruct after the fact.

Driving agent fleets from somewhere other than the desk

Once a single agent run takes twenty minutes to an hour, supervising it at the desk stops making sense. Three entries near the top of this week's Product Hunt board give three versions of the same answer: move the console to your phone, or route the work to whichever subscription still has quota left.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
Kilo Code for iOS and Android	Product Hunt weekly leaderboard, September 14–20: rank 6, tagged Android, iOS and Developer Tools. 2	Free tier — the sign-up page offers "Kilo Auto Free" with no credit card. The underlying agent is MIT-licensed open source, and the gateway passes 500+ models through at zero markup, with bring-your-own keys and local models supported. 16	Extends an existing coding agent to a phone: start agent sessions, control them, and review pull requests from a mobile device. The same agent already runs in VS Code, JetBrains IDEs, a CLI, cloud agents and Slack, and specialised modes cover ordinary coding, architectural planning, debugging and code review. 16	A mobile companion usually means a thin notification client. Here the phone is another surface onto the same sessions, so a review started at the desk can be finished in a queue without restarting the agent. 16	Start a long agent task at the desk, then open the same session on your phone and merge or reject one proposed change. The test is whether the diff is genuinely reviewable on a small screen or only visible on it.
Superset Mobile	Product Hunt launch snapshot, September 21: rank 1 on the board, marked Launching today. 17	Free Options — the app is included with Superset Pro, and the launch offers the first month free with the code MOBILELAUNCH. 17	Orchestrates coding agents from an iPhone. The mobile app picks up the same workspaces as the desktop application, so a task started on the computer can be reviewed, diffed and merged away from it. It works with Claude Code, Codex and other terminal agents. 17	The desktop product is built around running many agents at once with each task isolated in its own sandbox. The mobile client inherits that same workspace instead of offering a separate, thinner queue. 17	With one agent already running, open the mobile app and check whether the workspace listing matches the desktop. If you have to re-establish the session on the phone, the mobile client is a viewer rather than a control surface.
Weave Router 2.0	Product Hunt weekly leaderboard, September 14–20: rank 12, tagged Open Source, Developer Tools and Artificial Intelligence. 18	Free Options. 18	A subscription-aware router for coding agents. It runs Claude models inside Codex and GPT models inside Claude Code on plans you already hold, sending each request to whichever subscription still has quota. A classifier scores the task's complexity and switches models only when the saving exceeds the cost of rebuilding the cache. 18	Most routers optimise for price per token against a single provider. This one treats subscription entitlement — which plan still has headroom, not which endpoint is cheapest — as the routing constraint, and adds quota arbitrage between providers on top. Weave says Router 2.0 matches its comparison model on Terminal-Bench 4.0 and SWE-Atlas at half the cost and twice the speed; those are the vendor's numbers, not an independent result. 18	Install the router for one week and log which provider each request went to. The real question is how often it moved a request to a weaker model to save quota, because that is the cost this design hides best.

Kilo Code and Superset both sell the same basic promise — your agent sessions, reachable from the phone — and differ in whether the phone sits on top of an open-source agent or a managed one. Weave Router 2.0 is not a console at all; it is the thing you install when the constraint is subscription quota rather than model quality.

The commercial layer under AI products

The two highest-ranked launches on the September 14–20 Product Hunt board were not models or agents. Both sell the parts a builder needs after the product works — checkout, tax, metering, customer data — and both pitch themselves at people shipping AI products specifically.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
tiun.	Product Hunt weekly leaderboard, September 14–20: rank 1. 2	Free. 19	One system for authentication, payments, customer data and analytics, installed with a single command and aimed at letting a builder launch a paid product the same day they start building. 19	The pitch is consolidation rather than any single capability. The alternative is composing separate services for auth, billing and analytics, which the launch page frames as more dashboards, more integrations and more places to break. 19	Install it into a throwaway project and create one paid product with one price. Then check how hard it is to export the customer records, because consolidation is exactly the trade you are making.
CREEM 2.0	Product Hunt weekly leaderboard, September 14–20: rank 2. 2	Paid. CREEM acts as merchant of record and charges platform fees on sales; sign-up is through creem.io. 20	Handles checkout, tax, invoices, fraud and chargebacks for software and digital products, and adds affiliates, revenue splits and payouts on the same balance. The billing model is built for how AI products charge — prepaid credit packs drained per generation, call or minute, usage-based prices with a free allowance and a cap, and free trials. Everything in the dashboard is scriptable through an MCP server, an API and a CLI. 20	Stripe and Paddle also sell merchant-of-record billing, and both can express usage pricing. The difference here is the shape of the meter: credit wallets and per-generation billing are first-class objects rather than configurations on top of a generic subscription, and the agent-facing surface means the same account can be driven from inside an editor. 20	Set up one credit-pack product and one metered product in a sandbox, then read a single month's invoice. If tax and the meter appear correctly on the invoice, the merchant-of-record claim is doing real work; if you have to reconcile them by hand, it is not.

Both sell the same underlying promise, which is that a founder should never touch tax logic. tiun. consolidates the widest surface including authentication and keeps the entry price at zero; CREEM 2.0 goes deeper on the metering that AI products specifically need. Neither is free to leave once you have customers in it, which is the cost both launch pages are quiet about.

Editorial view: what to test first

Two patterns run through this week's releases, and they pull in opposite directions.

The first is inspection. Four of the seven themes are about putting something between an agent and the world it touches: a registry that vets the skills it installs, three skills that police the prose and the answer shape it produces, a review pipeline that checks whether its comments point at the right lines, and a session record that says which agent made which commit. The reasoning chain is visible in the repositories themselves — the registry cites a scan finding critical vulnerabilities in more than 13% of marketplace skills, and the review tool publishes a defect benchmark rather than a demo. The direction of travel is clear. Whether any of these checks catches anything real in your repository is a question only your own first test answers.

The second is the opposite: agents leaving the desk. Onto a phone, into a logged-in browser, on top of a routing layer that decides which subscription pays. These are convenience and cost releases rather than capability releases, and they are the ones that quietly widen what an agent can reach.

If you test one thing this week, pick by the friction you can actually name:

If your agent's skills directory has grown without review: install the agent-skills CLI, then run agent-skills audit and find out what is already in there before adding anything else.
If your team ships AI-written prose to customers: add No AI Slop for the named patterns and Humanizer for the texture — one or the other, not both on the same draft.
If code review is the bottleneck rather than code generation: put Open Code Review in CI on your noisiest repository and count dismissed comments, not found issues.
If your repository has accumulated debt no one has triaged: give Hyrax AI a repository you own but do not depend on this week, and merge nothing until a proposed failing test actually reproduces its problem.
If you are shipping audio and cannot justify per-generation pricing: try Voicebox, and check the model download size against the laptop first.
If your agents already run for an hour at a time: open the same session in Kilo Code or Superset Mobile and see whether the diff is genuinely reviewable on a phone.
If you scroll back through terminal history to remember why a commit exists: the commit message is the only thing that survives today, and Atlas is the tool that assumes it should not be.

The check tools and the reach tools fail in the same way: they only help if you read what they produce. A registry you do not audit, a review pipeline whose comments you dismiss without reading, and a session log nobody opens are three ways of paying for reassurance. Test the one whose output you will actually read next week.

References

1
GitHub Trending, weekly
github.com
2
Product Hunt weekly leaderboard, September 14–20, 2026
producthunt.com
3
agent-skills repository
github.com
4
Claude-Red repository
github.com
5
Humanizer repository
github.com
6
No AI Slop repository
github.com
7
i-have-adhd repository
github.com
8
Open Code Review repository
github.com
9
OpenSpec repository
github.com
10
Hyrax AI on Product Hunt
producthunt.com
11
Voicebox repository
github.com
12
YuE2 repository
github.com
13
HyperFrames repository
github.com
14
BrowserSkill repository
github.com
15
Atlas repository
github.com
16
Kilo Code official site
kilocode.ai
17
Superset Mobile on Product Hunt
producthunt.com
18
Weave Router 2.0 on Product Hunt
producthunt.com
19
tiun. on Product Hunt
producthunt.com
20
CREEM official site
creem.io

At a glance

Auditing the agent skill supply chain — agent-skills (Tech Leads Club), Claude-Red
Output discipline: the anti-slop stack — Humanizer, No AI Slop, i-have-adhd
Repository-scale review and specification — Open Code Review, OpenSpec, Hyrax AI
Local-first voice, music and video — Voicebox, YuE2, HyperFrames
Agents in a real session, with a record — BrowserSkill, Atlas
Driving agent fleets from somewhere other than the desk — Kilo Code mobile, Superset Mobile, Weave Router 2.0
The commercial layer under AI products — tiun., CREEM 2.0

Auditing the agent skill supply chain

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
agent-skills (Tech Leads Club)	GitHub Trending weekly: +1,161 stars (6.6k total). 3	MIT licence. Free, with no paid tier, published on npm as `@tech-leads-club/agent-skills` and usable either through `npx` or a global install. 3	A curated skill registry and CLI for AI coding agents. The catalogue ships no binaries, every skill is scanned with Snyk Agent Scan before publication, prompts are human-curated, and the CLI adds sanitisation, path isolation, symlink guards, an atomic lockfile and an operation audit log. The README cites a Snyk scan finding critical vulnerabilities in more than 13% of marketplace skills as the reason the layer exists. 3	Open marketplaces publish whatever a contributor uploads and rely on the user to judge it afterwards. This registry makes review a gate before publication, and its CLI keeps a local audit trail of what your agent installed, which is the record you need when a skill later misbehaves. 3	Run `npx @tech-leads-club/agent-skills` and install one skill into a scratch project. Then run `agent-skills audit` and confirm the install is logged. If you cannot find what your agent installed, the audit trail has not earned its place yet.
Claude-Red	GitHub Trending weekly: +2,649 stars (6.6k total). 4	MIT licence. Free; the install is a clone into the agent's skills directory, with no service or account behind it. 4	A curated library of 78 offensive-security skills across 23 categories. Each skill is a structured `SKILL.md` file that primes the model with expert-level methodology for one attack surface, from SQL injection to shellcode and EDR evasion. The agent auto-loads a matching skill when a conversation triggers it, so mentioning SQL injection pulls in the relevant playbook. 4	Generic red-team prompt collections leave the agent to improvise method. This library encodes the method per attack surface and lets the agent select it by trigger, which is the same mechanism a malicious skill would use — the clearest available demonstration of why a skills directory is an attack surface. 4	Clone it into a throwaway container with no credentials mounted, then ask the agent to plan a SQL injection test against a deliberately vulnerable local target. You are testing whether the skill fires and how much surrounding context it pulls in — not whether the exploit works.

Output discipline: the anti-slop stack

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
Humanizer (blader)	GitHub Trending weekly: +3,045 stars (50.8k total). 5	MIT licence. Free; `npx skills add blader/humanizer --global`, after which the skill answers to `/humanizer`. It also installs as a Claude Code plugin or a Claude Desktop skill. 5	Rewrites AI-sounding text until it reads like a person wrote it, without changing what the text says. Because the whole skill is Markdown, it works with any agent that supports skills. 5	Detectors score a passage and leave the editing to you. Humanizer performs the edit, which makes the interesting part of the test the diff rather than the score. 5	Paste a paragraph you know a model wrote into `/humanizer` and read the diff line by line. The test fails if any factual claim, number or qualifier changed.
No AI Slop (petergyang)	GitHub Trending weekly: +1,819 stars (10.9k total). 6	MIT licence. Free; `npx skills add petergyang/no-ai-slop --skill no-ai-slop --global --yes`. 6	Removes more than 20 named AI writing patterns — the "It's not X. It's Y." construction, "What nobody tells you is…", "The future isn't coming. It's already here." — while trying not to flatten the writer's own vocabulary, cadence and humour. 6	Humanizer targets the general texture of model prose. This skill targets a written list of named constructions, which makes its work auditable: you can check each listed pattern off rather than judging the result by feel. 6	Run it over something you wrote yourself, not over model output. The skill has to remove the listed patterns without stripping the quirks you put there on purpose, and that is the failure worth watching for.
i-have-adhd (ayghri)	GitHub Trending weekly: +5,249 stars (49.4k total). 7	MIT licence. Free; installed as a plugin or skill for a Claude-style agent, following the repository's install instructions. No diagnosis is needed to use it. 7	Stops a coding assistant from burying its answer. The skill enforces action-first output: the command to run, the file and line to change, numbered steps, and no closing pleasantries. The repository's own before-and-after shows a three-paragraph explanation of an outdated token library collapsing into one install command, a numbered three-step edit, and a single line telling you what to paste back if a test fails. 7	The usual brevity prompt asks the agent to sound terser. This skill changes the order of the answer instead, so the instruction you act on arrives before the reasoning that justifies it. 7	Give the agent a task you have already solved by hand. Read only the first line of the reply and check whether it alone was enough to act on.

Repository-scale review and specification

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
Open Code Review (Alibaba)	GitHub Trending weekly: +15,504 stars (38.7k total) — first place on the all-language weekly board. 8	Apache-2.0. Free and self-hosted: `npm install -g @alibaba-group/open-code-review`, then point it at an OpenAI- or Anthropic-compatible model endpoint. Requires Git 2.41 or newer. 8	A code-review CLI with a hybrid design: a deterministic pipeline decides which files to review and how to group them, then an LLM agent reviews each bundle. Related files are bundled into one review unit and each bundle runs as a sub-agent with isolated context, which keeps very large changesets tractable. Separate positioning and reflection modules check that every comment points at the right line. The `ocr scan` command reviews whole files where there is no diff to read. 8	General-purpose agents reviewing code cut corners on big changesets and drift on line numbers, because the method lives in natural-language instructions. Alibaba's own benchmark — 50 repositories, 200 real pull requests, 10 languages, 1,505 annotated defects — reports higher precision and F1 than a general-purpose agent on the same model using roughly one ninth of the tokens, with lower recall as a deliberate trade toward fewer false alarms. Those figures are the vendor's, not an independent result. 8	Run it on a pull request you have already reviewed by hand, using the same model you would have given a general-purpose agent. The interesting number is how many comments you have to dismiss, not how many it produced.
OpenSpec (Fission)	GitHub Trending weekly: +1,593 stars (69.7k total). 9	MIT licence. Free; `npm install -g @fission-ai/openspec@latest`, which needs Node.js 20.19 or newer. 9	Adds a lightweight specification layer so a human and an agent agree on what is being built before code exists. A new beta feature, Stores, keeps those specs in a repository of their own, so one plan can span several code repositories and a platform team can own specs that product teams read. 9	Requirements that live only in a chat history are unreviewable and vanish with the session. OpenSpec makes the agreed plan a file that travels through git, which is a different claim from prompting the agent to plan more carefully. 9	Start a spec for a small feature on a project that already has one, then read the resulting change file and check whether it caught an error case you would have had to fix after the code was written.
Hyrax AI	Product Hunt launch snapshot, September 21: marked Launching today. 10	Free Options. The free plan is a mini audit, five findings and five fixes a month; paid tiers sit above it. 10	Connects to GitHub and audits the entire codebase across six engineering domains — security, correctness, performance, maintainability, architecture and operations. It builds context around the project's architecture and conventions, writes a failing test to prove a reported problem is real, then writes the fix and runs the repository's own tests, build and lint before opening a pull request for a human to review and merge. 10	Most AI review tools comment on the pull request in front of them and stop there. This one works on the repository behind the pull request and hands back a branch rather than an opinion. 10	Point it at a repository you own but do not depend on this week, and read the first proposed pull request before merging anything. The claim to check is the failing test: if the test does not actually reproduce the reported problem, the rest of the pipeline is decoration.

Local-first voice, music and video

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
Voicebox (jamiepine)	GitHub Trending weekly: +2,346 stars (55.3k total). 11	MIT licence. Free, with downloads for macOS (Apple Silicon and Intel), Windows and Docker from voicebox.sh. 11	A local-first voice studio. It clones a voice from a few seconds of reference audio, generates speech in 23 languages across seven TTS engines, dictates into any text field from a global hotkey, and gives an MCP-aware agent a voice of your choosing. A bundled local model refines output, and per-profile personas keep characters apart. Post-processing covers pitch, reverb, delay, chorus and compression, and the Chatterbox Turbo engine accepts paralinguistic tags such as `[laugh]` and `[sigh]`. 11	The two cloud incumbents each cover half of the loop — one sells speech out, the other sells dictation in. This one covers both in a single application and keeps models, reference audio and captures on your machine. 11	Clone a voice from a 10-second clip you own the rights to, then dictate one sentence into an unrelated app with the hotkey. Watch the first-run model download size before committing a laptop to it.
YuE2 (multimodal-art-projection)	GitHub Trending weekly: +2,427 stars (9.9k total). 12	Open source with a split licence. Code, agent skill and documentation are Apache-2.0; the model weights are CC BY-NC 4.0 with an added creator permission — free for personal, creator and musician use including monetising output, free for non-commercial academic use, and companies must negotiate a commercial licence. 12	Frontier-scale music generation. An autoregressive/non-autoregressive mixture-of-transformers predicts a symbolic score and semantic music tokens, flow matching generates acoustic latents, and a VAE decodes them into 48 kHz stereo. Creation, cover and edit modes differ only in where the score comes from — the model, a transcribed recording, or an edited composition — and the repository ships a `yue2-music` skill so an agent can drive it. 12	Most music generators are audio-only black boxes. Keeping the score as an intermediate artefact makes the composition editable and reusable, which is what allows the zero-shot cover and the agentic editing path at all. 12	Install it on a machine with 24 GB of VRAM and generate one short instrumental. Then check whether the retained score file alone is enough to reproduce the same piece, which is the whole promise of a symbolic stage.
HyperFrames (HeyGen)	GitHub Trending weekly: +2,546 stars (52k total). 13	Apache-2.0, with no per-render fees and no commercial-use threshold. Free; `npx hyperframes init my-video`, `npx hyperframes preview`, `npx hyperframes render`. 13	Renders video from plain HTML. A composition is an HTML file with data attributes; animation comes from GSAP, CSS, Lottie, Three.js, Anime.js or a custom runtime; and `index.html` plays in a browser with no build step. The same input produces the same frames, which is what makes it usable in CI. 13	Both this and Remotion render through headless Chrome and FFmpeg, and the repository says so; the difference is the authoring model. Remotion asks for React components, HyperFrames asks for HTML that an agent can already write, with a non-interactive CLI by default. 13	Build a ten-second composition, render it twice, and compare the frame checksums. If the output is not byte-stable, automatic rendering in a pipeline will produce surprises.

Letting an agent into a real session, and keeping the record

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
BrowserSkill (Tencent)	GitHub Trending weekly: +4,077 stars (6.2k total). 14	MIT licence. Free. The agent-install path is a single instruction file the agent follows; manually you run the repository's `install.sh` with your shell to get the `bsk` CLI, then load the extension from the Chrome Web Store. Chrome and Edge are supported today and Firefox is planned. 14	Connects a shell-capable agent — the repository lists Cursor, Claude Code, Codex, OpenClaw, CodeBuddy and others — to your already signed-in browser. Browser work runs in a separate visible Agent Window so you keep using your own browser, and an agent that needs a tab you already have open must borrow that tab explicitly and hand it back when the task ends. 14	Headless frameworks such as Playwright start from no session and need separate test accounts. This one starts from a real login and turns tab access into an explicit, returnable loan, which is the part that makes it auditable after the fact. 14	Either install it, or — cheaper — read `AGENT_INSTALL.md` first and check what the agent is being told to do to your machine. Then have it read a page behind a login you are already signed into and confirm the rest of your browsing is undisturbed.
Atlas (pacifio)	GitHub Trending weekly: +1,556 stars (4.7k total). 15	Apache-2.0. Free desktop build: a `.dmg` for macOS 13 or newer, an `.msi` for Windows 10 or newer. Linux builds from the same Tauri codebase but is untested, and the project is self-hostable. 15	Source control for coding agents. Atlas records which agent session produced which commit, keeps the reasoning alongside the change, and lets you query changes across agents in one place instead of switching tools. Notes are Markdown, canvases are JSON, sessions are JSONL, and the editor itself is a file on disk, so nothing is locked in. A transcript import backfills existing Claude Code history. 15	Version control records the diff; the prompt, the abandoned attempts and the tool calls are gone the moment the scrollback clears. Atlas keeps that second record and attributes each commit to the session that made it. 15	Run two different agents against one repository, then ask Atlas which session produced which commit. If it cannot answer that, the attribution is not working and nothing else in the tool matters.

Driving agent fleets from somewhere other than the desk

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
Kilo Code for iOS and Android	Product Hunt weekly leaderboard, September 14–20: rank 6, tagged Android, iOS and Developer Tools. 2	Free tier — the sign-up page offers "Kilo Auto Free" with no credit card. The underlying agent is MIT-licensed open source, and the gateway passes 500+ models through at zero markup, with bring-your-own keys and local models supported. 16	Extends an existing coding agent to a phone: start agent sessions, control them, and review pull requests from a mobile device. The same agent already runs in VS Code, JetBrains IDEs, a CLI, cloud agents and Slack, and specialised modes cover ordinary coding, architectural planning, debugging and code review. 16	A mobile companion usually means a thin notification client. Here the phone is another surface onto the same sessions, so a review started at the desk can be finished in a queue without restarting the agent. 16	Start a long agent task at the desk, then open the same session on your phone and merge or reject one proposed change. The test is whether the diff is genuinely reviewable on a small screen or only visible on it.
Superset Mobile	Product Hunt launch snapshot, September 21: rank 1 on the board, marked Launching today. 17	Free Options — the app is included with Superset Pro, and the launch offers the first month free with the code MOBILELAUNCH. 17	Orchestrates coding agents from an iPhone. The mobile app picks up the same workspaces as the desktop application, so a task started on the computer can be reviewed, diffed and merged away from it. It works with Claude Code, Codex and other terminal agents. 17	The desktop product is built around running many agents at once with each task isolated in its own sandbox. The mobile client inherits that same workspace instead of offering a separate, thinner queue. 17	With one agent already running, open the mobile app and check whether the workspace listing matches the desktop. If you have to re-establish the session on the phone, the mobile client is a viewer rather than a control surface.
Weave Router 2.0	Product Hunt weekly leaderboard, September 14–20: rank 12, tagged Open Source, Developer Tools and Artificial Intelligence. 18	Free Options. 18	A subscription-aware router for coding agents. It runs Claude models inside Codex and GPT models inside Claude Code on plans you already hold, sending each request to whichever subscription still has quota. A classifier scores the task's complexity and switches models only when the saving exceeds the cost of rebuilding the cache. 18	Most routers optimise for price per token against a single provider. This one treats subscription entitlement — which plan still has headroom, not which endpoint is cheapest — as the routing constraint, and adds quota arbitrage between providers on top. Weave says Router 2.0 matches its comparison model on Terminal-Bench 4.0 and SWE-Atlas at half the cost and twice the speed; those are the vendor's numbers, not an independent result. 18	Install the router for one week and log which provider each request went to. The real question is how often it moved a request to a weaker model to save quota, because that is the cost this design hides best.

The commercial layer under AI products

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
tiun.	Product Hunt weekly leaderboard, September 14–20: rank 1. 2	Free. 19	One system for authentication, payments, customer data and analytics, installed with a single command and aimed at letting a builder launch a paid product the same day they start building. 19	The pitch is consolidation rather than any single capability. The alternative is composing separate services for auth, billing and analytics, which the launch page frames as more dashboards, more integrations and more places to break. 19	Install it into a throwaway project and create one paid product with one price. Then check how hard it is to export the customer records, because consolidation is exactly the trade you are making.
CREEM 2.0	Product Hunt weekly leaderboard, September 14–20: rank 2. 2	Paid. CREEM acts as merchant of record and charges platform fees on sales; sign-up is through creem.io. 20	Handles checkout, tax, invoices, fraud and chargebacks for software and digital products, and adds affiliates, revenue splits and payouts on the same balance. The billing model is built for how AI products charge — prepaid credit packs drained per generation, call or minute, usage-based prices with a free allowance and a cap, and free trials. Everything in the dashboard is scriptable through an MCP server, an API and a CLI. 20	Stripe and Paddle also sell merchant-of-record billing, and both can express usage pricing. The difference here is the shape of the meter: credit wallets and per-generation billing are first-class objects rather than configurations on top of a generic subscription, and the agent-facing surface means the same account can be driven from inside an editor. 20	Set up one credit-pack product and one metered product in a sandbox, then read a single month's invoice. If tax and the meter appear correctly on the invoice, the merchant-of-record claim is doing real work; if you have to reconcile them by hand, it is not.

Editorial view: what to test first

Two patterns run through this week's releases, and they pull in opposite directions.

If you test one thing this week, pick by the friction you can actually name:

If your agent's skills directory has grown without review: install the agent-skills CLI, then run agent-skills audit and find out what is already in there before adding anything else.
If your team ships AI-written prose to customers: add No AI Slop for the named patterns and Humanizer for the texture — one or the other, not both on the same draft.
If code review is the bottleneck rather than code generation: put Open Code Review in CI on your noisiest repository and count dismissed comments, not found issues.
If your repository has accumulated debt no one has triaged: give Hyrax AI a repository you own but do not depend on this week, and merge nothing until a proposed failing test actually reproduces its problem.
If you are shipping audio and cannot justify per-generation pricing: try Voicebox, and check the model download size against the laptop first.
If your agents already run for an hour at a time: open the same session in Kilo Code or Superset Mobile and see whether the diff is genuinely reviewable on a phone.
If you scroll back through terminal history to remember why a commit exists: the commit message is the only thing that survives today, and Atlas is the tool that assumes it should not be.

References

1
GitHub Trending, weekly
github.com
2
Product Hunt weekly leaderboard, September 14–20, 2026
producthunt.com
3
agent-skills repository
github.com
4
Claude-Red repository
github.com
5
Humanizer repository
github.com
6
No AI Slop repository
github.com
7
i-have-adhd repository
github.com
8
Open Code Review repository
github.com
9
OpenSpec repository
github.com
10
Hyrax AI on Product Hunt
producthunt.com
11
Voicebox repository
github.com
12
YuE2 repository
github.com
13
HyperFrames repository
github.com
14
BrowserSkill repository
github.com
15
Atlas repository
github.com
16
Kilo Code official site
kilocode.ai
17
Superset Mobile on Product Hunt
producthunt.com
18
Weave Router 2.0 on Product Hunt
producthunt.com
19
tiun. on Product Hunt
producthunt.com
20
CREEM official site
creem.io

New AI Tools Weekly #17: AI tools audit what agents install, write, and reach

At a glance

Auditing the agent skill supply chain

Output discipline: the anti-slop stack

Repository-scale review and specification

Local-first voice, music and video

Letting an agent into a real session, and keeping the record

Driving agent fleets from somewhere other than the desk

The commercial layer under AI products

Editorial view: what to test first

References

Related content

New AI Tools Weekly #17: AI tools audit what agents install, write, and reach

At a glance

Auditing the agent skill supply chain

Output discipline: the anti-slop stack

Repository-scale review and specification

Local-first voice, music and video

Letting an agent into a real session, and keeping the record

Driving agent fleets from somewhere other than the desk

The commercial layer under AI products

Editorial view: what to test first

References

Related content