AI tools: how to try them yourself, all gathered here
These are the ones we featured on Instagram. Each time a new episode comes out, one more gets added here.
The guides on this page are free. Most of the tools are free to use, and anything that costs money is listed on each card under "Good to know".
Each card shows the words you wrote in the comments. If you find those words, that's the episode you watched.
Every tool below was made by its own company or developer. We didn't make them, and we have no partnership with them.
Terms can change, so check the official link once more.
01the words you wrote in the commentsAlternatives
Chinese coding tools you can use instead of Claude Code
Xiaomi, Alibaba, Z.ai, Moonshot, and Huawei have released tools that let you code from the terminal, like Claude Code. But MiMo Code from Xiaomi is the only one you can use for free right away without logging in, and the rest need a login, a subscription, or an API key (as of 2026-09-28).
What's good about itMiMo Code (Xiaomi) — The official page says, in its original wording, "No login, ready out of the box… on par with Claude Sonnet 4.6 — free" (the performance claim is Xiaomi's). zCode (Z.ai) — Uses the GLM-5.3 model. Z.ai's model documentation says the previous model, GLM-5.2 GLM-5.2, scored 81 on Terminal-Bench 2.1, while Claude Opus 4.8 scored 85 (the 81 in the video is this GLM-5.2 score). Qwen Code (Alibaba) — Open source (Apache-2.0), and the repository's own SWE-bench Verified average is 77.8%. Kimi Code CLI (Moonshot) — Open source (MIT). All the scores are numbers measured by the company that made it.
Good to knowNot all of them are free. zCode requires a paid subscription(a paid Lite plan · log in with a Z.ai account). Qwen Code's free login (Qwen OAuth) ended on April 15, 2026, so now it only works if you use an Alibaba Cloud Coding Plan or another company's API key. Kimi Code asks you, the first time you run it, to log in or enter an API key. Huawei CodeArts was an "invite-only test" when the video came out. Now you apply through a Huawei Cloud account for a free trial(a Chinese service). Before you put in company code or personal information, check each company's data policy.
Start with the easiest one: MiMo Code. On Mac or Linux, use the first line. On Windows, use the second line in PowerShell. Type only one of them.
After installing, go to your project folder and type mimo and the first-run guide will appear. If you use npm, you can also type npm install -g @mimo-ai/cli instead.
2Qwen Code (Alibaba) - the tool is free, the model needs a key
npm install -g @qwen-code/qwen-code@latest
You need Node.js 22 or later. In your project folder, type qwen , then /auth to pick the company and API key you want to use (you can also connect a model that runs on your own computer, like Ollama).
Run AI on your own computer inside Claude — Ollama × Claude Desktop
When you connect Claude in the Ollama app, the Claude Desktop screen stays the same, but the AI that answers switches to a model on your own computer(or an Ollama cloud model). Ollama announced this on its official blog on 2026-08-25.
What's good about itIf you pick a model on your computer, your text doesn't leave your machine and your usage limit isn't used up. Ollama has remote data collection (telemetry) turned off by default, and it states that it does not retain your data (Zero Data Retention). If you disconnect, Claude goes back to its original settings.
Good to knowFor now, it only works on Mac — the official Ollama docs say Windows is "in progress" (checked 2026-09-28). The announcement came from Ollama, and Anthropic's docs don't mention Ollama (it works through Claude's general-purpose external connection feature). A model on your computer runs only as fast as your computer, and its answers may be weaker than Claude's models. Cloud models require an Ollama login, and some are on a paid plan..
These steps are for Mac. They follow the official Ollama docs in order.
1Install Ollama
Download Ollama for Mac from the official site and install it. If you don't have Claude Desktop, Ollama will suggest installing it when you connect.
Open the Ollama app, then click Apps → Claude Connect and follow the prompts to restartClaude. Choose the model you want to use from Settings → Apps and click Restart Claude .
4Switch back to normal
Ollama Apps If you disconnect there, Claude's previous settings come back. They also come back if you quit Ollama. To do it from the Terminal, use the one line below.
A 27-billion-parameter AI on your own computer, for free — Qwen3.8-27B + LM Studio
Just install one free app LM Studio and you can download Alibaba's open model Qwen3.8-27B(27 billion parameters) and run it right on your own computer.
What's good about it After the first download, it runs without internet and needs no payment. The model is Apache-2.0 an open model, so you can use it commercially, and it can read images too. LM Studio is also free for work and business use, with no separate contract.
Good to know The LM Studio model page shows this model's The minimum memory is listed as 16GB (as of 2026-09-28). The model files are big, so they take up a lot of storage space, and the more compressed a version is, the lighter it is, but the rougher its answers get. On a Mac, only Apple chips (M1 and later) work. Intel Macs do not. Because it runs on your own computer, it can be slower than cloud AI.
The steps follow the video exactly: install the app → search → pick a version → chat.
1Install LM Studio
Download the version for Windows, Mac, or Linux from the official site and install it.
In the app, use the Search (magnifying glass) on the left to find qwen3.8-27b. Several versions with different levels of compression (for example, 4-bit and 8-bit) will show up.
3Pick the highest version your computer can handle
A bigger number (bits) means better answers but more memory use. LM Studio shows whether each version fits your computer, so download the highest version that shows as a fit.
4Load the model in the chat window
Chat At the top of the tab, pick the model you downloaded, and it starts running on your own computer.
Free AI from many companies in one place: FreeLLMAPI
It gathers the free allowance each company offers in one place, and when one runs out, it automatically moves on to the next. According to the repository description, that is 34 companies, 635 free models, and about 7.4 billion tokens per month combined(as of the README on 2026-09-28).
What's good about it You don't pay for each model separately. In your own program, you only need to enter one address and one key. The tool itself is free (MIT) and has about 29,000 GitHub stars (as of 2026-09-28). You can also connect it to coding tools like Claude Code and Codex.
Good to knowIt's not "done with one setup" — for each company you want to use, you have to get a free API key yourself and enter it. The 7.4 billion is each company's free allowance Combined total, and the repo notes that there is no top-tier model, and that late in the day the limits on the better models fill up first, so answers get weaker. personal experiments(per the repo's disclaimer), and each company's own terms of service still apply. For a review of the terms, see the repo's advice to avoid Coherein its notes. Claude modelsare not on the free list (this just plugs other models into a tool called Claude Code). The free install gets the new model list toolother models into it). The new model list is updated as the free setup 30 days later ; to get it right away, you need the paid plan ($19 a year or $49 for life).
Pick just one of the three easy paths . Numbers and commands are based on the README as of 2026-09-28.
1Windows and Mac: use the installer (easiest)
On the Releases page, download the Windows .exefile, or the Mac .dmgfile (arm64 for Apple chips, x64 for Intel), and install it. Click the icon in the taskbar (menu bar on Mac) and you will see your unified key.
When it finishes, open http://localhost:3001 in your browser.
3Add each company's free key
On the admin screen's Keys page, add the API keys for the companies you want to use, and set the order for falling back from one to the next on the Fallback Chain page. The top of the same Keys page has unified key. Use only one account per company (a repo rule).
4Plug it into your tools
If a program accepts an OpenAI-style address, put this in the address field http://localhost:3001/v1and put your unified key in the key field. For Claude Code, the one line below writes the settings for you (it backs up your existing settings first).
These are four questions that ask what Claude has picked up about you from all the conversations you've had so far. Paste the four prompts below into the same chat window, one after another, in order.
What's good about it Claude isn't digging up anything new. It rereads what you've already said and sums it upfor you. It runs from your habits and priorities, to your decision patterns, to two versions of the next 5 years, to one paragraph to read in the morning. It works on the free plan too.
Good to knowThis is not a psychological test — it's only a guess based on what you've said in your chats. It draws on two things: (1) the conversations in that chat window right now, and (2) what's saved in memory, if memory is turned on. So if you open a new window, the earlier answers lose their basis. If there's hardly any conversation, or memory is turned off, the answers will be shallow. Memory is on by default for Free, Pro, and Max, and in Settings → Memoryyou can look at what's saved, edit it, or delete it. Anthropic states that sensitive topics like health, religion, and politics are not saved by default (checked 2026-09-28).
In one chat window on claude.ai or the Claude app,paste the prompts in order, 1 → 4. The original text in the video is in English, and you can paste it in Korean too.
11. Five things I never told you
Send this to Claude
Based on every conversation we have ever had, tell me the 5 things you figured out about me that I never directly told you. Do not hold back to make me feel good.
(Original in English) based on every conversation we have ever had, tell me the 5 things you figured out about me that I never directly told you, and do not hold back to make me feel good.
22. How you think and decide
Send this to Claude
Tell me the patterns you see in how I think and make decisions, and the one thing I am in denial about.
(Original in English) tell me the patterns you see in how I think and make decisions, and the one thing I am in denial about
33. Two versions of the next 5 years
Send this to Claude
Predict my next 5 years if nothing changes. Then predict them again if I fix the one pattern that is holding me back the most.
(Original in English) predict my next 5 years if nothing changes, then predict them again if I fix the one pattern that is holding me back the most.
44. One paragraph for the morning
Send this to Claude
Now, based on everything you know about me, write the one paragraph I should read every single morning.
(Original in English) Now write the one paragraph I should read every single morning based on everything you know about me.
07the words you wrote in the commentsGoogle Skills
Google's official skills for Claude. Pick and install only Firebase, Cloud Run, and Gemini.
Google wrote skills in its official repository google/skills that teach Claude how to use Google tools. 153skills have been released (Apache-2.0, checked 2026-09-28). You can install just the three firebase-basics · cloud-run-basics · gemini-api mentioned in the video.
What's good about it Instead of guessing, Claude works by following the steps Google wrote, step by step. Each skill also comes with a note saying "don't use this for these kinds of jobs," which helps keep you from using it in the wrong place.
Good to knowThe skills are free,but Cloud Run and the Gemini API may charge you based on your Google Cloud usage(you are billed for whatever goes over the free limit). firebase-basics is for signing in to the Firebase CLI and creating a project.It does not cover app user sign-in (Firebase Auth) or deployment. gemini-api requires active Google Cloud credentials.
This is for your own computer where Claude Code is installed. Type the commands in a terminal outside of Claude Code.
1See the list first
You can see what is included without installing anything.
If you say things like "Create a new Firebase project for me" or "Deploy this app to Cloud Run," Claude follows the steps in the matching skill.
Cloud Run needs a gcloud CLI and a Google Cloud project, and you need to turn on the Cloud Run Admin and Cloud Build APIs (prerequisites listed in the skill).
08the words you wrote in the commentsVoice editing
Edit video just by talking: Google Gemini Omni Flash
Just say something like "Make the violin disappear" in a sentence, and Google's video model edits the video for you. It remembers earlier scenes, so you can change the background, lighting, and clothes with your words, again and again (example from Google's docs).
What's good about it You don't have to learn editing software. You can create a new video from text, and then keep editing it in a back-and-forth conversation. Each time you can make a clip of 3 to 10 seconds, from 360p to 4K, in both landscape and vertical (9:16). The official release (gemini-omni-1.1-flash) came out in August 2026.
Good to knowThe developer API has no free allowance — the Google pricing table shows the free tier as "Not available," and it costs about $0.10 per second of video (per the 2026-09-28 pricing table). For regular users, per Google's announcement It runs on a paid subscription to the Gemini app or Google Flow (AI Plus, Pro, or Ultra) . Videos it makes carry an invisible AI mark (SynthID). The Arcads service shown in the video is an ad-video service from a company separate from Google, and it has no public price list or free plan. The old name gemini-omni-flash-preview shuts down on September 30, 2026.
This is the path for developers (Python). You need a Gemini API key with paid billing turned on. If coding feels like too much, you can use the same sentence as-is in the Gemini app (paid subscription).
1Get an API key
Create a key in Google AI Studio and turn on billing. Put the key in the environment variable GEMINI_API_KEY .
Make the video with your first request, and in the second request, include the previous request ID (previous_interaction_id) and the sentence describing what to fix. This is exactly the example in Google's docs.
Send this to Claude
import base64
from google import genai
client = genai.Client()
res1 = client.interactions.create(
model="gemini-omni-1.1-flash",
input="A violinist playing on a rooftop at sunset",
response_format={"type": "video", "aspect_ratio": "9:16"},
)
res2 = client.interactions.create(
model="gemini-omni-1.1-flash",
previous_interaction_id=res1.id,
input="Make the violin invisible.",
)
with open("edited.mp4", "wb") as f:
f.write(base64.b64decode(res2.output_video.data))
arcadsFRESHR (the company name in the terms, France)
09the words you wrote in the commentsCourse map
Free courses from the company that makes Claude: the course map
This is Claude Academy (academy.claude.com) , run by Anthropic. It has courses on Claude Code, Cowork, Claude on the web, and the API, and it is free.
What's good about it These are explanations written by the people who made the product, not a summary passed around online. You don't need to log in to browse or start a course. To save your progress, log in with a free Claude account(no paid plan or work account needed, per the official FAQ). Each course has quizzes, and when you pass them all, you get a completion badgewith a public verification link you can post on LinkedIn and elsewhere.
Good to know The completion badge is not a certification. The exam-based certification (Claude Certified …) is a separate, paid product taken through Pearson VUE. Courses and hours are based on the official page as of 2026-09-28 and may change. This is an Anthropic service and is not affiliated with us.
Course map Pick them in the order mentioned in the video (hours are as listed officially).
1If you're new, start with the intro course
Claude 101(2.5 hours). If you'll use Claude Code, take Claude Code 101(1.5 hours). For Cowork, where you hand off whole tasks, take Introduction to Claude Cowork(2.5 hours). If you'll build a product with the API, take Claude Platform 101(1.5 hours).
2If you already know the basics — the developer course
Building with the Claude API(9 hours · 8 quizzes) is the longest. MCP, subagents, and skills aren't inside it. They are covered in separate, shorter courses: Introduction to Model Context Protocol (1 hour) · Model Context Protocol: Advanced Topics (1.5 hours) · Introduction to subagents (45 minutes) · Introduction to agent skills (1 hour).
3If you're using it at an organization
The materials for using Claude at work are collected in their own Claude for work bundle.
Log in with a free Claude account, take the course, and pass every quiz in it, and your completion badge appears right away (unlimited retakes — official FAQ).
Build a scroll-animated website with Claude — three steps
Add an animation library (Motion) and a design rules skill (UI UX Pro Max) to Claude Code, then pull in ready-made parts from 21st.dev, and you get a site that moves as you scroll. Below, in the order of the video, are the install commands and the skill address. 21st.dev Pull in the parts, and you get a site that moves as you scroll. Below, I listed the install commands and skill addresses in the order of the video.
What's good about it Claude Code is Anthropic's coding tool that runs in the terminal. Motion (formerly Framer Motion) is an animation library for React and JavaScript (GitHub stars 33,759 · MIT). UI UX Pro Max is a design skill that picks color combinations, font pairings, and UX rules for you (131,038 stars · MIT). 21st.dev is a collection of components that lists "12,000+ crafted React components, templates, and shadcn themes". Figures are as of the public pages on 2026-09-28.
Good to knowNot everything is free. The official docs say "A paid subscription is required." for Claude Code, so you need a paid plan. The 21st.dev pricing table has no separate free plan, and unlimited component copying and installing is listed under Builder ($6/month · billed annually), and AI features under Builder + AI ($15/month · billed annually) . Motion and UI UX Pro Max are MIT open source. Prices may change, and all four tools belong to their own companies and developers and have no partnership with us.
The steps follow the video exactly: Claude Code → Motion → UI UX Pro Max → 21st.dev. You need Node.js (to install Motion) and a paid Claude plan.
1Install Claude Code
This is the install method the official docs recommend. On Mac or Linux, run the top line. On Windows PowerShell, run just the bottom line. After installing, type claude in the folder for your site and log in.
Enter two lines in the Claude Code chat, one after the other (as written in the README). After that, just ask it to build a screen, and it picks color and font rules from this skill.
Pick a component you like on 21st.dev, copy the code (or prompt), paste it into Claude Code, and ask, "Fit this into the spot for ○○ on my page." The "Lego assembly" in the video is this step.
Based on the pricing table, unlimited copying and installing starts with Builder ($6 per month, billed annually). The public page doesn't say how many you can get for free.
Three things to install first when vibe coding with Claude Code: Ponytail, graphify, and OmniRoute
Ponytail(a skill that stops Claude from adding code and settings you didn't ask for) · graphify(a skill that maps your whole codebase) · OmniRoute(a tool that bundles several AI models into one gateway and switches to another model when you hit a limit). The repositories and install commands are listed below.
What's good about it Give it the same task and the results come out tighter, and when you ask "if I change this, what breaks?", you get an answer without opening every file. If one model hits its limit, your work keeps going. All three are public repositories and free: Ponytail (MIT, about 147,000 stars), graphify (Apache-2.0, about 122,000 stars), and OmniRoute (MIT, about 71,000 stars), as of 2026-09-28.
Good to know They are made by different people and have no partnership with us. Download them only from each official repository. The Python package name for graphify is graphifyy(with two y's). This is temporary while the project reclaims its original name, according to the README. OmniRoute is a free tool, but the free allowance of the models it connects depends on each company, so it changes often, and each company's terms apply as written. Claude Code itself is paid (starting at $20 per month for Pro).
Claude Code is assumed here. You can also install the three separately.
2graphify: install in the terminal, then draw the map in Claude Code
uv tool install graphifyy
graphify install
If you don't have uv, you can also use pipx install graphifyy . Then, in the Claude Code input box, type /graphify . and a project map (graph.html and others) is created in the graphify-out folder. After that, you can ask things like "If I change this function, what does it affect?"
After you install it, a gateway starts on your computer at http://localhost:20128 . Connect the companies you want to use in the admin screen, then launch Claude Code through this gateway with omniroute run claude --model company/model as in the README example. Per-tool setup is in the repository docs.
A website that moves in 3D as you scroll: Higgsfield MCP + Claude Code
For the images and videos on your site, use Higgsfield(an AI image and video service) makes them, Claude Codethen Claude Code builds a page that moves as you scroll. Below, the four steps from the video are written out as real commands.
What's good about it You don't have to hunt for videos or images separately. With a single Higgsfield MCP, you can call more than 30 image and video models, including Seedance 2.0, right inside Claude (Higgsfield explainer).
Good to knowIt's not free. Higgsfield uses plan credits(new accounts come with a few free credits), and Claude Code is also paid. Fable 5 may be billed separately (usage credits) instead of counting toward your plan's included usage, depending on your planand in that case Claude Code asks for your consent first. Max reasoning effort uses more of your usage because it thinks longer. A one-line prompt is not guaranteed to finish the job: you'll need to look at the result and ask for fixes a few more times. Higgsfield isn't in Claude's connector list, so you enter the URL yourself to connect it.
Claude Code as of this writing. First, sign up at higgsfield.ai.
11. Connect the Higgsfield MCP: one line in the terminal
claude mcp add --transport http --scope user higgsfield https://mcp.higgsfield.ai/mcp
Open Claude Code, then /mcp pick higgsfield and log in. On the claude.ai website, go to Settings → Connectors → Add custom connectorand enter the same URL.
33 and 4. Describe the site in one line + 'generate everything with Seedance 2.0'
Open Claude Code in an empty folder and enter something like the example below. Change only what's inside the brackets to your own topic (this is an example sentence).
Send this to Claude
[Our cafe's coffee beans intro] Build me a one-page website. Make it feel cinematic, with 3D depth and a rotating camera as you scroll, and write the motion in CSS. Generate all the images and short videos with Seedance 2.0 through the Higgsfield MCP. When you're done, also tell me how to run it so I can view it locally.
13the words you wrote in the commentsConnection Hub
Four connections (MCP) to plug into Claude: search, reading, browser, and apps
Perplexity(real-time web search) · Firecrawl(pull a page's full text from a single URL) · Playwright(open a real browser to click and fill in forms) · Composio(connects to 1,500+ apps, including Gmail and Slack). Below, the commands to add to Claude Code are listed one by one.
What's good about it Once installed, Claude answers with what's on the web right now, works with services that have no API by driving the screen, and does real work in apps like Gmail and Slack.
Good to knowNot all four are free(Based on official pricing as of 2026-09-28.) Playwright is free, open-source software (Apache-2.0). Firecrawl can extract page text and search without a key, but with usage limits, and if you sign up, you get 1,000 free credits per month. Composio's free plan includes up to 100,000 calls per month(no credit card needed). Perplexity is pay-as-you-go with no free allowance(a few dollars per 1,000 searches). Anthropic has not run a security check on these connections. Put your API key only in the key spot of the commands below, and never paste it into the address bar or a chat window.
Claude Code pricing (paid). Type the commands one line at a time in the terminal. When you're done, restart Claude Code and check that it connected with /mcp . You don't need to connect all four. I recommend starting with the free ones, Playwright and Firecrawl.
1Playwright: Control a Browser (Free)
claude mcp add playwright npx @playwright/mcp@latest
After signing up for Composio, open AI Client Setupin the dashboard and copy the connection settings (URL and key) for Claude Code. Apps like Gmail and Slack will show that app's login window the first time you use them.
A Site That Turns a Paper Link Into a Video Explainer: arXivisual
Add five letters, isual, after arxivin an arXiv paper's address (arxiv.org → arxivisual.org) and it becomes an explainer page with animation and narration.
What's good about it You can go through a hard-to-read paper with moving visuals for each section. If the paper has already been made, just open it. It was built by a team of Carnegie Mellon students, and the repository describes it as a TartanHacks 2026 winner.
Good to know Papers that have already been made open for free. Making a new one has a daily limit and a human check. Also, this repository has no license file, and each arXiv paper has its own license, so you need to check on your own before redistributing or using one commercially.
It works right in your browser, with no sign-up or install.
1Add Five Letters to a Paper Address
In the address of the paper you're reading, arxiv after isual into the address bar and open it. Example: the Transformer paper
The gallery has papers that were processed earlier, so you can view them right away without waiting. A new paper takes time to make, and there is a daily limit and a human check.
arXivisualArmaan Gupta · Raj Shah · Nikhil Hooda · Ajith Bondili (a student team at Carnegie Mellon)
15the words you wrote in the commentsScreenshot to Code
An open-source tool that turns screenshots into code: screenshot-to-code
Capture a screen and drop it in, and it rebuilds it as web page code. Besides screenshots, it also accepts mockups, Figma designs, and screen recording videosand exports to HTML, Tailwind, React, Vue, and more, six options in all.
What's good about it You can get the code for a screen you like and edit it to fit your needs. The source is public (MIT license, about 80,000 GitHub stars as of 2026-09-28), so you run it on your own computer. If Chromium is installed, it can open the page it made and check it on its own(if you run it with Docker, it is included automatically).
Good to know The program is free, but it doesn't run for free. To use it on your own computer, you need an API key from one of OpenAI, Anthropic, or Gemini, and you may be charged by that company for what you use. To use a screen recording video, a Gemini key is required(per the README). There is also a free route (Ollama), but the creator wrote that it gives poor results and is not recommended. To try it right away without installing anything, there is the hosted version run by the creator screenshottocode.com(a sign-up service; check the official page for pricing). Copying someone else's site as is can be a copyright problem, so use it for reference and practice.
What you need: Git, Docker (Docker Desktop), and one API key. The Docker route is the shortest (per the README, 2026-09-28).
1Download
git clone https://github.com/abi/screenshot-to-code.git
cd screenshot-to-code
2Add your key and start it
In the first line, put your keyin the key spot. For an Anthropic key, change it to ANTHROPIC_API_KEY=, and for a Gemini key, change it to GEMINI_API_KEY=.
echo "OPENAI_API_KEY=sk-your-key" > .env
docker-compose up -d --build
You can also add the key after opening the app, in the gear icon (Settings) .
3Open in your browser
http://localhost:5173 Open it, drag and drop a screenshot, and pick the code option you want (for example, HTML + Tailwind).
16the words you wrote in the commentsReels Process
Four steps to make a Reel with Claude Code
It turns a reel that did well into scene snapshots, → writes a new hook and script, → renders the visuals with code, → and handles comments and DMs with Instagram's official features. That's four steps. Below, each step shows the actual command you type.
What's good about it If you split up what you do and what the tools do, you don't have to start from scratch every time. Claude Code is an official tool from Anthropic, with about 148,000 GitHub stars(as of 2026-09-28).
Good to know Claude Code is paid (starting at $20 a month with Pro). Claude can't take a video file as-is — so you need to pull scene snapshots (frames) and captions out of the video and give those to it, and then it analyzes them (up to 100 images per API request, or 20 per message in claude.ai). It doesn't make voices or music on its own either — you have to record them or call another AI service, and most of those have a small free allowance or are paid. The code-based rendering tool Remotion is free for individuals and companies with 3 or fewer employees, and larger companies need a paid license. Automatic comment replies and DMs come from Instagram's featuresrather than Anthropic, and you can send only one reply per comment, and only within 7 days after the comment is posted. Use other people's reels for analysis only. Use it.
What you need: Claude Code(paid). Installation is one line from the official docs. On Mac and Linux, use the first line. On Windows, use the second line in PowerShell.
21. Break down a reel that did well into scene snapshots and captions
You can't feed in a video as-is, so you break it into images. A public plugin, watch takes the video, pulls out scene snapshots and captions, and shows them to Claude (type one line at a time in the Claude Code input box).
/plugin marketplace add bradautomates/claude-video
/plugin install watch@claude-video
/watch reel URL or video file path What happens in the first 3 seconds?
The first time you run it, it will guide you through installing yt-dlp and ffmpeg. Instagram often won't let you download without logging in — in that case, screen-record it on your phone and enter that file's path.
There's a free set of skills for reel analysis and scripts in your own voice, available publicly (reel-analyzer · reel-scripter · reel-builder). Claude doesn't know your writing style, so you need to paste in at least 5 of your past posts or captions for it to imitate your style.
Look at the scene snapshots and captions pulled out earlier and sort out: 1) what grabbed attention in the first 3 seconds, 2) when the cuts change, and 3) what keeps people watching to the end. Then, in the voice of my writing below, use the same structure to write a roughly 30-second reel script about [my topic]. For each sentence, also write the "text to show on screen" and the "scene to show." Don't make up any numbers or facts I haven't given you.
[Paste at least 5 of my posts here]
43. Render the screen with code — captions and motion in code, voice and music with other tools
Remotion is a tool that builds video out of React code. Create a new project and install the agent skill that Remotion officially released, and Claude Code will build scenes by its rules.
For the voice, recording it yourselfis the easiest and it's free. To use an AI voice or music, get an API key from that service and tell Claude, "Use this API to make a voice track from the script and add it." (Pricing differs by service.) If coding feels like too much, you can also put it together in an editing app like CapCut·Vrew instead.
54. Comment replies and DMs — an official Instagram feature
"Comment a word and get a DM" means calling Instagram's official API (private replies) from code. A professional (business or creator) accountand a connected Meta developer app are required, and you can send only one message per comment, and only within 7 days after the comment is posted. Tell Claude, "Following this document, write code that sends a DM when someone comments [word]," and include the link to the document.
A bundle of skills for Claude Code and Codex that write drafts of your LinkedIn posts, comments, and replies in your own voice (sergebulaev/linkedin-skills · MIT).
What's good about it The skills split the work: writing (20 opening hook formulas), comments, replies, a 7-day publishing plan, and profile cleanup. Before you post, the humanizer cuts long dashes down to about 1 per 100 words and flags and fixes paragraphs packed with AI-sounding words like leverage (since the 2026-09 revision, it leaves old words like delve alone by default). Every skill shows you a draft first and only acts once you approve.
Good to know The code is free (MIT), but using the skills and plugins in claude.ai or the desktop app requires a paid plan (Pro, Max, Team, or Enterprise) and turning on code execution (README). By default, you copy the draft and paste it in, and posting straight to LinkedIn requires connecting a separate service (Publora). This is not a tool that gets you past AI detectors. It only shows how much the verdicts of five detectors (GPTZero, Originality.ai, ZeroGPT, Sapling, and Copyleaks) disagree. Based on the README as of 2026-09-28.
Choose the option that fits of the three easy paths where you use it.
1Claude Code
Enter two lines in the chat window, one after the other (original text from the README).
Go to Customize → Plugins → Add → Add marketplace → Add from a repository, enter the address below, sync it, then find it in Discover and add it. This is the step the video describes as "enter the address in the Plugins tab, then add it after syncing."
sergebulaev/linkedin-skills
3Ask for a post
After installing, just mention a LinkedIn task and the right skill turns on automatically and shows you a draft first.
Send this to Claude
Write a draft LinkedIn post about one thing I learned this week. Polish it with the humanizer before I post.
Hera, the AI motion designer that turns your writing into motion graphics
A web service that turns a written description of the scene you want into an editable animation (hera.video · Y Combinator Summer 2025 batch).
What's good about it You set the direction by adding reference images (up to 5), reference videos, and logo files. You fix timing, camera movement, and color by describing the changes, and adjust fonts and sizes yourself in the editor. It has over 100 templates and exports to mp4, webm, and gif in 16:9, 9:16, and 1:1. It has an API and an MCP server, so you can call it from Claude Code too.
Good to know You can start on the free plan, but conditions like limits and watermarks can only be checked on the in-app plan screen, since there's no public pricing page. It is not "unlimited." The API key you use to connect Claude Code is also described in the official docs as "obtained on request," so you may not get it right away. Hera is a service of hera.video and has no partnership with us.
It works right in the browser. Connecting Claude Code (step 3) is optional.
1Open the Hera app and log in
Go to app.hera.video and log in. You can also start from a template.
2Describe the scene in writing, then fix it with words
Describe the scene you want in a sentence or two, and upload your logo or reference images. If the result isn't quite right, don't start over. Just ask for changes to the timing, camera, or color feel in plain words. You adjust fonts and sizes yourself in the editor.
Send this to Claude
A 5-second 9:16 video on a black background: our logo pops out from the center, and below it the words "Launching today" appear one letter at a time
3(Optional) Call it from Claude Code
After you get an API key in Hera's developer settings, it's one line in the terminal (straight from the official MCP docs). Put your key where it says "my API key."
claude mcp add --transport http hera https://mcp.hera.video/mcp --header "x-api-key: my API key"
An open-weight video model you run on your own computer: LTX-2.5
A 22B video generation model whose weights were released by Lightricks (August 2026 · ltx.io).
What's good about it Give it text, photos, or clips, and it makes a video and generates sound along with it. Add a spatial upscaler and it goes up to 4K. The weights and a LoRA training tool (GitHub Lightricks/LTX-2) are public, so you can fine-tune your own style. You can get started right away with the official ComfyUI workflow, and downloads on Hugging Face over the last month are about 1.6 million (as of 2026-09-28). Weights: huggingface.co/Lightricks/LTX-2.5 (agree to the terms before downloading).
Good to knowIt is not completely free — it uses the LTX-2.x Community License: free for commercial use if your annual revenue is under 10 million dollars, and a paid agreement if it's above that. It is not an OSI open-source license, so be sure to read the terms. The minimum specs listed in the docs are an NVIDIA graphics card with 32GB or more of VRAM · 32GB of RAM · 100GB of storage · CUDA 13.0 or higher. LTX-2.5 is a Lightricks model and we have no partnership with them.
The easiest way is ComfyUI (per the official LTX guide). What you need: An NVIDIA graphics card with 32GB or more of VRAM · 100GB or more of free storage(official README). If your computer doesn't meet these specs, running it on your own machine will be difficult.
1Install ComfyUI
Download ComfyUI Desktop from the official site and install it.
In ComfyUI, go to Manager(Ctrl+M) → Install Custom Nodes → search for "LTXVideo" → Install → restart ComfyUI. The models you need are downloaded automatically the first time you use it (agree to the terms on Hugging Face before downloading).
Open an example workflow that comes with the installation, add your text or photos, and run it. If you run out of graphics memory, follow the README's Low VRAM guide (splitting the model across memory).
Figure out what to build before writing code: GitHub Spec Kit
A free tool from GitHub that, instead of having AI jump straight into code, makes it write up what to build first (github/spec-kit · MIT).
What's good about it Describe what you want to build in plain words, and it creates a spec before any code, then leaves a document at each step in this order: plan → tasks → implement. A clarify step asks you questions first when something is vague. It works with more than 30 coding agents, including Claude Code · Copilot · Gemini · Cursor, and it has 139,141 GitHub stars (as of 2026-09-28).
Good to know The tool is MIT open source, but the coding agent and model you use with it are billed separately. It requires Python 3.11 or higher and uv. Command notation differs by agent. In Claude Code, you call commands with a hyphen, like /speckit-specify,(the /speckit.specify you see throughout the docs is just notation for explanation). It's a tool that sets up the order of writing the spec first, but it doesn't guarantee the output will come out right. Spec Kit is a GitHub project and we have no partnership with them.
What you need: Python 3.11+ · uv · Claude Code. Steps 1 and 2 are in the terminal. From step 3 on, you work in the Claude Code chat window.
1Install Spec Kit
It's one line in the terminal (straight from the README).
Create a project for Claude Code, then launch Claude Code in that folder.
specify init my-project --integration claude
cd my-project
claude
3Start by writing what you want to build
In the chat window, type /speckit-specify followed by what you want to build, in everyday language. If anything is vague, /speckit-clarify will ask you follow-up questions.
Send this to Claude
/speckit-specify I want to build a photo organizer app that groups photos by date and shows them as albums.
4Plan → Tasks → Implement
Run them one at a time, and read the document each step produces before moving on. After /speckit-plan, you can also name the tech you want to use (for example, Vite and JavaScript). After implementing, /speckit-converge checks whether the result matches the spec (README: repeat implement ↔ check until it says "Converged").
21the words you wrote in the commentsSave Your Limit
If you hit your Claude usage limit fast, here are three free tools that help you use fewer tokens
These tools either draw a map of your code ahead of time so the AI doesn't have to dig through whole files, or remember what you did in past sessions.
What's good about it Graphify — Turns your code and docs into a knowledge graph and has the AI ask the graph first before it searches through files. The code map is drawn on your own computer without an LLM (GitHub stars 116,845 · Apache-2.0). claude-mem — Uses AI to summarize what you did in each session, then feeds the context you need back in at the next session (stars 93,924 · Apache-2.0). CodeGraph — Indexes your code structure ahead of time and updates it automatically when files change. It's 100% local, and its README says the maker measured an average 62% token reduction (7 repositories · August 2026) (MIT). As of 2026-09-15.
Good to knowThis doesn't mean you'll never hit your limit — results vary depending on how you use them. Graphify uses an AI model when it processes documents, PDFs, and videos. For claude-mem, if you log in during setup, the hosted memory feature is a 30-day free trial, and after that it goes back to your own plan's pricing. The CodeGraph README also notes that in multi-turn conversations, more context may be left over at the end. All three tools are open source projects from their own makers and have no affiliation with us.
Four free things to install before you use Claude Code
Install these before you hand coding over to AI. They're free tools that make the AI write shorter code, pick the right model, and let you ask about your whole project at a glance.
What's good about it Ponytail — A skill that trims AI agent code down to shorter code; its README benchmark table lists code size -54% · tokens -22% (GitHub stars 138,866 · MIT). OmniRoute — An AI gateway that bundles 352 providers (150+ free) behind one address and switches to the next model when you hit a limit (stars 66,344 · MIT). Graphify — Turns your whole codebase into a knowledge graph you can ask questions about (stars 116,847 · Apache-2.0). Agent Skills — 25 skills organized by stage: define, plan, build, verify, review, and ship (stars 94,447 · MIT). As of 2026-09-15 on GitHub.
Good to know The tools are free and open source, but the coding agent and model fees you pair them with are separate. OmniRoute's "free providers" also follow each company's own free limits and terms. The benchmark numbers are the values listed in the Ponytail README and vary depending on how you use it. All four tools are open source projects from their own makers and have no affiliation with us.
Four tools to use with Claude Code when you build a website
Add an animation library, a design rules skill, and a component collection to one coding tool, and even your first site will look less plain.
What's good about it Claude Code is Anthropic's coding tool that runs in the terminal. Motion (formerly Framer Motion) is an animation library for React and JavaScript (GitHub stars 33,759 · MIT). UI UX Pro Max is a design skill that picks color combinations, font pairings, and UX rules for you (131,038 stars · MIT). 21st.dev is a collection of components that lists "12,000+ crafted React components, templates, and shadcn themes". Figures are as of the public pages on 2026-09-28.
Good to knowNot everything is free. The official docs say "A paid subscription is required." for Claude Code, so you need a paid plan. The 21st.dev pricing table has no separate free plan, and unlimited component copying and installing is listed under Builder ($6/month · billed annually), and AI features under Builder + AI ($15/month · billed annually) . Motion and UI UX Pro Max are MIT open source. Prices may change, and all four tools belong to their own companies and developers and have no partnership with us.
The steps follow the video exactly: Claude Code → Motion → UI UX Pro Max → 21st.dev. You need Node.js (to install Motion) and a paid Claude plan.
1Install Claude Code
This is the install method the official docs recommend. On Mac or Linux, run the top line. On Windows PowerShell, run just the bottom line. After installing, type claude in the folder for your site and log in.
Enter two lines in the Claude Code chat, one after the other (as written in the README). After that, just ask it to build a screen, and it picks color and font rules from this skill.
Pick a component you like on 21st.dev, copy the code (or prompt), paste it into Claude Code, and ask, "Fit this into the spot for ○○ on my page." The "Lego assembly" in the video is this step.
Based on the pricing table, unlimited copying and installing starts with Builder ($6 per month, billed annually). The public page doesn't say how many you can get for free.
Gemini API Agentic Video Understanding — It picks out only the scenes it needs from a long video
Google has opened up a new way of processing video in the Gemini API 「agentic」. The default method also looked at video by pulling out one frame per second . The new method, based on your question, picks only the scenes it needs and looks at those more densely and in sharper detail .
What's good about it Per Google's announcement Up to 88% fewer tokens · up to 66% lower cost · up to 7% better quality(for long videos). Pricing It's the same as the regular Gemini API token pricing,with no separate fee for this feature.
Good to knowIt's not a free feature — you pay API charges for what you use. It's a developer featurefor calling the API from code. There are four models you can use Gemini 3.8 Flash · 3.7 Flash · 3.6 Flash · 3.5 Flash Lite (per the 2026-09-28 docs). For short videos under 5 minutes, the docs recommend the default mode instead.
This is for Python. JavaScript and REST examples are in the official docs too.
1Get an API key
Create an API key in Google AI Studio, then set it as a computer environment variable GEMINI_API_KEY .
3One line in the video input: "processing": "agentic"
After you upload the video, in the video input set processing to agentic (exactly as in the official docs example).
import time
from google import genai
client = genai.Client()
video_file = client.files.upload(file="my_video.mp4")
while video_file.state.name == "PROCESSING":
time.sleep(2)
video_file = client.files.get(name=video_file.name)
interaction = client.interactions.create(
model="gemini-3.8-flash",
input=[
{"type": "video", "uri": video_file.uri,
"mime_type": video_file.mime_type, "processing": "agentic"},
{"type": "text", "text": "What are the three main claims of this video?"},
],
)
print(interaction.output_text)
To check that it's really turned on, look at whether the result's steps has a processing_call · processing_result in it (per the docs).
ECC — How to install a whole Claude Code setup bundle at once
A Claude Code setup collection that bundles skills, subagents, and hooks, released on GitHub for free (MIT)according to the README. It includes 68 agents, 292 skills, and 94 legacy-style commands.
What's good about it Instead of writing every skill and subagent yourself, you can start from a bundle that people already use. The README says the setup security checker (AgentShield) was "built at the Claude Code hackathon (Cerebral Valley x Anthropic, February 2026)."
Good to know Claude Code is paid — a paid plan is required (from Pro at $20 a month). It mainly targets Claude Code, and Codex is supported separately. Cursor is in beta and Gemini is experimental,so only some features work (see the README support table). The creator is a winner of the Anthropic x Forum Ventures hackathon in September 2025— ECC itself did not win an award. The README warns about fake copies, so download only from the official repository or npm . The numbers are as of the 2026-09-28 README.
What you need: Node.js 18 or later, Git, and Claude Code. Pick one of the two options below of the three easy paths (README: don't install both on top of each other).
1If you only use Claude Code
It's one line in your terminal.
npx ecc-universal setup
2To install it on Codex and other tools at the same time, use
the guided install, where you pick options as it asks.
npx ecc-universal install --guided
The README shows a command with a version number (ecc-universal@2.2.2), but as of 2026-09-28, npm only has up to 2.2.1 published, so typing it as written will fail. Type it without a version number, as shown above.
Greptile — an AI code reviewer that reads your whole repository and reviews your PRs
Once you connect a repository, it uses your entire codebase to automatically leave review comments on every pull request (PR).
What's good about it In a benchmark Greptile published, it says it caught bugs 82%(in the same table: Bugbot 58% · Copilot in the 50% range · CodeRabbit 44%). With one click, you can send a problem it found to Claude Code·Cursor·Codex·Devin and have it fixed.
Good to know The benchmark was measured by Greptile itself. Pricing — each month you start with 50 free credits, and Pro is 30 dollars per seat per month(50 credits per seat included) · extra credits are 1 dollar each (from the official pricing page). To review, your code goes to Greptile's servers — if it's company code, check your company's policy first.
1Sign in with GitHub (or GitLab)
Go to greptile.com and sign in with the account that has your code.
2Install the GitHub App on the repository you want reviewed
This gives the Greptile app permission to read the repository and leave reviews on PRs.
3Open a PR and get an automatic review
From a review comment, one click sends the problem's context to Claude Code or another tool to fix it.
When Claude Code gets slow and hits its limit fast — three things to check
What takes up space from the start (context) · hooks quietly running along · the CLAUDE.md it reads every time, you open these three places yourself.
What's good about it You only need to look at two commands and one file. The official docs give a standard, so it's easy to decide what to cut.
Good to know Claude Code is paid (starting with Pro at $20 a month). The % and hook counts in the video are values from an example session, so they differ from what you see on your screen. Hooks are not always bad — the official docs also say that handing repeated work to hooks and skills actually saves tokens.
A read-only list that shows your configured hooks, with a count for each event A read-only list that shows your configured hooks, with a count for each event.
From the official docs: "Aim for under 200 lines per CLAUDE.md file. Longer files use more context and get followed less closely." If it runs over, split the rules up so they're only read when needed.
Run an Instagram Reels factory with Claude skills: a free Skill Pack released by its creator
The creator of the demo in the video, jayantcreates.ai(Jayant Joshi · Ootto) , has released skills that follow the same workflow as a free (MIT) repository. Each skill handles one step: finding topics → hooks → scripts → Reels outline → captions → comment and DM text.
What's good about it/content-factory One skill runs the rest in order. If you don't have Claude Code, you can paste the prompts from the skill files into claude.ai (see the README).
Good to knowThe skills alone can't do the fully automatic setup shown in the video. The README also says, "Skills are the manual version, and the automatic version is Ootto" (Ootto is a commercial service, and its pricing is not public). Actually rendering the video takes outside tools like Remotion or Runway, and auto-posting to Instagram, auto-replying to comments, and sending DMs take outside connections like Composio. An automatic Instagram DM can be sent only once per comment, and only within 7 days after the comment is posted. Claude Code is paid (starting at $20 per month for Pro).
1Install: one line at a time in the Claude Code input box
Call it like the example below, then tell it your niche, your @account, a sample Reel URL (optional), and your comment keyword.
/content-factory
3What the skills do / what you have to connect separately
With skills alone: topics, hooks, scripts, a scene-by-scene outline, captions, hashtags, comment replies, and DM text. You have to connect separately: video rendering(Remotion·Runway) · Instagram auto-posting, comments, and DMs(Composio and others), and running everything for you (Ootto, paid).
Break down a Reel scene by scene and get a script in your own voice: a free skill released by its creator
The creator of the skill in the video, jayantcreates.ai(Jayant Joshi · Ootto) , has released skills that do the same job as a free (MIT) repository, which are reel-analyzer(analyzes a Reel's scenes and captions), reel-scripter(a script in your voice), and reel-builder(a scene-by-scene outline). The repository has no skill named exactly "viral-reels" like the one in the video.
What's good about it It pulls out what a hit Reel did in the first 3 secondsand where it switched cuts, then turns that into a script on your topic, in your voice. The claim that it was "built from 200 accounts" comes from the creator's explanation in the video.
Good to know The skill is analysis, plus a script and a scene-by-scene outline. Claude can't take a video file as-is, so you have to pull out the scene images (frames) and the captions separately (step 2). Adding a voiceover and rendering the actual video take other tools (step 5). Someone else's Reel: for analysis only. don't post it as-is after rewriting it. Claude Code is paid (starting at $20 per month for Pro).
is what this guide assumes. If you don't have Claude Code, you can open the skill file in the repo and paste its prompt into claude.ai (see the README).
1Install the skills: one line at a time in the Claude Code input box
The public plugin watch downloads the video, pulls out scene images and captions, and shows them to Claude.
/plugin marketplace add bradautomates/claude-video
/plugin install watch@claude-video
/watch reel URL or video file path What happens in the first 3 seconds?
The first time you run it, it walks you through installing yt-dlp and ffmpeg. Transcribing speech in videos that have no captions needs separate setup (see the README). Instagram often won't let you download without logging in — in that case, screen-record it on your phone and enter that file's path.
On a Mac, use brew install yt-dlp ffmpeg. The last line pulls one image every 0.5 seconds. claude.ai accepts up to 20 images per message, so pick the ones where the scene changes.
In Claude Code, point it at the folder of images and the captions you pulled, and run reel-analyzer to start the analysis. If you're not using the skill, just send the text below as-is. Run 3 to 5 Reels that did well through the same process, and you'll see what they have in common.
Send this to Claude
Look at this Reel's scene images (0.5-second intervals) and captions, and organize the following in a table.
1) The first 3 seconds: the text on screen, the first thing said, and the device that grabbed the eye (number, question, twist, etc.)
2) When the cuts happen (in seconds) and what changed at each cut
3) The devices that keep people watching to the end (teasers, lists, countdowns, etc.)
4) The action the viewer was told to take at the end (comment keyword, follow, etc.)
5) A one-line formula for the pattern that could carry over to other topics
Don't guess at anything you can't confirm from the images or captions. Write "Can't confirm" instead.
5Write a script in your own voice
Claude doesn't know how you talk. It can only imitate you if you paste in things you've written before, like posts and captions. In Claude Code, reel-scripter·reel-builder does this job.
Send this to Claude
Below is what I've written before, posts and captions.
[Paste at least 5 of your posts here]
1) First, sum up the features of my voice in 5 lines (sentence length, endings and words I use often, expressions I never use).
2) Then, using the pattern you pulled out earlier, write a script of about 30 seconds for a Reel about [my topic].
3) Next to each sentence, write the "on-screen text" and the "scene to show."
Don't make up any numbers or facts I haven't given you.
6Putting it together as a video: from here on, you're outside Claude
For the voice, recording it yourself(free), or use an AI voice. For editing, use apps like CapCut(free to use, with a paid Pro plan), Vrew(has a free plan, with usage limits on AI voices and more). The repo's reel-builder suggests Remotion(for developers), which renders video from code, and AI video clips Runway(credit-based).
5 public repos that each add one ability to an AI agent, including how to install them
These are the five featured in the video: Agency Agents(a set of specialist agents for different fields), Browser Use(an agent that operates the browser directly), Graft(helps coding agents get lost less), agentmemory(carries working memory over), OpenMontage(Making videos with agents). All five have their code public on GitHub.
What's good about it Instead of building an agent from scratch, you can start by plugging in abilities that are already built. Below, for each repository, I've included one or two lines to install it and how to use it the first time. Star counts (2026-09-28): Agency Agents 154K · Browser Use 116K · OpenMontage 61K · agentmemory 28K · Graft 9K.
Good to know "All free" means the code is free. Most of them need an AI coding tool like Claude Code (paid, starting at $20 a month for Pro), and Browser Use and OpenMontage may also have separate fees for AI model, image, and video generation usage. The licenses differ too: MIT (Agency Agents, Browser Use, Graft), Apache-2.0 (agentmemory), and AGPL-3.0 (OpenMontage: if you modify it and offer it as a service, you must release your code). Graft's claim of "46% fewer tool calls and 42% fewer tokens" is the maker's own measurement.
Pick one and dig in. The commands are exactly as written in each repository's README (checked 2026-09-28).
1① Agency Agents: a set of specialist agents for different fields (MIT)
The easiest way is It's a desktop app(Windows, Mac, Linux). In the app you browse the list of agents and install them into Claude Code, Cursor, Codex, and more with a click. To use the command line instead, download the repository and run the install script.
git clone https://github.com/msitarzewski/agency-agents.git
cd agency-agents
./scripts/install.sh --tool claude-code
According to the README, it includes more than 230 specialist agentsorganized by field (the video's 272 agents and 22 fields come from counting folders). The script is for Mac, Linux, and WSL, so on Windows we recommend the app.
2② Browser Use: an agent operates the browser for you (MIT)
Paste the text below into Claude Code (or Codex, etc.) exactly as it is, and the agent installs everything itself and connects to your browser (original text from the README, "Path 2: CLI").
Send this to Claude
Install or upgrade browser-use to the latest stable version with uv using Python 3.12, run `browser-use skill install` to register the skill, and connect it to my browser. If setup or connection fails, follow https://github.com/browser-use/browser-harness/blob/main/install.md.
The library is free, but you pay separately for AI model usage and the cloud browser ($0.02 per hour). You can also run it for free by using a local model through Ollama (README FAQ).
3③ Graft: give your coding agent a map of your code so it gets lost less (MIT)
It takes two lines in your project folder. init asks which tool to connect to, builds a code map, and links it to Claude Code. To see what it will change first, run `graft init --dry-run`.
npm install -g @nanonets/graft
graft init
You need Node.js. It sends anonymous usage stats, and you can turn that off with `graft telemetry disable` (README).
4④ agentmemory: carry an agent's working memory over to the next task (Apache-2.0)
It's one line on Node.js 20 or later. The first time you run it, it walks you through choosing the tool to connect (Claude Code, Cursor, Codex, etc.).
npx -y @agentmemory/agentmemory@latest
On Windows, it's quicker to use WSL2 or Docker Desktop because installing directly on Windows means downloading the engine files by hand (README). It works without a key, but searching memory by meaning needs a setting turned on.
5⑤ OpenMontage: make videos with a coding agent (AGPL-3.0)
What you need: Python 3.10+, FFmpeg, Node.js 18+, and an AI coding tool. After you install it, open that folder with Claude Code and tell it something like "Make a 60-second animated explainer video about how neural networks learn."
git clone https://github.com/calesthio/OpenMontage.git
cd OpenMontage
make setup
Windows doesn't have make, so use the PowerShell install line in the README. You can make videos with just free footage (stock video), but if you use AI image or video generation, itcosts money. The example videos in the README cost about $1.33 to $5 each.
7 free repos that save you time when building AI agents, and how to set up each one
These are seven public repos you can grab and use when building an AI agent (an AI that works on its own), so you don't have to write everything from scratch. They cover memory, a knowledge store, browser control, a list of design resources, security skills, diagrams, and research skills.
What's good about it Most of them hook up with one commandto agents like Claude Code. Based on each repo's README (2026-09-28): agentmemory has 17 skills, the cybersecurity skills repo has 818 skills , and the diagram design and research skills repo has 166 skills. The harness engineering collection is not a set of skills. It is a list of reading links(about 500 lines).
Good to know All of them are free public repos, but you pay separately for the AI models they run. OpenViking needs keys for embedding and vision models, and Browser Use needs a model key such as OpenAI. OpenViking is AGPL-3.0,so if you modify it and offer it as a service, you must release the source. Despite its name, Anthropic-Cybersecurity-Skills is a personal, community project with no connection to Anthropic(the README says so). Use the security skills only to check your own systems. Use it.
This is based on Claude Code. npx skills add is a public tool that installs skills on several agents (the npm package "skills" by vercel-labs). Pick and install only what you need.
11. agentmemory: give your coding agent a memory
npx skills add rohitg00/agentmemory -y
For the full install that also starts the memory server, follow the plugin install section of the README.
33. Browser Use: let your agent open a browser and type for you
The README has a request you can paste straight into Claude Code (install, then "browser-use skill install", then connect your own browser). To use it directly from Python, run the one-liner below and then follow the README example.
44. Awesome Harness Engineering: a list of design resources (for reading)
It's not something you install; A list of links to articles and tools on agent architecture design. Open the repository and read only the articles you need.
Agency Agents: How to add dozens of specialist AIs for different fields to Claude Code at once
A repository that collects guides for AI experts by role, such as engineers, designers, marketers, and testers,all in for free (MIT)one place. Put the files in your Claude Code folder, and you can call on them with a request like "Help me in Frontend Developer mode."
What's good about it Each role has its own written-out job duties, workflow, tools, and success criteria,so the answers are more specific than with a single assistant just swapping hats. The original repository has, as of its 2026-09-28 README, more than 230 agents, and you can pick and install only the departments you need.
Good to know The "68 agents, 9 departments" in the video is based on a copy of the original (itallstartedwithaidea/agency-agents, not updated since April 2026). Today, get the original (msitarzewski/agency-agents)instead. It is still being updated. Legal is not a dedicated department. It is one "compliance check" agent (in the Support department), and the Sales department exists only in the original. Agents are role guides,so Claude Code does the actual work. You need Claude Code, plus paid (starting at $20 per month for Pro).
What you need: Git and Claude Code. On Windows, run the commands in a Git Bash window (the install script is a bash script).
1① Get the original repository
git clone https://github.com/msitarzewski/agency-agents.git
cd agency-agents
Add everything, add only the departments you want, or copy a department folder by hand (README "Option 2: Use with Claude Code").
./scripts/install.sh --tool claude-code
./scripts/install.sh --tool claude-code --division engineering,marketing
cp engineering
if (!k) return;
var el = document.getElementById('c-' + k);
if (!el) return;
el.classList.add('mine');
var t = el.querySelector('.mine-tag');
if (t) t.textContent = '← You’ve already seen this episode';
if (!location.hash) el.scrollIntoView({ block: 'start', behavior: 'instant' });
} catch (e) {}
})();
Want the next free tool?
We share one free AI tool like this every day. Follow the account that sent you this reel so you don't miss the next one.