Top AI Agent Projects : Olostep, Superagent, oMLX, Almanac & screenpipe Transcript, AI Summary & Key Points

Everyday developers build new AI tools and keeping up with all of them takes time. This is our weekly project update video where we cover the top trending AI agent projects this week so you can find what is worth using in one place. You will discover trending AI tools and projects across coding, web data, and local AI without wasting time. Let's get started.

>> Before we jump into today's project updates, here's a quick announcement for everyone. We've launched a brand new YouTube channel called AI Agent Studio dedicated entirely to AI Agent projects, tutorials, and tools. So, if you're interested in staying up to date with the latest AI agent open source projects, learning how to build your own agents, or exploring cuttingedge agent frameworks, make sure to check it out.

Subscribe now to get weekly videos, in-depth guides, and realtime project breakdowns. The link is right there in the description. Don't miss it. All right, let's get into today's video. >> Project number one, Top View Motion Studio. Product brief to connected launch video. Top view motion studio is a guided AI workflow that turns a product brief into a connected launch video.

It solves the cost and weeks of timeline work a motion graphics launch normally takes in professional software. You write a brief covering what the product is, who it is for, and the moments to show. Upload product photos, UI screenshots, logos, and brand assets. Then pick a motion style, a 4 to 60-second duration, and one of six aspect ratios. Those inputs pass into canvas where an AI agent and a dedicated motion design skill organize pacing, transitions, typography, and scene continuity around your real product

context, not a single text prompt shot. It suits AI tools, apps, SAS, and brand campaigns. It is built for product marketing and design teams. Write a brief and create your first launch video. Project number two, OLOSEP, web data infrastructure for AI agents. OLOSEP is web data infrastructure that lets AI agents search, scrape, crawl, structure, and monitor the whole internet through one API.

It solves the browsers, proxies, and brittle scrapers teams otherwise maintained to turn messy pages into data a model can use. You call simple endpoints, scrape any URL into clean markdown, HTML, screenshots, or JSON to your own schema. Crawl a whole site, map every link, batch up to 10,000 URLs in minutes. Ask questions and get answers grounded in cited sources, or set natural language monitors that alert you on change by email, web hook, or SMS.

It renders JavaScript, handles anti-bot checks, and bills only for successful requests. native Python and Node SDKs, a CLI, and an MCP server. Connect Claude, Cursor, N8N, and more. It is built for AI developers and data teams. Grab a free API key and make your first call. Project number three, now Super Agent, a home for your coding agent. Super Agent is an open-source Mac app that gives your coding agent a real workspace to work in.

It solves the blind spot of terminal agents where you cannot see the browser they drive or steer them from away. Each task gets its own chat that keeps context and survives restarts. And any chat can run on its own git work tree, so commits land on a branch while your checkout never moves. A pane shows a browser the agent drives with your loggedin sessions.

A streamed iOS simulator it taps and swipes and the same page as desktop and phone at once. It knocks from the corner instead of stealing focus. And the paired iPhone app lets you approve or prompt from anywhere over an endto-end encrypted relay. It runs on Claude Code today on the subscription you already have with nothing leaving your Mac and no account.

It is built for developers. Download it and give your agent a screen. Project number four, Sens Scroll. A teleprompter that follows your voice. Sensroll is a browser-based AI teleprompter that scrolls your text as you speak. It solves the problem of fixed speed prompters where you chase the scroll, lose your place, and sound robotic. You paste in a speech or notes, hit present, and real-time speech recognition matches your words, fading each one out and keeping your place centered on screen.

When you improvise, it waits, then locks back on the moment you return, even if you jump to another section. It handles over 60 languages, including scripts that mix languages, and a built-in recorder captures takes with the text overlaid on your camera. It runs in the browser with no install. Your text stays on your device and only mic audio is streamed for transcription with recordings saved locally.

It is built for presenters, educators, podcasters, and creators. Paste a passage and try it free. Project number five, Shu Mlx, fast local models serving on Mac. Shu MLX is a native Mac OS inference server built on Apple's MLX for running local models on Apple Silicon. It solves the longwait coding agents hit where shifting context throws away the cache and every turn recomputes from scratch.

It persists each attention cache block to your SSD in a two-tier setup, keeping hot blocks in RAM and cold ones on disk. So when an agent circles back to a prefix, it is restored in milliseconds across requests and restarts, cutting time to first token from a minute plus to under 5 seconds. Continuous batching serves concurrent requests at once and it loads language vision embedding and re-ranker models together.

It offers open AI and anthropic compatible endpoints. So it drops into clawed code, open claw and cursor with a config the dashboard generates plus tool calling and MCP. A menu bar app runs it. It reads your existing hugging face models and it is open- source. It is built for developers running local AI. Download the app and start serving. Project number six, Nashen 7052 VC pitch deck analyzer.

Investor grade feedback on your pitch deck. The 17052 VC pitch deck analyzer is a free multimodal AI tool that reviews a startup's pitch deck the way an investor would. It solves the slow silent feedback loop of fundraising where investors pass without saying why and founders iterate in isolation for weeks. You upload your deck and it reads it slide by slide judging not just the text but slide design, data visualizations, financial models, and narrative flow across thousands of investor attributes.

It returns a grade, a slideby slide breakdown, and a prioritized fixed list separating what is objectively weak from what is merely a mismatch of sector or stage in under a minute. Because it is trained on tens of thousands of real decks and the funding decisions that followed, its feedback tracks what makes investors say yes rather than design opinion.

It is built for early stage founders preparing to raise. Upload your deck and see what investors will think. Project number seven, H4 Preview. Open model built for agentic productivity. A4 preview is 10 cents openweight flagship language model built for agentic coding, office work, and research. It targets the way models lose the thread across long tool use sessions and whole code bases.

It uses a mixture of experts design activating only a slice of its parameters per token and a context window past a million tokens. So it sustains long agentic runs, entire repositories and multi-hour tool traces without cost exploding. Tencent co-designed it with its own code buddy and workbuddy products training on data from software, gaming, finance and security experts.

So it plans, calls tools, and keeps executing complex tasks. It generates a playable game prototype from one request, reasons through scientific problems, and even helped optimize its own training. Released under an open license on Hugging Face with a quantized variant, it is priced for heavy token use. It is built for developers and teams running Long Horizon agents.

Try it and tell them what breaks. Project number eight, Coher Parse 5 turns enterprise documents into AI ready data. Coher parse is a document vision model that turns complex enterprise files into structured data your AI can search or act on. It solves the messy first step of any knowledge application where scan pages, tables, and diagrams resist clean extraction.

You send a document and parse reads it with highfidelity OCR while also understanding meaning and context detecting and interpreting embedded tables, charts and images rather than just lifting text. It returns precise bounding boxes for each region so downstream tools can highlight, attribute sources and reason spatially. And it works across nine major commercial languages.

Its output is retrieval optimized improving chunking, indexing, search and citation quality. and it feeds agents full document context available through the cohhere API and model vault on SageMaker and Azure. It also deploys on premises or airgapped for control and security. It is built for enterprise teams processing high volumes of documents. Try it and give your pipeline a boost.

Project number nine, stats dashboardless site analytics for coding agents. Stats is agentnative site analytics that turns your coding agent into a proactive analyst. It solves the problem of dashboards you never check where the numbers that tell you if a change worked sit in charts you have to read. There is no dashboard. Your agent is the interface tagging every deploy and comparing before and after metrics, tracing traffic spikes and refer anomalies to their source and mapping dropoff funnels to pinpoint leaks.

You ask how did the launch go inside claude or cursor and get an answer with evidence not another tab. The tracker is a tiny cookieless script with no IP logging and no cookie banner and your agent installs it while writing features. Adding one HTML attribute per action. Setup as a single MCP command and one key covers your whole portfolio. Working with any MCP client, it is built for indie makers and small teams shipping fast.

Connect your agent and start free. Project number 10. Video agent by photoour. Turn text or images into videos. Video Agent by Fo is a browser-based AI video generator that turns a text prompt or a still image into a short polished video. It solves the cost and skill barrier of traditional editing where a clip once needed software time and a timeline.

You choose text to video or image to video, describe what you want, pick a style, and it generates the clip in minutes. Then you refine it with built-in editing. It brings several top video models into one dashboard. So you create without separate subscriptions and finishes the job with tools for texttospech voice over, auto captions, background removal and enhancement.

You edit by typing plain commands rather than learning complex controls and it exports in high resolution. It works on the web and on iOS and Android. It is built for creators, marketers, and small businesses. Enter a prompt and generate your first video. Project number 11, Caddy discovers repetitive work and automates it. Caddy is an agent that finds your most repetitive back office work and builds automations to do it across the tools you already use.

It solves the manual high volume grind of professional services where intake, email triage, filing, and data entry eat staff hours. It reads your systems, ranks what actually repeats, and counts every step. Then you teach it like a new hire by screen sharing and narrating the job while it pokes at edge cases. as you answer in chat. From that it builds the automation running deterministic code so the same input takes the same path every time with AI only where judgment is needed.

Every run replays decision by decision with the permission it used. Agents work under scoped access no broader than the person's own and your data never trains models. It connects to over 150 tools like Microsoft 365, Salesforce, Docusine, Cleo and Net Documents. It is built for law firms, RIIA, accounting firms, and insurers. Show it one process and start free.

Project number 12. Micro Duck. Open-source biped robot. You retrain yourself. Micro Duck is a small open-source biped robot from Pollen Robotics that you teach new behaviors yourself with reinforcement learning. It solves the closed fixed function nature of most consumer robots, where what it does out of the box is all it will ever do. It stands about 25 cm tall on 15 motors with a camera, lidar, two motion sensors, and a beak that grabs.

And it walks, sits, stands, kicks, and picks itself up off the ground the moment you unbox it. Every behavior is a policy you can read, fork, and retrain. You train it in a physics simulator, on your own machine, or on hugging face. Deploy to the real robot in one step, refine, and publish your move for the community. The full SDK, simulation, and training stack are on GitHub under a permissive license.

So, what the robot runs is what you can change. It is built for AI builders, makers, and hobbyists. Pre-order one and teach it a trick. Project number 13, Open Tag. One tag routes work across many models. Open Tag is a model agnostic AI co-worker that lives in Slack and Microsoft Teams. It solves the cost and lock in of tagging a single model for every task.

When most tagged work does not need a Frontier One, you mention it in a thread with full company context and it routes the job to whichever of more than 80 models fits, reaching for Claude, GPT or Gemini only when the task needs them and returns the result where the work started. Each run happens on its own sandbox cloud machine scope to the tools you connected and torn down when it finishes with permissions following the person who asked.

It volunteers to own tasks it sees repeating, holds anything leaving the company until you approve, and keeps a self-updating wiki sourced to the threads decisions came from. It is built for whole teams, not one user. Add it to Slack or Teams and start free. Project number 14, Firecrawl Developer Index. Technical docs search for coding agents. Firecrawl developer index is a search API that gives coding agents one endpoint over the technical sources they ask about.

It solves the way general web search buries the exact thread or doc that answers a coding question under blog spam and stale pages. It holds over 70 million primary sources, readmes, GitHub issues, merged pull requests, documentation, and open API specs, and most refresh within a day of publishing. You search semantically pasting an error message to get the issue and the pull request that fixed it and matched passages come back as markdown so there is no second scrape.

You scope results by type, repo, language, topic, license or stars. Setup is a single command that wires it into clawed code, cursor or codecs which then search on their own and it needs no API key to start over a skill CLI, MCP or the API. It is built for developers running coding agents. Run one command and give your agent the index. Project number 15.

Nail manic. An agent with a company brain. Almanac is an always on AI agent that works from its own computer with a self-updating wiki of everything your company knows. It solves the way chatbot integrations fetch on demand but never remember. So every task starts from scratch. You connect your tools and Almanac learns your people, customers, and projects.

Compiling the work happening. across Slack, Gmail, Calendar, Docs, GitHub, and Meeting Notes into a wiki it reads before doing anything. With every line linked back to its source, so you can check and correct it. It runs on its own machine with a real browser, files, and login. So, even tools without an integration, it just signs into and uses like you would.

You text it work and it texts you when done. And it notices what needs doing on its own, pausing at a login, payment, or judgment call to ping you or hand over the live browser. Accounts you connect stay yours. Only the understanding reaches the shared wiki. It is built for teams. Get started free and give your company an almanac. Project number 16.

Heterinbe operating system to build and monetize agents. Heterinbe is an operating system for AI agents that lets you build one, launch it and charge for it fast. It solves the plumbing tax where shipping a single agent means stitching together nine separate vendors for the model, voice, browser, sandbox, memory, tools, metering, and payments. Each with its own keys, limits, and bill.

One API covers runtime, memory, a browser, tools, models, and billing. So, you add the SDK with one line, and skip weeks of integration before writing the part that is yours. per tenant metering tags every session to an end user so you can rebuild your own customers from day one and its token compression trims prompt and contact spend for the same output.

You can also hire ready-made agents instead of building. It runs on a shared credit pool with no per seat fees and idle agents cost nothing. It is built for developers and founders shipping agent products. Install the SDK and launch your first agent. Project number 17, Screenpe. Local computer history for any agent. Screenpipe is a memory layer that captures your computer history and opens it to any AI you use.

It solves the way assistants start from scratch every session, blind to what you actually saw, said, and did. It runs on Mac OS, Windows, and Linux, continuously recording screen frames with OCR, app, and accessibility data, and microphone and system audio, then storing it all in a local database on your device. You search that history for anything or let an approved agent read it to draft weekly updates.

Summarize recent calls and update your tickets grounded in real evidence rather than guesses. Any model reaches it through a local API and MCP server. So claude codecs or your own agent can turn a finished meeting into notes and follow-ups on its own. You exclude apps, windows, and URLs and redact sensitive fields before anything is stored with cloud upload off by default and the source is open.

It is built for professionals and teams who want AI grounded in their work. Download it free and build your workflow memory. Project number 18, Spline V2 browser 3D design tool with AI agent. Spline is a web-based collaborative design platform for building productionready interactive 3D experiences with an AI agent working alongside you. It solves the way 3D normally needs heavy specialist software and a developer just to get a scene onto a page.

You model with primitives, layerbased materials and real-time rendering, then add interactivity and motion through a system of states, events, timeline animation, physics, and particles. editing by direct control or by asking the AI agent to make changes. Variables, web hooks, and APIs bring live data into a scene. A companion canvas called HANA handles interactive UI and graphics with its own AI agent, vector networks, and effects.

You ship the result to the web, iOS, and Android, embedding into web flow, framer, react, next, JS, swift or cotlin, and remix designs from the community to start. It is built for designers and teams. Get started free and make anything 3D. Project number 19. Play with Putty. Multiplayer vibe coding you build together. Play with Putty is a Google Labs experiment in collaborative vibe coding where several people build tools and websites together in real time.

It solves the solo single cursor nature of most AI building tools where only one person can shape an idea at a time. You and others enter the same space and AI turns your plain language prompts into working software as you go with instant sync and multiplayer editing so everyone's changes appear at once. You interact directly with the interface too, selecting any component and telling it in natural language how to change its look or behavior.

So the app takes shape without breaking creative flow. It is a labs experiment built around lighter, faster models that respond immediately. Exploring how making things becomes a group activity. It runs in the browser and opens through a weight list and wave-based rollout needing a personal Google account. It is built for anyone who wants to create together rather than alone.

Join the weight list and take your imagination multiplayer project number 20. Revalvo local first eval workbench for prompt engineering. Revalvo is a local first workbench for prompt engineering and LLM evaluation. It solves the gap between chat playgrounds which are fast but leave no record and hosted eval platforms which are rigorous but slow in server side.

You run the same prompt against every model in parallel and compare answers side by side then score them with dozens of built-in evaluators roughly 2/3 deterministic checks like exact match reax JSON schema and PII patterns and the rest modelbased judges and similarity scorers each labeled before you attach it. Uver version prompts like code with git style history and batch test against whole data sets before anything ships to production.

It runs in your browser with no account and no hosted database. So your API keys stay with you and you point it at open router or any open AAI compatible provider or run fully offline with a local model. Setup takes under a minute and it never marks up your API spend. It is built for developers shipping LLM features. Open it and run your first prompt. Thanks for watching. See you in the next