Top AI Agent Projects : OpenSwarm, Claude Haiku 5.5, NOVA CLI, Semwright & Velozity

Author: ManuAGI - AutoGPT Tutorials · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Top AI Agent Projects : OpenSwarm, Claude Haiku 5.5, NOVA CLI, Semwright & Velozity · Channel: ManuAGI - AutoGPT Tutorials · 22:06 · Watch on YouTube

Topics: AI agents, Cybersecurity and software security, AI developer tools, Multimodal models

Summary

This weekly roundup presents 20 AI-related projects spanning coding, voice, research, productivity, security, hardware, and data collaboration. The featured tools include agent frameworks and integrations, models and benchmarks, desktop apps, and platforms designed to help people and AI agents work together. The presenters describe each project’s purpose, key capabilities, intended users, and how to try it. The video closes with DataBench, an open-source workspace for people and agents to collaborate on data.

Key points

Open Swarm is a Mac desktop where multiple agents use browsers, apps, and tools in parallel on multi-step tasks.
Claude Haiku 5.5 is presented as Anthropic’s fastest small model, aimed at high-volume, cost-sensitive work.
Sakura Bench compares speech-to-speech models through live phone-call scenarios, measuring reliability, task success, accuracy, response time, and cost.
Nova CLI maps a code repository and can generate, refactor, test, and fix code from the terminal while retaining context across sessions.
ToneFold generates editable music on Mac and Windows, with user approval required before its producer agent writes parts.
SemWrite gives AI agents a typed, policy-checked interface to applications such as Figma, Chromium, Blender, and MLT video.
Strak Comply lets teams manage compliance checks and evidence through integrations and an MCP server.
Agent Guard scans AI agent skills for security risks before installation, while Sign Frame M3 tests MCP servers with real agent tools.
Velocity combines a multiplayer AI office with agents, calls, transcripts, tasks, and integrations for remote and hybrid teams.
DataBench is a self-hostable workspace where people and AI agents collaboratively use Python and SQL notebooks.

Chapters

Start	Chapter	Summary
00:00	Introduction and announcement	The video introduces its weekly AI project roundup and announces the AI Agent Studio YouTube channel.
01:00	Projects 1–5: Agents, models, voice, and coding	The roundup covers Open Swarm, Claude Haiku 5.5, Griffin, Sakura Bench, and Nova CLI.
06:05	Projects 6–10: Workspace, monitoring, inference, research, and music	The presenters introduce Claude for Google Workspace, PolyLane, Liquid Inference, Articose MCP, and ToneFold.
11:16	Projects 11–15: Calls, narration, software interfaces, decisions, and mobility	This section features Claw Call, Neappalude Narrator, SemWrite, Seemo, and Moonwalkers Dusk.
16:27	Projects 16–18: Compliance and agent security testing	The video covers Strak Comply, Agent Guard, and Sign Frame M3 for compliance workflows and agent-related security and testing.
20:03	Projects 19–20: Team collaboration and data work	The final projects are Velocity, a multiplayer AI office, and DataBench, a collaborative open-source data workspace.

Quotes

Free AI desktop where agents swarm. — 01:00

Sakura bench is an open benchmark that tests real-time speech-to-speech voice models on live phone calls. — 04:03

SemWrite, a semantic interface between agents and software. — 13:20

Agent Guard is a scanner that checks AI agent skills for security risks before they touch your codebase. — 17:59

DataBench is a self-hostable, web-based workspace where people and AI agents collaborate on data live. — 21:06

Concepts and entities

OpenSwarm, Claude Haiku 5.5, Griffin, Sakura Bench, NOVA CLI, Claude for Google Workspace, Liquid Inference, Articose MCP, ToneFold, SemWrite, Seemo, Agent Guard, Sign Frame M3, Velocity, DataBench

Full transcript (original language)

00:00 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, voice, and data. 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

00:30 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 cutting-edge agent frameworks, make sure to check it out. Subscribe now to get weekly videos, in-depth guides, and real-time 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, the NL Open Swarm.

01:00 Free AI desktop where agents swarm. Open Swarm is a free AI desktop for Mac where many agents work at once. It solves the limit of single agent tools that handle one task while you wait and hand anything bigger back to you. Each agent gets its own browser, apps, and tools. So, a swarm uses your real browser, apps, and connected tools to finish the whole job rather than explaining how. You watch every agent on one infinite canvas, pan and zoom between them, step in when you want, and pick up the finished work. With one ask opening

01:31 several browsers that each read their own pages in parallel. You describe a tool and an agent builds it into your launcher. And an agent marketplace offers pre-built multi-agent apps like lead finder and post harvester. You bring your own models and API keys. It is built for Mac users delegating multi-step research, outreach, and ops. Join the free waitlist. Project number two, Claude Haiku 5.5. Anthropic’s fastest, most capable small model, Claude Haiku 5.5, is Anthropic’s fastest

02:02 and most capable small model built for high-volume, cost-sensitive work. It solves the problem that quick, repetitive tasks become too expensive to run at scale on larger models. It reliably handles summaries, compactions, database queries, and classification. And because it is Anthropic’s fastest model to date, it suits speed-sensitive work like live customer support and browser use. It also pairs with Opus and Sonnet as a sub-agent on coding. So, a bigger model builds the deck while a Haiku sub-agent pulls one exact figure

02:32 from a long filing. It is the first Haiku class model with an adjustable effort setting, so you optimize for cost or intelligence, and it runs far cheaper than the previous Haiku. It shows strong gains in alignment and ships with updated SDK support for computer use and browser use. It runs on the Claude platform, AWS, Google Cloud, and Azure. It is built for developers running work at scale. Get started with Claude Haiku 5.5. Project number three, Griffin by Tavis, a model for face-to-face

03:02 conversation. Griffin is Tavis’s first human interaction model, a new class of model built to understand and generate real-time face-to-face conversation. It solves the way machines force you to meet them halfway, managing your timing, simplifying your words, and watching a blank face for signs it understood. Griffin listens while it speaks and reads expressions, gaze, and pauses, not just words, as one unified video-to-video system that perceives, decides when to respond, and generates speech and video all at once. It

03:33 interrupts and gets interrupted without losing its place, backchannels and nods at natural moments, and generates every pixel of the scene in real time from one reference image, cloning a voice from about 10 seconds of audio. In a live study, nearly half of people who talked to it over a 1-minute call believed it was a real person. It is built for natural human computer interaction. Griffin light is a research preview for select testers. Request access through the waitlist. Project number four, Sakura bench, test speech-to-speech

04:03 models on real calls. Sakura bench is an open benchmark that tests real-time speech-to-speech voice models on live phone calls. It solves the guesswork of picking a voice model from vendor claims where no number tells you how one holds up on an actual call. It runs every model through the same open-source phone agent with the same prompt, tools, and caller scenarios, changing only the model across a short clinic booking suite and a longer Medicare intake suite. A simulated caller works through every scenario three times on live

04:34 calls, and a call passes only when the agent reached the right outcome and every saved field matched. Models are ranked on reliability, task success, saved data accuracy, stalled calls, response time, and cost, with each result linking to the full transcript and scores, so any figure can be checked, and a tool compares any two head-to-head. It is built for teams choosing a model for voice agents. View the leaderboard. Project number five, Nova CLI v1.0, an autonomous AI

05:04 developer in your terminal. Nova is a code-base aware AI developer that lives in your terminal and builds, tests, and fixes code. It solves the constant context switching of modern development where jumping between IDE, browser, docs, a chatbot, Git, and the terminal breaks your flow and makes you re-explain the project each time. You authenticate from the command line, point it at a project, and it maps your repo, understanding your structure and file connections, so you build without loading files first. It generates and

05:34 refactors code, reads stack traces, and runs a self-healing loop that catches and patches bugs on its own. With clean atomic diffs, you review in the terminal. It keeps persistent context across sessions, works across large multi-file code bases, and reaches over 160 AI models running natively in the CLI with your workflow under your control. It installs in seconds. It is built for developers. Install it and start building. Project number six, Claude for Google Workspace. Claude

06:05 working inside Docs, Sheets, Slides. Claude for Google Workspace is an add-on that puts Claude to work inside Google Docs, Sheets, and Slides. It solves the break in flow from leaving your document to ask AI elsewhere and pasting the result back. One install adds Claude to all three apps where it works alongside you in the file you have open. In Docs, it drafts, tightens, and structures writing. In Sheets, it builds financial models, analyzes data, and pulls in live numbers. In Slides, it builds decks from

06:36 your layout and flags overlapping elements. You review every change, seeing suggested edits, highlighted cells, and new slides before you accept. And it keeps your theme, heading styles, and surrounding formatting intact. You pull context from other tools through connectors, save a working process as a reusable skill, set preferences, and drop in a PDF, CSV, or Office file for Claude to read. In the add-on, it sees only the file you open plus connectors you turn on, and admins control who installs it. It is in beta on all paid

07:07 Claude plans. It is built for teams and professionals. Install it from the Google Workspace Marketplace. Project number seven, PolyLane for Vercel. Finds, diagnoses, and fixes Vercel issues. PolyLane for Vercel is an AI monitoring and incident response tool that finds issues in your Vercel apps, works out why, and writes the fix. It solves the manual digging of production work where you chase alerts, read logs, and trace a failure to its cause by hand. You install the integration from Vercel in about a minute over OAuth with

07:39 no SDKs, API keys, or alert rules, and pick which projects it can see. It learns how each project normally behaves and raises an issue when something is off, grouping failed builds and repeat alerts so you hear about each problem once. During an investigation, it pulls the project’s request logs and metrics, checks recent deployments and config changes, and shows what broke and why. When the cause is in your code, it opens a pull request on the repo behind the project, and any write waits for your approval. Projects, domains, storage,

08:09 and firewall rules sit in one graph with your stack, and any coding agent can query it. It is built for developers and DevOps teams. Get started free. Project number eight, Liquid Inference. Providers auction every prompt on price. Liquid Inference is a live auction exchange for LLM inference built by Architect Financial Technologies. It solves the cost and guesswork of picking one provider where you pay a fixed rate and never see who could serve the same request cheaper. For every request, inference providers post competing

08:39 offers, and the job is awarded to the lowest priced one that meets your rules for cost, intelligence, latency, and data handling with a firm price cap locked before the first token, so you know the most it can cost. Each job returns a receipt showing the winning provider, the cap, the final charge, and how many competed, and it publishes live order books, per model quotes, and price history. It tests every provider with standard benchmarks and publishes the results, and providers can monetize spare GPU capacity by posting and changing prices. It is Open AI and

09:11 Anthropic compatible, so your code works unchanged with tools like Claude Code and Cursor. It is built for developers and inference providers. Swap your base URL and start. Project number nine, Articose MCP. Run synthetic user research from your assistant. Articose MCP is a server that brings Articose’s synthetic user research into your AI assistant. It solves the cost and weeks of traditional research and the extra step of leaving your chat to run it. By letting you start a study from the same place you already work. Articulated

09:42 interviews AI personas built on the big five model with 30 facets and enforced stance diversity, blind to your hypothesis, so they push back instead of telling you what you want to hear. Then returns a structured report of motivations, objections, confusion points, and prioritized recommendations in under 30 minutes. Over MCP, you connect Claude, ChatGPT, or any compatible client and run the full flow. Start and confirm a study, check its status, reuse a saved audience, then follow up with a single persona, put a

10:14 question to the whole panel, or ask a finished study and get its answer back. It is built for SaaS teams, agencies, and consultants. Connect your assistant and run your first study free. Project number 10, ToneFold, describe a song, edit every note. ToneFold is a free, open-source AI music composer for Mac and Windows. It solves the gap between AI tools that generate a finished track you cannot change, and the labor of writing every part by hand. You describe what you want to hear in plain language, and it writes chords, melody, bass,

10:45 drums, and more into a piano roll, plays it at once through a built-in synth, and lets you change any note. It works in layers and sections across 11 editable kinds, so you lock the chords you like and ask for a new baseline. In producer mode, a producer agent proposes a plan naming each layer and section, and nothing is written until you approve. With the approval enforced by the app, rather than a prompt. Then specialists write only their own parts. Generation, playback, editing, and export run on your machine. Only the chat uses a

11:16 model, your own Claude login, or a local Ollama. You export MIDI, a stereo mix, and per-layer stems, and an optional plugin runs it inside a DAW. It is GPL 3.0. It is built for producers and musicians. Download it free. Project number 11, Claw Call, AI that makes your phone calls. Claw Call is an AI that makes phone calls for you. It solves the time lost to phone trees, hold queues, and talking to a rep for simple errands. You give it a US number and a task, and

11:46 it dials, listens to the menu, and presses the right keys, waits through the hold queue, and talks to whoever answers. Then reports the outcome with the transcript it captured. When a decision genuinely needs you, it calls your phone, briefs you, and connects you only if you say yes, and it never agrees to anything you did not authorize, saying it is an AI whenever asked. You trigger a call by trying your own number, texting it from a US or Canadian phone, or installing its chat GPT plugin and asking in plain English. And it also works with Claude, Grok, a REST API, and

12:19 a hosted MCP server for agents. Reserved plans give you a private caller ID and an inbound AI assistant. It is built for anyone who dreads the phone. Try a free call. Eppalude already appeared in an earlier batch, batch 11, entry 14. It’s note taker feature. This entry is its narrator feature. So here’s a fresh write up for it. Project number 12, Neappalude Narrator. Listen to any text on Mac. Neappalude Narrator is a Mac text-to-speech feature that reads almost anything aloud. It solves the time cost

12:49 of reading, letting you listen while your eyes and hands are busy. You select text in any app and press a hot key to start reading without moving it elsewhere. Or bring Narrator a PDF, EPUB, DOCX, web link, clipboard text, image, or screen capture, each saved to your library. A floating player highlights the sentence or word being read, or shows one word at a time in RSVP mode, and you click a word to jump there or look away and just listen, at speeds from half to triple. EP1’s voice reads on your Mac, so its text and audio

13:20 stay there, while cloud voices send the text to Eppalude for a wider library. And the picker says where each runs before you press play. You can put a narration on the web to finish on your phone or share a link. It is built for anyone with a reading backlog. Start a free trial. Project number 13, SemWrite, a semantic interface between agents and software. SemWrite is an open-source semantic layer that gives AI agents a typed interface to real software. It solves the fragility of agents guessing screen coordinates or inheriting a broad

13:51 shell where an action is neither precise nor safe. Instead of clicking pixels, an agent calls a named operation on an application object, a scene’s objects, a document’s nodes, a timeline’s clips through one broker. Every request passes through policy that checks scope and permissions before anything runs. So, discovering a capability never grants authority to use it. And sensitive actions like input, clipboard, capture, and app changes each need an explicit grant. References identify objects within a session, jobs expose progress,

14:23 and results come with read back and execution evidence you can inspect. Application drivers cover Figma, Chromium, Blender, and MLT video. And it works over MCP, a CLI, and an SDK. You bring your own agent, and it starts read-only. It runs on Linux, macOS, and Windows. It is built for developers connecting agents to applications. Install it and connect your agent. Project number 14, Seemo, a judgment layer for software that acts. Seemo is a decision model from Temporal Labs that

14:54 returns calibrated probabilities instead of generating text. It solves the problem that software acting on an AI’s output has to parse an essay and guess how confident the model really was. You send a state, text, a screenshot, or video along with a set of typed questions, and Seemo answers each as a choice, a score, or a value, returning a probability distribution for every one from a single pass in milliseconds. Because the probabilities are calibrated, your code can branch on numeric thresholds, route a case for

15:25 human review, or sum them across a corpus, rather than trusting a confident sounding label. It fits anywhere software needs to make a fast, grounded judgment, from classification and moderation to routing and triage, giving an agent a reliable decision layer. It is built for developers building software that acts. Explore it and ask your first typed question. Project number 15, Moonwalkers Dusk, robotic shoes that speed your walk. Moonwalkers Dusk is a pair of motorized robotic shoes from Shift Robotics, the lightest

15:55 the company has built. They solve the slowness of walking, letting you cover ground faster without changing how you move. You strap them over your own shoes, and an intelligence layer called Sensa reads your movement intent and matches its response to your stride, so the assist feels like part of your own motion. It adjusts control for terrain like gravel, grass, and shallow puddles, and recognizes when you shift from a walk to a run, keeping pace automatically. A wrap-around light shows what the shoes are doing and glows on the ground in the dark. An a companion app plans routes, tracks distance, and

16:27 personalizes settings. They come in three sizes for a secure fit, charge over USB-C, and let you swap straps, hubs, and wheels without tools. They are built for commuters and anyone who walks a lot. Check availability on the Shift Robotics site. Project number 16, Strak Comply, compliance you run from your agent. Strak Comply is a security-first compliance platform for getting audit ready without the sales call. It solves the manual grind of SOC 2, ISO 27001,

16:57 PCI DSS, HIPAA, and NIST CSF, where teams drag screenshots into upload forms and guess what the auditor wants. You connect your stack from over 100 integrations like AWS, Google Workspace, GitHub, and Okta, and 100-plus tests run every day against your real systems, each one counting toward every control it proves. When a test fails, it names who or what to fix and writes the steps for your setup. And your auditor reviews evidence in a portal without needing an account. You run the whole program from

17:28 your coding agent over an MCP server. So Claude Code, Code X, Gemini CLI, Cursor, or Open Code can file evidence, rerun checks, and answer the auditor with its name on every change. It also surfaces which AI tools can reach your data and scores vendor risk. It is built for security-first teams. Start a free trial. Project number 17, Agent Guard, scans AI agent skills before you install. Agent Guard is a scanner that checks AI agent skills for security risks before they touch your codebase.

17:59 It solves the blind trust of installing a skill where a skill MD can quietly run shell commands, read secrets, or carry hidden instructions you never see. You paste a skill MD or its instructions into the scanner, and in seconds it returns a trust report with security and quality scores, a risk rating, and the capabilities the skill actually has from reading project files to executing commands and reaching the network. It combines deterministic pattern checks with AI-assisted analysis to surface

18:30 shell commands, file and environment access, external downloads, credential access, suspicious URLs, obfuscated instructions, prompt injection patterns, and destructive operations. Each finding explaining why it matters and what to do. It is free, needs no account, and the report includes only issues evidenced in the skill you scan. It is built for anyone installing agent skills. Scan a skill free. Project number 18, Sign Frame M3, tests MCP servers against real agents. Sign Frame

19:01 M3 is a PESTY framework that tests MCP servers against the agents your users actually run. It solves a blind spot in MCP testing. A server can expose the right tool, yet the agent answers from memory and never calls it, which a bare model API never reveals. M3 runs real Claude code, Codex, open code, or Pi binaries against your server, records every MCP tool call, and fails the test when your server was skipped or called wrong. So, you fix the tool description and rerun. Direct tests check schemas,

19:33 results, and errors with no model, while agent tests check which tools a real harness used, all as ordinary Pesti files kept in your repo. You run the same suite across agents to compare. Pin the version your users run today and the next one to catch breaks early, and gate pull requests in CI where a run with zero tests fails. Your coding agent can write the tests from a skill, and any ACP agent works, too. It is local first and open source. It is built for MCP server developers. Read the docs and

20:03 start testing. Project number 19, Velocity, a multiplayer AI office for teams. Velocity is a multiplayer AI office where your team and your AI agents work together in one place. It solves the tool switching and scattered context of remote work, where calls, notes, tasks, and calendars live in separate apps your agents never see. You walk into a cabin to talk to a teammate or an agent, and presence, video calls, transcripts, tasks, and guests all stay in the same window, so your agents already know what happened. You spawn

20:35 agents powered by the AI subscription you already pay for, like Claude code and they run on your Mac or a cloud computer with nothing sent, paid, or signed without you. A built-in agent called Speed needs no separate subscription and reads your email, transcripts, and files to create and capture tasks, do deep research on your own data, and deliver a morning brief. It offers over 100 integrations plus a cloud computer for legacy apps. It is built for remote first and hybrid teams. Open your free office. Project number

21:06 20, DataBench by Alcar open-source data workspace for people and agents. DataBench is a self-hostable, web-based workspace where people and AI agents collaborate on data live. It solves the split between human analysts and agents working on data in separate tools with no shared place to chat, run code, and see results together. Teammates and an agent share each chat live, and the agent runs in its own sandbox on a machine you choose using open code or clawed code. You work in reactive plain

21:36 Python and SQL notebooks that people and agents edit together, compatible with Marimo with a renderer for tables and charts, and connect to any async Amela compatible database. You run chats and notebooks on the host or add your own machines over SSH and share them per team with organizations, nested teams, roles, and every authorization decision recorded. A model gateway calls Anthropic or OpenAI with your own keys, and nothing else leaves your network. It is Apache 2.0 open-source. It is built

22:06 for data teams working alongside agents. View it on GitHub. Thanks for watching. See you in the next video.