Why it matters: Turns web search into an event-driven primitive other AI agents can build on.
Perplexity's Monitor API searches the web on a schedule and alerts agents via webhook only when relevant changes occur, using its own decision model to filter unimportant results. It turns search from one-off queries into persistent watchdogs.
Seedeo MCP lets ChatGPT and Claude connectors call AI media generators Suno, Kling and Veo, so assistant users can create music and video from within chat.
Resurf is a fully local knowledge library for Mac, iPhone and iPad that saves notes, links, images and documents and hands context to AI via MCP or a Mac CLI.
Webless Agents lets users describe an outcome in plain English; it builds a customer-facing agent grounded only in published content, with every instruction reviewable before launch, then answers questions and books meetings.
oovi runs a team of AI bots on the desktop: a leader plans, a researcher digs, a writer drafts and a critic checks, with approval before output. Free with your own key or from $10/month.
Omi CLI is a terminal coding agent that reads a repo, shows every edit as a diff for approval and runs tests until they pass, auto-selecting from 13 models under one plan; installable via npm.
Ulika is a visual editor for React, Vite or Next sites built with Claude, Cursor or Codex, letting anyone edit copy and images on the page and publish without touching code; open beta, free to try.
TeaserKit installs in one line and makes Claude Code, Codex, Cursor and other agents direct a launch video, ad or explainer from a repo or site, including concept, style, song and quality check; one-time purchase.
Squint lets users press a hotkey, drag a box around anything on screen and get a streaming answer beside it; now on Mac with native-language answers and one-press copy to coding agents. Free plan, also Windows and Linux.
Zapyso is a Mac app that records calls without a bot, auto-detects Zoom, Teams and Meet, and writes on-device transcripts and summaries with decisions and next steps. Also offers hold-to-dictate in any app. Free, needs Apple silicon.
AWS Quick now has a Meta ads AI connector, letting advertisers use natural language to create, manage and analyze campaigns, and combine Meta ad data with other sources in dashboards and decks.
Why it matters: Unfunded AI compute pledges raise doubts about the datacenter buildout's financing.
WSJ reports Wall Street pledged $500B for Nvidia but has deployed none of it, as analysts debate the AI trade selloff, Nvidia board governance and valuation.
Muru is a new AI-assisted Rust server from Graham Dumpleton, creator of mod_wsgi, aimed at serving ASGI and WSGI Python web applications in both development and production.
Pandex is an MCP server exposing 2.8 million European legal documents and EU law to AI agents, letting them query official legislation through a standard interface.
SurveyMonkey launches a Gemini integration that feeds survey and customer-feedback data into Google's assistant, letting teams analyze responses and draft follow-ups with AI. Small Business Trends reports the launch targets small-business users.
Snowflake now lets its enterprise AI agents produce maps as a native output, so agents can render geospatial results directly inside enterprise workflows.
Upfirst is packaging AI voice agents as a fully integrated phone system aimed at small businesses, combining call answering and telephony features in one offering.
FS2Crew's Generation AI PMDG 737 SOP 1 Edition is a flight-simulator add-on that uses AI voice to simulate a first officer and crew, letting pilots fly standard operating procedures with spoken callouts.
Prabhkiran Singh, cofounder and former CEO of Bewakoof, has launched Save8, an AI shopping assistant that compares prices to help consumers find deals.
Why it matters: A 2.8T-parameter open-weight model pushes what can be run and studied outside closed labs.
Moonshot launched Kimi K3, an open-weight model with 2.8 trillion parameters. A homelabber ran it on two RTX 3090s at 3.11 tokens/s using Optane memory, moving nearly half the model into a RAM drive to gain 41% speed.
Why it matters: Cheaper idle pricing makes always-on voice agents more economical to run.
OpenAI released GPT-Live-1.1 for voice agents via its API, cutting idle cost to $0.02 per minute. The update targets developers building real-time speech applications.
GladiosWAF is a web application firewall that uses AI instead of manually written rules to block attacks; the company says it democratizes zero-rule web security.
AlterContacts is a voice-first assistant for client-facing work: talk through a meeting and it updates the client's profile, tracks promises and renewals, and drafts follow-up messages for approval.
AI Context Flow keeps one shared memory over MCP so multiple AI tools and teammates work from the same version, with buckets by project or client, access controls and edit history.
Glassmail is a native Mac email client for Gmail, iCloud, Outlook and IMAP whose AI runs locally. Ask what needs attention today or dictate a reply in your own voice; no mail server, no subscription.
OpenVorld is an open-source harness that gives AI agents a visible pixel-art village, where each agent has a name, role and brain (Claude Code, Codex or any OpenAI-compatible API) and asks before running commands. Free and local.
Machina Labs, which builds AI-powered robotic manufacturing, launched Machina Rebellis, an automotive brand using robotics to make customized vehicle production faster and cheaper.
Vectigo, a Colorado-based division of an Irish company, released a roadway-embedded safety system using solar-powered sensor units called PODs to improve conditions for conventional and autonomous vehicles.
Why it matters: Rising AI server costs squeeze datacenter budgets and AI capacity plans.
Dell, HPE, Lenovo and Supermicro are demanding up to 45% more for AI servers after deals are signed, with the higher costs hitting trading firms and Nvidia systems.
Chinese chipmaker MetaX said its C-series GPUs now support the Hy4 model and released FlagOS W8A8 weights for the hardware. Results are reported from a 32-GPU test setup.
StarBot Spark is a US-built humanoid robot pitched for classrooms and living rooms, part of a wave of companion and education robots entering everyday settings.
Why it matters: Tesla's vision-only FSD update shows how much real-time AI a single in-car chip can handle.
Tesla's FSD v15 runs real-time AI video using about half of the AI4 chip's TRIP accelerator, per reports. Tesla also wants its cars to share traffic updates with Grok.
Why it matters: A mega-acquisition that would bring chip-efficiency software in-house for AWS.
Amazon is reportedly in advanced talks to acquire AI chip software startup Decart for about $7.5 billion, a move aimed at helping AWS compete with Nvidia on computing efficiency.
Houston-based Bot Auto launched Capacity Grid, a freight network matching shipper demand with autonomous trucking capacity, backed by agreements covering 200 million miles of driverless freight.
The Knoxville Police Department launched a Drone as First Responder program that dispatches semi-autonomous drones to 911 calls, reaching nearly all of Knoxville in 90 seconds or less.
Why it matters: Shows the debt and collateral risks building beneath the AI datacenter boom.
Congress is scrutinizing Nvidia's ties to CoreWeave, its key supplier and $6.3B backstop, while lenders backing SpaceX's $43B Nvidia GPU deal worry the collateral could end up in orbit and be unrecoverable.
Nvidia released GeForce Hotfix driver 621.35 to fix numerous excessive and unnecessary SSD writes that could shorten drive lifespans; Nvidia says RTX 50 cards are unaffected, though some owners dispute that.
Waddle-1 is a cloud reasoning agent for robots from a two-person YC startup. Waddle Labs says it scored 81 on Robocurve, beating GPT-6.2 Astra and Dream-1 on unseen bimanual tasks, and opened a preview API.
Why it matters: The recall highlights safety and regulation challenges for driverless fleets at scale.
Waymo issued a recall after its robotaxis drove into closed freeway lanes and struggled to follow manual traffic directions; several vehicles stalled during San Francisco flooding. The company is updating software.
Why it matters: Extends Nvidia's flagship AI and HPC platform roadmap to 2028.
Nvidia expanded its Vera Rubin platform at SC26 and unveiled a Rubin HPC GPU with over 50,000 CUDA cores expected in 2028. Nvidia also released cuPhoton for GPU-accelerated scientific data processing.
Why it matters: Removes doubts about financing for one of the largest AI datacenter buildouts.
Oracle said its OpenAI infrastructure deal is fully funded after a final capital raise. Separately, the WSJ reported OpenAI is partnering with NATO on cyber defenses.
ThreeJS-3D-Scenes-v2 is a Hugging Face dataset of 3,445 curated Three.js scene examples across 50 domains, including 1,431 voxel scenes, with prompts, reasoning traces, HTML/JavaScript responses and rendered references for training 3D-capable models.
Why it matters: Turns web search into an event-driven primitive other AI agents can build on.
Perplexity's Monitor API searches the web on a schedule and alerts agents via webhook only when relevant changes occur, using its own decision model to filter unimportant results. It turns search from one-off queries into persistent watchdogs.
The piece explains how Anthropic's Claude Code checks agent-written code through verification loops, describing the review steps that catch errors before changes land.
A Python repo that spins up 20 Claude Code agents on iOS Simulators and Android emulators via GitHub free runners to find broken flows, translation, layout and accessibility bugs.
Seedeo MCP lets ChatGPT and Claude connectors call AI media generators Suno, Kling and Veo, so assistant users can create music and video from within chat.
A Stepback newsletter argues the vibe-coding revolution is coming to all gadgets, with consumers coding custom features for devices rather than just apps.
pi-H3 is a Hugging Face Gradio demo that generates MiniMax H3 video in 2-4 steps using pi-Flow V2 few-step generation and DWD direct pixel prediction, skipping the video VAE decoder.
BeanPod is a native macOS app for managing recordings from voice recorder devices, offering on-device transcription, meeting notes and to-dos, a knowledge-base Q&A, sentence-by-sentence playback, English learning and MCP access for Claude and Codex.
Manus, a general AI agent app, updates its iPhone and iPad app to version 26.19.16 with Manus 2.0: remote control of your computer or cloud computer from the phone, plus Manus Studio collaboration for coding and app use.
MarkTechPost ranks personal AI agents including OpenClaw, Hermes Agent, Gemini Spark, OpenAI dots, Meta Muse and Perplexity Personal Computer on memory, security defaults and real cost.
A Hugging Face dataset repackages the CMU Graphics Lab Motion Capture Database as GLB files: 2,426 clips, 2.28M frames (10.5 hours) at 60 fps, six categories, retargeted onto a free character plus bare source skeleton.
Resurf is a fully local knowledge library for Mac, iPhone and iPad that saves notes, links, images and documents and hands context to AI via MCP or a Mac CLI.
Webless Agents lets users describe an outcome in plain English; it builds a customer-facing agent grounded only in published content, with every instruction reviewable before launch, then answers questions and books meetings.
oovi runs a team of AI bots on the desktop: a leader plans, a researcher digs, a writer drafts and a critic checks, with approval before output. Free with your own key or from $10/month.
Omi CLI is a terminal coding agent that reads a repo, shows every edit as a diff for approval and runs tests until they pass, auto-selecting from 13 models under one plan; installable via npm.
Ulika is a visual editor for React, Vite or Next sites built with Claude, Cursor or Codex, letting anyone edit copy and images on the page and publish without touching code; open beta, free to try.
Squint lets users press a hotkey, drag a box around anything on screen and get a streaming answer beside it; now on Mac with native-language answers and one-press copy to coding agents. Free plan, also Windows and Linux.
SitePoint benchmarks four local LLM serving tools — Ollama, LM Studio, llama.cpp and vLLM — for hosting a local model to several users, comparing throughput and behavior under concurrent load.
Zapyso is a Mac app that records calls without a bot, auto-detects Zoom, Teams and Meet, and writes on-device transcripts and summaries with decisions and next steps. Also offers hold-to-dictate in any app. Free, needs Apple silicon.
A command-line interface for Amp, a coding agent and development environment. Installable globally from npm with npm i -g @ampcode/cli; this release bumps the version with no listed changes.
Why it matters: Shows coding agents can produce competitive GPU kernels for under-supported hardware.
An AI agent built the Loom package of AMD kernels in 4.3 hours, nearly doubling Inkling 2 Mini decode speed and hitting 5.9x llama.cpp prefill at an 18k-token prompt, against what an AMD engineer called super nascent day-one support.
fast-anything is a tool that profiles Python, Node or Ruby apps, finds the 5% that is slow, and has a coding agent port just those hot functions to Rust, proving behavior matches.
ThreeJS-3D-Scenes-v2 is a Hugging Face dataset of 3,445 curated Three.js scene examples across 50 domains, including 1,431 voxel scenes, with prompts, reasoning traces, HTML/JavaScript responses and rendered references for training 3D-capable models.
pi-H3 is a Hugging Face Gradio demo that generates MiniMax H3 video in 2-4 steps using pi-Flow V2 few-step generation and DWD direct pixel prediction, skipping the video VAE decoder.
A Hugging Face dataset repackages the CMU Graphics Lab Motion Capture Database as GLB files: 2,426 clips, 2.28M frames (10.5 hours) at 60 fps, six categories, retargeted onto a free character plus bare source skeleton.
SenseNova-Loop-9B-GGUF is an Apache-2.0 quantized build of SenseNova's 9B image-text-to-text model, letting people run the multimodal model locally; the base model was also linked for download.
tiny-voca is a synthetic American English speech corpus with a female and male voice reading the same 17,371 sentences, about 19.1 hours of 24 kHz audio, meant to train the tiny-voca TTS model. Released under CC BY 4.0.
FastFrames-S is an open text-to-video model from Apple aimed at interactive-latency generation. It is built on Wan2.1-T2V-1.3B and published with code and a paper.
Inkling 2 Mini is a multimodal open model from Thinking Machines Lab that takes text and images and generates text, using looped reasoning to solve tasks with fewer tokens. At 26B parameters in a mixture-of-experts design, it runs locally on a 32GB Mac; transformers v5.24.0 adds support.
EVIDENT is a Hugging Face dataset of 220 GUI/web agent task records (150 Korean, 70 English) pairing user requests with intended outcomes and evidence checks, for evaluating agent trajectories.
A Python repo that spins up 20 Claude Code agents on iOS Simulators and Android emulators via GitHub free runners to find broken flows, translation, layout and accessibility bugs.
BeanPod is a native macOS app for managing recordings from voice recorder devices, offering on-device transcription, meeting notes and to-dos, a knowledge-base Q&A, sentence-by-sentence playback, English learning and MCP access for Claude and Codex.
fast-anything is a tool that profiles Python, Node or Ruby apps, finds the 5% that is slow, and has a coding agent port just those hot functions to Rust, proving behavior matches.
A new GitHub repo, AI Tutor, is a self-hosted learning assistant that turns study notes, PDFs and lectures into chapter paths and quizzes, only teaching sentences it can quote, with bring-your-own model key.
PiBot is an open-source, local-first AI agent team for your desktop, with multi-agent group chats, computer use, browser control, MCP, routines and approvals, running on macOS, Windows, Linux and Android. It drew 107 stars in its first week.
AI4XE/ai4ai is a TypeScript toolkit that lets AI agents manage themselves, covering runtime status, context renewal and self-duplication across agent harnesses. It drew 101 stars in its first week.
Inkling 2 Mini is a multimodal open model from Thinking Machines Lab that takes text and images and generates text, using looped reasoning to solve tasks with fewer tokens. At 26B parameters in a mixture-of-experts design, it runs locally on a 32GB Mac; transformers v5.24.0 adds support.
New open-source JavaScript repo collecting agent skills for the NestJS Node.js framework, letting AI coding agents work with Nest projects; gained 100 stars in its first week.
TraxAgent is a read-only MCP server and Claude Code plugin that lets Claude, ChatGPT, Claude Code, Cursor or VS Code show who spent your AI budget and what to fix; the repo drew 109 stars in its first week.
Manus, a general AI agent app, updates its iPhone and iPad app to version 26.19.16 with Manus 2.0: remote control of your computer or cloud computer from the phone, plus Manus Studio collaboration for coding and app use.
MarkTechPost ranks personal AI agents including OpenClaw, Hermes Agent, Gemini Spark, OpenAI dots, Meta Muse and Perplexity Personal Computer on memory, security defaults and real cost.
Resurf is a fully local knowledge library for Mac, iPhone and iPad that saves notes, links, images and documents and hands context to AI via MCP or a Mac CLI.
oovi runs a team of AI bots on the desktop: a leader plans, a researcher digs, a writer drafts and a critic checks, with approval before output. Free with your own key or from $10/month.
Squint lets users press a hotkey, drag a box around anything on screen and get a streaming answer beside it; now on Mac with native-language answers and one-press copy to coding agents. Free plan, also Windows and Linux.
Zapyso is a Mac app that records calls without a bot, auto-detects Zoom, Teams and Meet, and writes on-device transcripts and summaries with decisions and next steps. Also offers hold-to-dictate in any app. Free, needs Apple silicon.
JournalArta reports that users ending relationships with AI companions experience genuine emotional distress akin to human breakups, highlighting the emotional weight of long-term chatbot attachments and raising questions about how companion apps should handle separations.
SurveyMonkey launches a Gemini integration that feeds survey and customer-feedback data into Google's assistant, letting teams analyze responses and draft follow-ups with AI. Small Business Trends reports the launch targets small-business users.
DeepSeek's official iPhone and iPad assistant app updated to version 2.9.0, powered by its latest flagship model, with fixes for known issues. Free to use.
AlterContacts is a voice-first assistant for client-facing work: talk through a meeting and it updates the client's profile, tracks promises and renewals, and drafts follow-up messages for approval.
Glassmail is a native Mac email client for Gmail, iCloud, Outlook and IMAP whose AI runs locally. Ask what needs attention today or dictate a reply in your own voice; no mail server, no subscription.
SuperMailbox unifies Gmail, iCloud, Outlook, Yahoo and other IMAP accounts into one inbox with cross-language semantic search, shared team access and mail rules. It connects ChatGPT, Claude or other assistants via MCP, with separate permissions. From $4.99/month after a 7-day trial.
Meta's personal AI assistant app for iPhone and iPad adds scheduled tasks and reminders, email and calendar connection, and in-depth research reports in version 300.0.0.
Perplexity's Monitor API searches the web on a schedule and alerts agents via webhook only when relevant changes occur, using its own decision model to filter unimportant results. It turns search from one-off queries into persistent watchdogs.
The piece explains how Anthropic's Claude Code checks agent-written code through verification loops, describing the review steps that catch errors before changes land.
Seedeo MCP lets ChatGPT and Claude connectors call AI media generators Suno, Kling and Veo, so assistant users can create music and video from within chat.
A Stepback newsletter argues the vibe-coding revolution is coming to all gadgets, with consumers coding custom features for devices rather than just apps.
Webless Agents lets users describe an outcome in plain English; it builds a customer-facing agent grounded only in published content, with every instruction reviewable before launch, then answers questions and books meetings.
Omi CLI is a terminal coding agent that reads a repo, shows every edit as a diff for approval and runs tests until they pass, auto-selecting from 13 models under one plan; installable via npm.
Ulika is a visual editor for React, Vite or Next sites built with Claude, Cursor or Codex, letting anyone edit copy and images on the page and publish without touching code; open beta, free to try.
SitePoint benchmarks four local LLM serving tools — Ollama, LM Studio, llama.cpp and vLLM — for hosting a local model to several users, comparing throughput and behavior under concurrent load.
A command-line interface for Amp, a coding agent and development environment. Installable globally from npm with npm i -g @ampcode/cli; this release bumps the version with no listed changes.
An AI agent built the Loom package of AMD kernels in 4.3 hours, nearly doubling Inkling 2 Mini decode speed and hitting 5.9x llama.cpp prefill at an 18k-token prompt, against what an AMD engineer called super nascent day-one support.
Muru is a new AI-assisted Rust server from Graham Dumpleton, creator of mod_wsgi, aimed at serving ASGI and WSGI Python web applications in both development and production.
Pandex is an MCP server exposing 2.8 million European legal documents and EU law to AI agents, letting them query official legislation through a standard interface.
Snowflake now lets its enterprise AI agents produce maps as a native output, so agents can render geospatial results directly inside enterprise workflows.
The New Stack reports Nvidia used 20,000 training examples to fix mistakes in its internal AI and saw its coding score jump, detailing the data and training approach behind the improvement.
A Google guide for AI coding agents on adding Google Analytics to iOS and Android apps, covering setup steps so agents can instrument screens, events and conversions correctly.
AMD added support for its XDNA 2 architecture NPUs to Unsloth Desktop, aiming to let 100+ agentic AI applications run large models on PC neural processing units.
AI Context Flow keeps one shared memory over MCP so multiple AI tools and teammates work from the same version, with buckets by project or client, access controls and edit history.
OpenVorld is an open-source harness that gives AI agents a visible pixel-art village, where each agent has a name, role and brain (Claude Code, Codex or any OpenAI-compatible API) and asks before running commands. Free and local.
Shipyard is a local CLI and MCP server for Claude Code, Cursor and Codex that connects accounts once, then automatically sets up domain, DNS, hosting, database, email and payments for each new app in about an hour.
Stitch Loop is an experimental Google Labs product agent that connects to a GitHub repo and product data, proposes and builds fixes in a cloud sandbox, and opens pull requests for human review.
ClawDOS is a project that gets an AI coding agent running on vintage DOS hardware, demonstrated on a 1981 IBM PC XT. It shows agent tooling adapted to extremely constrained machines.