New AI products and companies, with their launch sources.
64 launches
Recent voices, likes and reposts, with a small boost for editorial picks. Attention fades with a 24-hour half-life. This measures attention, not product quality. Scores reflect current attention for launches in this edition. Equal scores use the latest discussion, then a stable order.
AIA solo vibe coder built skill-issue.in, a directory of viral skills for AI agents covering design, frontend, backend, motion graphics, automation, social media, and finance. Users copy an install command and run it on their agent, and can save favorites for later.
Why it matters: Each skill comes with an install command to copy and run on the agent
AITypeLLM, an open-source project, introduces what it calls the first workflow model, which runs an entire workflow in a single API call. It turns a JSON Schema into constrained decoding on SGLang, so fields can only return values from their allowed enums. The listed JevBench results show 84.42% accuracy without thinking and 98.70% with thinking for TypeLLM running on Qwen3.8-27B, a checkpoint used without fine-tuning.
Why it matters: Runs a whole multi-step workflow from one JSON Schema as a dependency graph in a single API call
AISolidCraft is an open-source parametric 3D CAD application built in pure Rust, available for Windows, macOS, Linux and in the browser. Its built-in MCP server lets AI agents model, measure and export parts headless or in the running app. The author says it is early but usable for simple parts, and that 75 of 76 public reference parts match their analytic or FreeCAD-computed volume, area, centre of mass and bounding box.
Why it matters: Built-in MCP server lets AI agents model, measure and export parts headless or in the running app
AIBeancat launches at beancat.ai as an AI agent for media production that handles prompt engineering and asset organization across generation tools. The team built it as an internal tool to make original animated episodes, which drew millions of views. The service is free to start.
Why it matters: Handles prompt engineering and asset organization across multiple generation tools
AIWakeBee is a menu bar app that keeps a Mac awake while agents, builds, or renders run, including with the lid closed. Users set their own limits before closing the lid. The app is open source.
Why it matters: Keeps a Mac awake with the lid closed, with user-set limits
AIArtificial Analysis invites users to try AA-3D Voting, which shows two 3D scenes built by unnamed models from the same brief. Users explore each scene for 30 seconds, then pick one to reveal which models made them.
Why it matters: Users pick between two unnamed models' 3D scenes, and the models are revealed after the vote
AIIMD launches a form where users type any request and 1,000 AI agents work to build it, such as a coin, Uniswap hook, company, oracle, website, or audit. The post says the current price is 0.5 $IMD per job, about $5.
Why it matters: Dispatches a request to 1,000 AI agents for 0.5 $IMD per job
AIbuy, a marketplace where AI agents shop for themselves, is now live on Celo. Agents can pay in USDC, USD₮, or USA₮ for more than 700 tools, including compute, data, and APIs. The launch partners are SQD, Cencori, and Chainstack.
Why it matters: Agents pay in USDC, USD₮ or USA₮ on Celo for 700+ tools
AICesium announces the Cesium ion MCP Server, which connects 3D geospatial data directly to AI tools so users can build, query, and test without leaving their code editor. The release uses the Model Context Protocol (MCP).
Why it matters: Lets users build, query and test 3D geospatial data from inside their code editor using MCP
AIMigma launches a hosted version of Mailcheap, an open-source newsletter tool, so users need no SES approval, MongoDB, Redis or server. Users sign in with Migma, add a website and domain, write or auto-design an issue, and send, with Migma handling sending, warmup, compliance and design. The post says pricing matches AWS, with code published on GitHub.
Why it matters: hosted version that handles SES approval, sending, warmup, compliance and design, at the same price as AWS
AIA solo developer built a chess game into Claude Code so users can play against bots while the agent runs. The post links to the project on GitHub as chessmod.
Why it matters: Chess games playable inside Claude Code while agents run
AIGoogle Gemma credits developer @_arahim_ for building and open-sourcing DigUp, a Mac app for searching files by their content rather than their names. DigUp runs Google DeepMind's EmbeddingGemma 2 on-device and supports searching video moments, podcast minutes, PDF pages, photos and code in any language. The app is free and open source, with its code on GitHub.
Why it matters: Searches files by meaning and content rather than file names, running EmbeddingGemma 2 locally
AIAgor, a hosted platform now in open beta, runs existing agent harnesses such as Claude Code, Codex and Gemini on a shared spatial board. Teams pick the agent and model per task, bring their own API keys or subscriptions, and pay providers directly without token markup. Recurring jobs can be packaged as AI teammates with set roles, versioned knowledge documents, MCP tool connections and optional schedules, and can take requests from Slack, Discord, GitHub and Shortcut.
Why it matters: Runs multiple agent harnesses on one shared spatial board, with per-task agent and model choice
AIAPUS AI has open-sourced Reflex Agent under the MIT license, a personal agent built on its Reflex decision model and Continuum memory engine. APUS AI says its internal tests show response speeds up to 3 times faster than Claude Code and Codex. The product ships as a desktop app, a web app and an SDK.
Why it matters: Built on the Reflex decision model and Continuum memory engine; APUS AI says internal tests show up to 3x faster response than Claude Code and Codex
AIYap is a tool that gives AI agents the ability to record a screen, use a camera, and process videos. It has no UI or built-in AI, offering only those primitives. The post says it works with Claude, Codex, and other agents.
Why it matters: Gives agents screen recording, camera and video editing primitives with no UI or built-in AI
AIoBrain launches docs.obrain.cloud, documenting a biological connectome translated into a quantized, deterministic spiking neural network deployed on Arc. The company says it open-sourced the code, with licensing terms documented publicly, so researchers and developers can examine, reproduce, and build on the architecture.
Why it matters: a biological connectome translated into a quantized, deterministic spiking neural network deployed on Arc
AIContext.dev has open-sourced sgrep, a Rust-based local semantic code and document search CLI that uses a roughly 32MB static embedding model and works offline after the first search. On SWE-bench Verified with GPT-6, the author reports that sgrep's pipeline mode cut LLM API costs by 21.3% versus ripgrep and solved 70 of 100 attempts against 68. The author also notes that sgrep trails a semantic mode of another tool on CoSQA snippet retrieval.
Why it matters: Runs a roughly 32MB local embedding model that works offline after the first search, so code never leaves the machine
AISolo developer Ian released Covenant 0.1.0 on Hex under the MIT license, a Contract layer for Elixir. It reads component contracts written in Markdown and YAML under contracts/, and fails the build when obligations stop being proved by tagged tests. The post also describes mix covenant.check for contract coverage, decision records with checkable conditions, and mix covenant.mutate for mutation testing credited to individual obligations.
Why it matters: Reads component contracts written in Markdown and YAML under contracts/ and fails the build when obligations are not proved by tagged tests
AIDeslop is an npm package, ai-deslop, that provides Oxlint rules rejecting low-evidence AI-generated TypeScript patterns, such as chained type assertions, object-typed parameters, and unknown return types. The rules are vendored from dmmulroy/anti-slop under the MIT license, alongside additional deslop/* rules. Users enable them through a recommended config that works with Bun, Deno, or Node.js.
Why it matters: Rules that reject low-evidence AI-generated TypeScript patterns such as chained type assertions, object-typed parameters and unknown return types
AIMermoz introduces an agent-native mail client for macOS, promising a faster alternative to Apple's Mail.app. The beta is open to the first 100 people who register at mermoz.xyz.
Why it matters: agent-native mail client for macOS
AILiteLLM has open-sourced agentchat, a Python SDK that connects an AI agent built with any framework to Slack, Teams, iMessage, WhatsApp, and Discord. The SDK handles OAuth, webhook signatures, socket connections, deduplication, and session isolation, so developers can focus on agent logic. The post describes a pipeline of channel adapters feeding a core that deduplicates, locks sessions, and dispatches to an application handler, with replies sent back through the original channel.
Why it matters: Handles OAuth, webhook signatures, socket connections, deduplication and session isolation for agents built with any framework
AIA solo developer releases moo, a native AppKit launcher for macOS that avoids Electron and is free and open source. One hotkey searches apps, files, math, clipboard, windows, and AI, and it installs via Homebrew with brew install moomoi/moo/moo.
Why it matters: Native AppKit build without Electron
AIAgent37 (YC F26) says its sandboxes for running AI agents cost under 10% of most sandbox providers, averaging $0.37 per sandbox per month. The company says it now offers a free plan, with over 90% of new signups using it to host agents such as Hermes, OpenClaw, and Claude Code, at infrastructure costs of $0.05 to $0.10 per free user per month.
Why it matters: Sandboxes averaging $0.37 per sandbox per month, under 10% of most sandbox providers' costs, according to the company
AIAtticus is a free, open-source hold-to-talk voice assistant for macOS that runs through Claude Code on a paid Claude Pro or Max plan. Simple commands such as opening apps and moving windows skip the AI and run instantly, and the source reports about 0.14 cents of plan usage per AI ask. Apple's on-device speech keeps audio on the Mac, and risky commands like deletions and purchases require user confirmation.
Why it matters: Simple commands such as opening apps and moving windows run without the AI
A first public release of a standalone AI app, API, open-source tool or hardware product from a startup or independent maker. Betas, waitlists and preorders are labelled. Routine updates, integrations, redesigns, fundraising and relaunch campaigns stay out.
“New company” requires evidence of a company’s debut with its first product. Otherwise a qualifying standalone launch is labelled “New product.” First spotted is the date of the earliest launch report we have, unless the launch date is explicitly established.
Our sources and reviewed suggestions determine coverage; this is not an exhaustive record of every launch. Buzz measures attention. Source counts describe coverage. Joining the tracking list gives no ranking advantage.
Looking for funding and other startup news? Visit the Startups news tab.