Startup Buzz

AITypeSafe AI's Jev routes each task to the best tool or model for an AI agent, according to @TypesafeAI's post. AIML API says its Jev Router connects agents to over 10,000 tools and models through one MCP and claims results 96% cheaper than Claude. The post is a promotion that invites users to try the router.

Why it matters: One MCP connection that reaches 10,000+ tools and models, with TypeSafe AI's Jev choosing the tool per task

AILauren Tan (@poteto) says her Grok-based bot Tibot monitors mentions and keywords on X, sends a sorted daily or hourly digest, filters spam, translates non-English posts, and can auto-block trolls and abusive accounts. She says Tibot never posts on her behalf, and she still writes her posts and articles by hand.

Why it matters: Sorted daily or hourly digest of mentions and keywords, with translation of non-English posts

AIVirtuals introduces the Virtuals App for Android, which lets users reach any asset, chain, and agent through its Butler agent. Butler can build, trade, hedge, and rebalance on its own from a phone. The app is in closed beta on Google Play.

Why it matters: Butler agent reaches any asset, chain and agent, and acts on its own from the phone

AIAtai Barkai announces OpenDuet, a self-hostable proactive AI agent that works with any agent harness and can be cloned as a template. The post lists live asset generation, human-in-the-loop approvals, computer use, and mobile, web, and channel support. It links to the GitHub repo and is built on CopilotKit and AG-UI. The quoted post from @claudeai separately describes Claude Duet, a research preview of a proactive agent app.

Why it matters: works with any agent harness and can be cloned as a template

AILARPCHECK says a paste-in URL, X post, or contract address is opened by an AI agent in a real browser, which reads the chain and project X account and returns a story score from 0-100 and a holder rating in about a minute. The site adds that users can now trade from the coin page, and that its token $LARPCHECK currently reads 66/100 "LIKELY REAL" with holders rated OK. The post notes that each check's first run is free.

Why it matters: AI agent opens a pasted CA, X post or link in a real browser and returns a 0-100 story score and holder rating in about a minute

AIA solo developer built a partial neuron atlas for Talkie-1930, a historical language model created by @status_effects and @AlecRad. The project was inspired by recent persona work from @Sauers_ and @jfonsecarivera. The atlas is available at talkieneurons.vercel.app.

AINace has introduced Drex for Search, a decision-powered retrieval tool for agentic search through its File Search API. Nace reports 62% fewer steps to find answers, 56% fewer tokens, and a 4% performance gain across Claude Code, Codex, OpenCode, Hermes, and PI. The figures come from six professional work benchmarks, including OfficeQA with 529 real-world queries.

Why it matters: Narrows retrieval by folder, file and page before passing relevant pages to the answering model

AIAditya Umale announces mausbot, a self-hostable proactive AI agent that works with any agent harness. The post lists live asset generation and editing, human-in-the-loop approvals, computer use, and mobile, web, and channel access. The code is in the OpenMausBot GitHub repo, and a hosted VM option is available at mausbot.com/pro.

Why it matters: works with any agent harness

AIEon launches the Data Resilience Initiative, letting companies apply to use its backup and recovery service free for 90 days. The service includes isolated daily backups, ransomware-scanning backups, and help restoring data after an attack. Eon's founder Ofir Ehrlich says any company can start deploying Eon today and calls on other cybersecurity firms to join.

Why it matters: Ransomware scanning runs on backup copies rather than live servers

AIA developer says they built a simple motion capture app using FlowHMR, ExMR, and GVHMR. The app includes basic foot locking and motion retargeting for characters with different body proportions. It also imports Mixamo .fbx, .pmx, and .vrm files.

Why it matters: Basic foot locking and motion retargeting for characters with different body proportions, with import of Mixamo .fbx, .pmx and .vrm files

AIOptimAI launches Router, an OpenAI-style chat completions API that lets builders choose a model that Router never swaps, with API keys that carry monthly caps and per-request charge limits. Providers can serve open models from their own computers through OptimAI Core Node, set their own prices, and earn credits for each completed request. MCP and x402 pay-as-you-go for individual requests are listed as coming soon.

Why it matters: Router never swaps the model a builder chooses, and providers set their own prices

AIADRs4AI is a repository template that stores human-AI deliberation as versioned Git records. Each question carries its options, the AI's recommendation, rationale, confidence, and falsifier, so a human answers by accepting or overriding. The project also offers a VS Code extension, an MCP server, and an iOS client for viewing and answering open questions.

Why it matters: Trails replace pull requests, with customized evals scored by agents on each code push

AIThomas Dohmke says Entire is shipping today as a Git forge built for collaboration between AI agents and humans. He says the team built GoEntire on Entire and invites developers to build with it.

Why it matters: A Git forge built specifically for collaboration between AI agents and humans

AIBrady Long says Autoharness is a plugin for Claude Code that learns skills from real sessions without manual SKILL.md writing. According to the post, it merges near-duplicate skills, rewrites skills when corrected, archives unused ones, and never touches skills the user wrote or installed. The GitHub README in the image describes it as an MIT-licensed tool, version v0.5.3, for Linux and macOS with Python 3.11+ and cites a CORE-Bench result rising from 42% to 78% with a different harness.

Why it matters: Creates and maintains Claude Code skills automatically from real sessions, without manual SKILL.md writing

AIEpoch AI has launched the AI Chip Imports explorer, which estimates AI compute arriving in each country quarter by quarter since 2022 using customs records. It estimates the US received 19.8 million H100-equivalents, or 72%, of the 27.4 million total through June 2026, while China, including domestic production, received 2.6 million, or 10%.

Why it matters: Estimates AI compute arriving in each country quarter by quarter from customs records

AIinClinico Agent, an agentic AI system, predicts clinical trial outcomes with an efficacy AUROC of 0.828 and a safety AUROC of 0.809. It is built on over 100,000 trial records and live ClinicalTrials.gov queries. The system is now available to industry partners, and a head-to-head prospective benchmark against five frontier LLMs covered 3,235 ongoing Phase I–III trials.

Why it matters: Predicts clinical trial efficacy and safety outcomes, with efficacy AUROC 0.828 and safety AUROC 0.809, built on over 100,000 trial records and live ClinicalTrials.gov queries

AIDesign Arena introduces Code to Video Arena, where two models each write HTML, CSS, and JavaScript to direct a 15–45 second film in a single webpage. Each film is rendered at 1080p and 60fps, with sound synthesized through the Web Audio API and normalized to the same loudness. A leaderboard is coming soon.

Why it matters: Models direct films entirely in HTML, CSS and JavaScript, with sound synthesized through the Web Audio API

AIUnibar announces a Mac launcher app built to run alongside T3 Code. Users can see all their agents and statuses with one keystroke, then reply to or steer them by voice or text.

Why it matters: One keystroke shows all agents and statuses, with voice or text replies

AIA solo founder says they built an AI chat interface at scrolltide.co that is animated and interactive, with a design they say does not look like a template. The post does not give pricing, features, or models.

Why it matters: Animated, interactive design that the maker says does not look like a template

AIA solo founder built substack-to-pdf, a site that converts a Substack link into an LLM-readable PDF or MD file for sharing posts in discussions. The output is also printable for humans, and agents can use the site directly from a link.

Why it matters: Accepts a Substack link and outputs a PDF or Markdown file, and agents can use the site directly from a link

AIGlasser has launched a pay-per-call data service that gives AI agents access to more than 2,000 paid API endpoints from over 80 providers. Agents can request data such as company records, SEO and social media metrics on demand, without a monthly plan for each provider, and see each call's price before running it.

Why it matters: Agents see each call's price before running it and can pick among providers selling the same data at different prices

AIPaperScorer, a NeurIPS '26 project, predicts whether an anonymized submission will be accepted, and its authors say it beats GPT, Gemini, and Jev with well-calibrated probabilities. The tool is available to try at paperscorer-ai.com, with details in a blog post and a thread.

Why it matters: Predicts acceptance from anonymized submissions with calibrated probabilities, per its authors

AIA solo founder built whiteknuckle.app, a driving app that routes big rigs and campers by difficulty rather than duration. The app warns about grade, tight turns, and clearances based on the rig's height and length, and models turning with truck bed length and hitch dimensions. The founder says the app drew 100 signups in five hours after a post in an RV Facebook group, and that it took about a week to build.

Why it matters: Routes ranked by difficulty, with warnings for grade, tight turns and clearances based on the rig's height, length, bed length and hitch dimensions

AIArtificial Analysis has released AA-3D Explorer, a browser tool for comparing how different models build the same 3D scene. Users pick a task from one of seven domains and switch between models to view each result. The scenes can be steered, flown through, and clicked, the same way the human panel reviews them before voting.

Why it matters: Lets users steer, fly through and click each model's 3D scene, as the human review panel does

AIA solo developer has built Strike Copilot, a free tool that backtests any Hyperliquid trader before a user copies them, replaying the same 90 days at 10x and 5x leverage. The post says the 10x copy was liquidated 27 times, while the 5x copy was liquidated 4 times and made more than double the profit. The accompanying article says the toolkit runs inside Claude Code on the user's own computer, requires explicit confirmation before any account change, and never places or withdraws orders.

Why it matters: Backtests any Hyperliquid trader before copying, replaying the same 90 days at different leverage levels

AIAlpha Capital introduces My Alpha, an AI-powered trading journal that helps traders track, analyse and improve their performance. The tool is free for Alpha account holders, according to the post.

Why it matters: Free for Alpha account holders

AIGhostHand lets agents use a real iPhone connected to a Mac. The agent sees the screen over USB and taps, swipes and types over Bluetooth, so it can use any app, including those with no API. The post links to ghosthand.dev.

Why it matters: Agents can use any app on a connected iPhone, including apps with no API

AIClypmint introduces a b-roll search tool that finds real footage of a script's exact person, place, or moment for 1¢ per search. The company says it ranks first on B-Roll Bench accuracy, searches the Internet Archive, Wikimedia, NASA and the Library of Congress in one query, and finds exact moments inside YouTube videos. The open-source plugin works with Claude Code, Codex or any agent, and users pay only when it finds a result.

Why it matters: Searches the Internet Archive, Wikimedia, NASA and the Library of Congress in one query and finds exact moments inside YouTube videos

AIScholarXIV announces ScholarXIV for Astronomy, a product where users search stars, galaxies and other objects to see their type, position, distance, sky images, wavelength breakdowns and light curves. Users can also explore exoplanets and solar system bodies such as planets, comets and asteroids, with an AI agent to help interpret results.

Why it matters: Search by object to see type, position, distance, sky images, wavelength breakdown and light curves, with survey switching and download

AIBrowserbase introduces the BELL stack for building personal agents, pairing Browserbase browser infrastructure, Vercel's Eve harness, Linq for iMessage, and Stripe's Link for agentic payments. The stack is open source, and Browserbase is funding it with $10,000 for agentic payments.

Why it matters: Combines Browserbase browser infrastructure, Vercel's Eve harness, Linq for iMessage and Stripe's Link for agentic payments in one stack

AIA developer shares a web tool that adds swash glyphs to existing typefaces, currently limited to Google Fonts. The post says it was built with Argon, Antigravity and AI Studio, and links to type-renderer.ai.studio.

Why it matters: Draws swash glyphs on existing typefaces, currently limited to Google Fonts

AI0xArchive has open-sourced a quant research skill that lets coding agents backtest trading strategies on its Hyperliquid data. The skill compares each strategy against three fake baselines and logs every trial. Its 7-day momentum example reports a Sharpe of 2.43 in-sample and 1.81 out of sample, with a deflated Sharpe of 0.93 that falls below its 0.95 bar, so it is labeled as holding out of sample rather than an edge.

Why it matters: Tests each strategy against three fake baselines and logs every trial

AIANYbotics launches ANYmal X, which it calls the world's first legged inspection robot certified for explosive atmospheres, meeting ATEX in Europe and IECEx globally. The company says the robot runs autonomous inspections in oil, gas, and chemical plants on its Shift operations platform, with more than 50 customers running over 120,000 inspections a month.

Why it matters: Certified for explosive atmospheres under ATEX in Europe and IECEx globally

AIproc is a semantic command-line tool for managing processes on macOS, Linux and Windows, letting users target processes by port, PID, name or path. Its commands include proc kill :3000, which ends whatever is listening on port 3000 with SIGKILL, and --dry-run previews the action. Every command prints one JSON document with --json, and exit codes 0 through 4 keep fixed meanings for scripts and coding agents.

Why it matters: targets processes by port, PID, name or path, with --dry-run previews and JSON output on every command

AIMastra announces Knowledge, a standalone service that holds company and team context and gives it to agents when needed. It connects to systems like Slack and Monday and keeps itself current, with separate knowledge bases per function and permissions at individual, team, and org levels. Mastra is opening early access to eight more design partners, first come, first served.

Why it matters: Separate knowledge bases per function with permissions at individual, team and org levels