Builder Radar
A monthly brief on what AI products people are launching. Trends, emerging categories, and products worth watching.
We looked at 551 AI products in September. The sharpest signal came from one launch: TypeSafe released Jev, a closed model that answers with typed decisions instead of text. Within days open-source copies were on GitHub, one of them past 26,000 stars. Around it, the consumer layer kept renting new model names as domains. European inference providers started selling who owns the GPUs, not just where they sit.
551
Products spotted
Open copies of a closed model, within days
Top trend
A closed model launch got open copies within days
TypeSafe AI released Jev in mid-September: a hosted model that answers with typed decisions (a choice, a score or a yes/no with a confidence figure) instead of generated text. It costs $0.042 per million input tokens with output free. New accounts come through an early-access list. Within days at least seven open-source versions were on GitHub. Laya, built on ModernBERT for 100+ languages, had passed 26,000 stars when we looked on 27 September, nine days after its first commit. Kev, from Jared Palmer, is a family of Qwen-based models whose server matches TypeSafe's own API: the official SDK works against it unchanged. DataCamp had a roundup of the open alternatives out by 23 September.
A new model now gets a ring of look-alike domains in its first week
At least 35 sites in September carried someone else's model in the domain: Seedance, Seedream, MiniMax, Flux, Kling, Ideogram, Wan, Hailuo and Lyria among them. In August we counted at least 20. Most let you try the model in a browser and sell credits on top. The new case was Jev. Seven separate Jev-named sites reached us in two weeks, one of them openly calling itself an unofficial relay, for a model its maker still gates behind early access. The domain layer used to form around image and video releases. It now forms around developer-facing models too.
Gateways split by what they disclose
The OpenAI-compatible gateway was August's most-copied product. It kept coming in September. This month the field split clearly. One group publishes a named company, a licence and list prices: AISIX from the Apache APISIX team is Apache-2.0 with a free managed tier. Kontinent in Berlin routes only to EU providers. The other group advertises Claude and GPT well below the model makers' own prices, from "70% off" to plans promising many times the usage of a Claude Max subscription. Anthropic and OpenAI do not sell at those prices, which leaves the buyer to guess where the capacity comes from.
European inference now sells ownership, not only region
Two September launches made jurisdiction part of the product rather than a line on a policy page. Kontinent tags every model in its API with a sovereignty tier that separates EU-owned providers from the EU region of a non-EU parent. A caller can filter on it in code. Heabsy, from Slovakia, serves its main model on GPUs it operates inside the EEA, under a company with no US parent. It never logs prompts. Models it routes to other providers are labelled as routed. The pitch has moved from where the data centre is to who owns the machine.
Agents are getting their own drive and their own keychain
PuppyOne gives agents a shared, Git-versioned workspace with access scoped per agent, reachable over MCP, a CLI, Git or SSH. Aident Loadout keeps the credentials for 1,000+ apps in a vault. An agent acts through OAuth or delegated keys without handling the raw secrets. Every action leaves an audit receipt. In August the agent stack grew memory and package managers. In September it grew the plumbing a careful team asks for before an agent touches real accounts.
Worth a look
New AI products and tools that caught our eye in October 2026
TypeSafe Jev β
Hosted decision model that answers with typed choices, scores and yes/no instead of text
$0.042 per million input tokens with output free. Every answer carries a confidence figure your code can threshold. Its launch set off the month's wave of open copies.
Laya β
Open-source multilingual decision model built on ModernBERT
Passed 26,000 GitHub stars nine days after its first commit. Answers choice, score and yes/no questions in 100+ languages in one forward pass. Apache-2.0, installs with pip.
Kontinent β
OpenAI-compatible router over EU providers with a sovereignty tier on every model
Lets a caller filter by who owns the provider in code, instead of trusting a policy page. Prepaid in euros, from Berlin.
Heabsy β
EU inference API on its own GPUs inside the EEA, with no prompt logging
Its flagship Qwen3.8 27B costs $0.04 per million input tokens. Models it routes to other providers are labelled as routed, with a note that their compute may leave the EEA.
PuppyOne β
Shared, Git-versioned context drive for AI agents
Scoped access per agent over MCP, CLI, Git or SSH. Apache-2.0, from Singapore. It is the second product built for agent memory we have listed, after screenpipe.
AISIX β
Apache-2.0 AI gateway in Rust from the Apache APISIX team
One OpenAI-compatible API across 100+ providers plus an Anthropic-compatible endpoint, with a free managed tier at 100,000 requests a month.
What people are building
Based on ~551 AI products spotted across directories and launch platforms in October 2026
We looked at 308 AI products in August. Eight of them were the same piece of infrastructure: an OpenAI-compatible gateway that puts GPT, Claude and Gemini behind one key and a prepaid balance. At least twenty more were websites named after a model somebody else built. The more original work sat around the agent: what it costs, what it remembers and how you prove it did the job.
308
Products spotted
OpenAI-compatible gateways
Top trend
The AI gateway is the most-copied piece of infrastructure right now
Eight of the 26 infrastructure products we tracked in August were OpenAI-compatible gateways: one base URL, one key, a model list spanning GPT, Claude, Gemini and Grok, prepaid credits and a usage log. On the landing page they are nearly indistinguishable. The difference shows up when you visit logged out. Some publish every model with its per-token price, link working docs and name the company taking your money. Others show a single sentence and put everything behind a sign-in. For a product that is a thin layer over someone else's models, that public detail is most of what there is to compare. Before topping up a balance, check three things: a published price list; a named operating company with a jurisdiction; whether a first-party model is priced far below what its own maker charges.
Cutting the agent bill is turning into its own layer
July brought three tools for spending less on tokens. August added four more, each working on the bill from a different side. Varion compacts history and prunes tool schemas before a request reaches the provider, returning a header so you can check the savings yourself. Vynaris routes each request to the smallest model its evals support and returns a receipt: model served, provider cost, routing fee. Weckr wraps your existing client to track cost and margin per end customer. Openspender goes furthest. It gives agents their own wallets, capped per request and per day, to pay for models and tools over the x402 and MPP payment protocols. Nobody provisions a key for each service. Two months in, this looks less like a feature of other products and more like a layer of the stack.
Agent evaluation now scores the path, not the answer
BenchGen builds simulated copies of enterprise systems and scores an agent's whole trajectory across a multi-step task, then turns the failures into training data. Its argument is that grading the final output tells you little when an agent took dozens of steps to get there. HAIEC comes at the same problem from the compliance side: adversarial tests against live endpoints, with signed evidence mapped to SOC 2, the EU AI Act and New York City's law on automated hiring tools. SandBase Harness records agent sessions with approvals and replay. Three products, one shared assumption: checking that the answer looked right is no longer enough.
Agents are getting a memory and a package manager
screenpipe records selected screen, audio and meeting context locally and serves it to agents over MCP and a local API. Its GitHub repo passed 21,000 stars when we looked in August. Linkly AI indexes local documents, images, audio and video into a search engine agents can query. Two skill registries arrived as well. agentskills.codes listed more than 19,000 installable skills for Claude Code, Codex and Copilot in August, each with source-health scans. Another skills marketplace was submitted four times by the same sender. The shared idea: an agent should carry context between sessions and pick up new capabilities the way a developer installs a package.
The model-name website has become the default consumer launch
Image and video tools were 81 of the 308 products we looked at in August, the largest group again. At least 20 sites across image, video and music carried someone else's model in the domain or product name: Wan, Seedance, Nano Banana, Hailuo, MiniMax, Sora, Flux and Lyria among them. The shape has spread to music, with MiniMax Music and Lyria front-ends next to a run of consumer song generators. Most are a credit subscription over an API anyone can call directly, which makes them quick to launch and easy to copy. The name is the marketing: it catches search traffic for a model in the week it ships. Alongside them ran a long tail of one-job web utilities: crop an image to a circle, lower an image's quality, turn an image into text.
Worth a look
New AI products and tools that caught our eye in August 2026
Solheim AI β
Private EU-hosted LLM instances billed at a flat monthly fee instead of per token
The one new billing shape of the month. Each project gets an endpoint sized by concurrency and context window; a busy month costs the same as a quiet one. Open-weight models only, from EUR 15/month in August 2026.
Openspender β
Managed wallets that let AI agents pay per request for models, search and data
Built on the x402 and MPP machine-payment protocols, with spend caps per request, per day and in total, enforced before a payment is signed.
BenchGen β
Benchmarks AI agents in simulated enterprise environments
Scores the whole trajectory of a multi-step task rather than the final answer, then turns the failures into training data. Can be deployed air-gapped.
screenpipe β
Local-first screen and audio capture that gives AI agents searchable memory
Past 21,000 GitHub stars in August. Serves context over MCP and a local REST API. An agent can query what you worked on without the data leaving your machine.
Anycloud β
CLI and Python SDK that run AI jobs, services and VMs across cloud accounts you already own
Orchestrates rather than resells: credentials and billing stay in your own AWS, Azure, GCP, Lambda or Vast accounts. A different answer from the GPU marketplaces to where a job should run.
agentskills.codes β
Open registry of installable skills for Claude Code, Codex, Copilot and other agents
More than 19,000 skills in August, each with source-health and capability scans, plus a working API. The clearest example of the month's package-manager idea.
What people are building
Based on ~308 AI products spotted across directories and launch platforms in August 2026
Five European inference providers landed in July, from Italy, Norway and three from Germany. That is more than any month since March, and they are no longer pitching sovereignty as a compliance checkbox. They publish per-token rates, ship OpenAI-compatible endpoints, and expect you to compare them to Together and Fireworks on price. Underneath that, the image and video wrapper wave grew rather than slowed, and a new tell emerged for spotting it.
249
Products spotted
European inference providers
Top trend
European inference stopped selling compliance and started selling price
regolo (Italy, run by the hosting company Seeweb), SimpleLLM (Germany), KugelAudio (Berlin, text-to-speech), IONOS AI Model Hub (Germany) and WAYSCloud (Norway) all arrived within a few weeks of each other. What changed is what they lead with. GDPR and data residency are still on the page, further down. At the top is a per-token rate and an OpenAI-compatible base URL, which means they expect to be compared against Together and Fireworks on cost first and jurisdiction second. GreenPT, which arrived at the start of August, publishes a per-model price table down to cached-input rates.
KugelAudio trained its own voice models instead of reselling
Most European inference providers rent GPU capacity and serve open-weight models. It is a real business with a thin moat, and it describes almost every provider in this group. KugelAudio is the exception: it trained its own text-to-speech models in Berlin, published one under MIT, and competes on latency against ElevenLabs and Cartesia. One company is not a trend. It is the harder path, and nobody else in July took it.
The image and video wrapper wave got bigger, and easier to identify
Image and video generators were 72 of the products we looked at, comfortably the largest group and larger than June. The tell has sharpened into four things that travel together: the domain carries a major vendor's model name, pricing is credit subscriptions in the $9 to $40 band, an API is advertised but never linked to documentation, and no company is named anywhere. Seedance, Hailuo, Wan, Nano Banana and Flux each appeared across multiple unrelated domains. One went further and presented ByteDance's Seedance 2.0 as its own proprietary model, without mentioning ByteDance anywhere on the page.
Three ways to spend less on tokens, none of them switching provider
KV Cache Store precomputes the KV cache for long prompts so you skip prefill on repeated context. Tokenade sits between a coding agent and the model and rewrites requests before they are sent, billing on tokens saved. Parity Layer shadow-tests cheaper models against your baseline and routes to them once your confidence threshold is met. Three products, three mechanisms, all arriving the same month. Three is not a category yet, but a year ago the answer to a large bill was to find a cheaper provider, and none of these do that.
Agent tooling moved down the stack
None of the agent products that landed were frameworks for writing agents. Wire stores documents and records that agents query over MCP, running hybrid retrieval internally. ego lite is a browser where agents work in isolated spaces alongside you, reusing your real logged-in sessions instead of re-authenticating. Redcells runs adversarial probes against models you own. Tura is a local coding agent under AGPL. The framework layer looks settled enough that the interesting work has moved to what agents run inside and against.
Worth a look
New AI products and tools that caught our eye in July 2026
regolo β
OpenAI-compatible inference API run on Italian infrastructure with zero data retention
Operated by Seeweb, an Italian hosting company that has been around far longer than the AI boom. Serves open-weight models from EU data centres, and rents GPUs hourly if you want to deploy your own.
KugelAudio β
Real-time text-to-speech in 26 languages, trained and hosted in Europe
Trained its own voice models and published one under MIT. Priced per audio minute and built for voice agents, where the constraint is time to first audio.
KV Cache Store β
Build, share and reuse precomputed KV-cache artifacts to skip redundant prefill
Treats the KV cache as a distributable artifact with a registry, digests and model metadata. An unusual place to look for savings, and the open-source CLI works on its own without the hosted registry.
Wire β
Hosted context containers that agents query over MCP, with hybrid retrieval built in
Delivers retrieval as MCP tools, so an agent gets scoped access without you building the layer. BM25, embeddings, HyDE and reranking all sit behind one search call.
ego lite β
Chromium browser where humans and AI agents browse in parallel through isolated Spaces
A local-first answer to headless browser infrastructure. Agents reuse the sessions you are already logged into, driven by an MIT-licensed skill that works with several coding agents.
Runware β
Unified API for image, video, audio and 3D generation running on custom inference hardware
Owns its hardware and scheduler, which is the line between this and the media-generation wrappers filling the same category. You can also upload custom LoRAs and checkpoints.
What people are building
Based on ~249 AI products spotted across directories and launch platforms in July 2026
Previous editions
Get Builder Radar in your inbox
Monthly brief, no spam. Next edition: October 2026.
Is your product missing?