Yesterday's Top Launches: 5 Tools from September 12, 2026
New developer tools are split over whether AI agents should run in the cloud, on phones, or not use models at all, and Typewise Nova applies the agent approach to customer support by resolving tickets instead of deflecting them.
Yesterday’s slate of new developer tools leaned hard into agents, though the more interesting story was the disagreement underneath. Two of the five assume intelligence lives in the cloud and want you to watch it closely. One argues the opposite, that the model belongs on the phone and should never phone home. Another skips models entirely and audits the page you already shipped. Voting was still open when I wrote this, so treat the ordering below as my own read rather than a settled leaderboard.
Typewise Nova
Customer support is where AI agents get sold most confidently and deliver least convincingly, and Nova is a direct attack on that gap. The pitch is resolution over deflection: instead of bouncing a ticket, it looks up the order, applies your policy, and closes the request across email, chat, WhatsApp, social, and voice, then hands judgment calls to a person with the full thread attached.
What stands out is Nova itself, an operator you talk to. “Create a specialist for billing and payments” drafts one that reads Stripe and your order system, with refunds over a certain amount held for human approval. Setup is conversational, claimed at around 15 minutes.
The pricing model deserves attention. Unresolved requests are free, a partial hand-off counts half. That ties revenue to actual outcomes rather than volume, which is rare in this category. Beurer reports 90% autonomous resolution, Lehner Versand 95%. I’d want to see how the plain-language setup holds up once 3,500 integrations meet a genuinely messy enterprise stack, but the incentives here are pointed in the right direction.
AI Observability by OpenObserve
Agentic apps are expensive and opaque, and OpenObserve’s framing nails it: your agent cost $40 and took 34 seconds, but why? A single request fans into hundreds of spans, and per-span metering punishes exactly that shape of workload. This bills per gigabyte instead.
The Agent Graph maps every model, tool, and datastore a request touches, color-coded by error rate, and session replay shows where latency and money actually went. Evaluations run continuously on live traffic with LLM-as-judge, and reviewed traces can be distilled into datasets for testing future versions. Because LLM traces land beside your logs, metrics, and RUM, a slow agent can be traced to the pod or vector store behind it without a second tool.
It is a crowded space, and OpenTelemetry-native rivals are numerous. Bundling AI observability into the same platform as everything else is the real differentiator, particularly for platform and SRE teams already drowning in dashboards. Self-hosted is free to 50GB.
Desert Ant Labs
A pleasant counter-current: instead of one giant model doing everything, Desert Ant Labs publishes a family of small, task-specific models that run on-device through a single native SDK. Voz transcribes ten minutes of audio in two seconds on an iPhone. Uhm strips filler words. Redact filters PII locally before anything leaves the phone. Schemer turns unstructured text into typed JSON.
The economics are the hook. No tokens, no per-use billing, free up to 100k monthly active devices per platform, so you never meter your users. Catalog quality is uneven, and several entries are one-liners with no public benchmarks, so treat the lesser-known models as unproven until you test them yourself.
FreeScan.app
Paste a public URL, get 40 checks across SEO, AEO, GEO, security, accessibility, and design. No signup. Results come back as four category scores plus Fixes, Opportunities, and Insights, each finding paired with evidence and a reason it matters.
The genuinely useful bit is the handoff. Export the workspace as Markdown, or connect through MCP so a coding agent can read your private findings and request a rescan. Some teams describe shipping straight off the fix list.
FreeScan is upfront that this is a signal tool, not penetration testing or accessibility certification, and it isn’t a substitute for strategy. At $19/month, Pro adds site-wide audits, a fix board, weekly runs, and uptime monitoring.
Modeinspect
Most software gets designed twice, once as a picture and again in code, with intent drifting in between. Modeinspect puts a design canvas on top of your actual codebase, with your real components, tokens, live data, states, and breakpoints, then writes changes back as scoped, type-safe diffs you can open as a PR.
The claim is 45+ days from design to ship collapsed to about 10. Components are 1:1 rather than redrawn look-alikes, and tokens are enforced so nothing off-system sneaks in. Quentin Le Bras of Prelude says his designers explore on the real codebase and open the PR themselves.
This only pays off if your design system is already disciplined. Point it at a codebase full of hardcoded colors and magic numbers and the enforcement promise gets thin fast. For design engineers at that level, though, it removes a handoff rather than adding a tool.