Two things are happening at once in the recent wave of AI releases. Underneath, agents are getting routers that collapse dozens of vendor accounts into a single key. On the surface, AI features are arriving inside ordinary applications, mail clients and desktop tools rather than as separate chat assistants. This piece covers both ends.
One key instead of forty accounts
The plumbing problem for any serious agent is that every capability comes with its own vendor, key, SDK and billing relationship. Monid describes itself as OpenRouter but for agent tools: one base URL and one key reach more than 2,000 tools across 72+ providers, covering web search and scraping, people and company enrichment, social platforms, reviews and market data, and video, image and voice generation. It also publishes a SKILL.md that an agent can fetch and install for itself.
Orthogonal takes a similar approach with a unified API, an MCP server and a CLI over 60+ providers and more than 1,000 endpoints, and you pay for the calls you make. It publishes a skill.md you paste into Claude Code, Codex, Cursor or OpenClaw setup so the agent can configure itself without you reading documentation. The honest caveat with any router is that provider quality behind it is uneven.
Search deserves its own mention because it is the call agents make most. Keenable is independent web search infrastructure for AI, with its own crawler, index and ranking that improves as more agents use it, and it publishes its numbers: 100B+ documents, under 250 ms p95 in US East, and pricing from one dollar per 1,000 requests at 100 RPS or more. It also runs an agent storefront where an agent can buy access autonomously or with a human approval, which is a preview of how agent-facing commerce is likely to work.
A shared memory for agents themselves
Askpod is a neutral shared corpus of firsthand observations written and read by agents. The premise is simple: one agent should not have to rediscover what another already learned about an API, a product or a service, and agents need somewhere to compare tools that is not a vendor-written comparison page where the vendor ranks first. The corpus is served as markdown, JSON and llms.txt so an agent can read it in whichever format it prefers.
AI arriving inside ordinary applications
The more interesting consumer trend is AI appearing as a feature rather than a destination. Penguin Mail is a mail and calendar app for Linux and macOS that brings Gmail, Microsoft and any IMAP, POP3 or JMAP account plus contacts into one application on your own computer. It is free software under GPL-3.0-or-later, syncs from the system tray on Linux, understands Gmail categories, and includes an optional assistant that summarizes threads, drafts replies and manages the calendar. The assistant being optional is the detail worth noting: the mail client works without it.
Radix is an agentic workspace for enterprise knowledge work, delivered as a desktop app. You pick or ask for a workspace, describe the result you want, and it reads, searches and writes until it is done and puts the findings into a page. It reads Word, Excel, PowerPoint, Pages, Numbers and Keynote, asks before opening anything you have not shared, and never changes the original file. It also keeps notes across one continuous conversation, so a topic from weeks ago is still reachable.
Narrow jobs done in one step
Some of the more useful recent releases do exactly one thing. Premortem red-teams startup ideas: you paste a concept and six agents attack it simultaneously across market, technology, competition, unit economics and more, and a seventh writes the memo. It is not a substitute for a domain expert, but as a cheap way to find the obvious hole before committing a quarter to it, it is a reasonable first pass.
Coverage Cat is an AI-native insurance broker for high-earning households, covering home, umbrella, auto and renters. You can compare options in the portal with a licensed broker that states it does not sell leads, or let an agent do it: the site publishes an agent card for PACT-capable personal agents, an MCP manifest, llms.txt and agents.txt. This is one of the clearer examples of a consumer business designing for agents as customers rather than tolerating them.
Creative work in one window
RxFilm Studio is a native macOS app that scores, narrates, captions and renders film, with an agent in the director's chair. One window covers generated music cues, multi-speaker narration tuned line by line, transcription with translation and retiming exported as VTT or SRT, stills and reference art, and a real timeline that renders up to 4K at 60 fps. Keeping the whole pipeline in one application is the appeal; the cost is that it is macOS only and the agent behaviour depends on the model integrations it ships.
What this batch tells you
The router layer is consolidating fast, and if you are building an agent that needs search, scraping or enrichment, going through Monid, Orthogonal or Keenable is now cheaper than integrating vendors one at a time. The application layer is moving in the opposite direction: fewer standalone assistants, more AI features inside tools people already open every day, usually with the option to turn them off.
