Most agent launches stay inside the editor. This batch does not: it points the agent at the desktop, the phone, the kernel and the application itself, and in one case at your infrastructure bill. These are the launches where the agent's scope is the machine rather than the repository.
The whole desktop as a tool surface
Moching is the maximalist version: a desktop AI operator with 219 built-in tools covering screen perception, keyboard and mouse control, browser automation, an Office suite and AI media. One sentence in, finished work out. It writes code through to compilation and build with LSP-level understanding, drives Word, Excel, PowerPoint and PDF from natural language, and can manage many machines at once for batch operations. The privacy story is the interesting part given the scope: local SQLite storage, data that stays on the device, your own API keys, and support for any model API as well as locally deployed models. It runs on Windows and macOS.
Kryzz AI is the same instinct in a much smaller package: a private, local-first Android AI workspace with chat, voice and autonomous agents, running on your own API keys. There is no account, no subscription, no telemetry and no bundled keys, and conversations stay on the phone. It is about a 19 MB APK built with Kotlin and Jetpack Compose, Android 9 and later, MIT licensed, with a workflow picker covering Auto, Deep search, Wide search, Document, Spreadsheet, Database and Code, plus a voice session overlay.
Housekeeping for machines running agents
Baa exists because of a problem anyone running agents locally recognises: most of the RAM was not going to the apps you had open, it was going to dev servers, builds and agent leftovers you forgot about. Baa is a menu bar app that warns you when your Mac is about to run out of RAM or disk, finds what is quietly eating it, and frees it in one click. It is small and single-purpose, and does exactly one thing well.
Bomfather is the security counterpart and operates at the opposite depth: an open source runtime security agent you deploy on Linux servers to protect the resources your workloads depend on, such as GPUs, databases and filesystems. It enforces policy with eBPF, so it runs in the kernel rather than in userspace where another process could tamper with it, and the project reports 1% to 3% overhead against much higher overhead for confidential computing. It is a background process needing no changes to your infrastructure or programs, and one small policy file can cover a whole fleet through inheritance.
Science on your own hardware
Liatir is a local-first desktop environment for bioinformatics, built with Rust, Tauri 2 and SvelteKit. It brings native tools, locally managed AI models, plugins, visual pipelines and scientific viewers into one application, with API connectors, external workflow integrations and a controlled local MCP interface for connecting other applications. Local analyses keep working offline once dependencies are installed, and workspaces retain inputs, execution logs, results and provenance. It publishes scientific showcases that document methods, measured results and limitations - including a comparison of pretrained single-cell foundation models against established baselines, where the pretrained models did not show a uniform advantage.
One spec, everything downstream
Octri takes an OpenAPI spec all the way to production: a docs site and API guides, SDKs in ten languages, an MCP server that exposes your API as tools for AI agents, and monitoring that captures errors from your SDKs and your API. The free tools are usable on their own - a spec audit that scores how well a spec will build an SDK with a fix for every finding, a parser that validates, bundles and dereferences, and a converter from Swagger, Postman and RAML - and there are public samples built from real OpenAI, Anthropic, GitHub and other specs.
An application that repairs itself
Recast is the most experimental launch in this group and the one worth watching. It is a Smalltalk application that uses an LLM to repair itself while it runs: a failing call supplies the stack, method source, receiver state and tests, the model proposes a repair, and a separate VM verifies it before the application replaces the method in its live image. Subsequent calls use the new code with no rebuild and no restart. The demo is a vending machine with a deliberately broken method - the first call fails, investigation and verification run asynchronously, and the next call succeeds. Smalltalk is a good substrate because classes, methods and execution contexts are objects the running system can inspect. The author is explicit that this is an experiment about how much of a maintenance loop can become part of the application itself.
Knowing the bill before it arrives
Manifestack is a free skill for your coding agent that answers a question most teams ask too late: what will this stack cost at 50k users? It prices every service at your usage on the vendor's own pages rather than a static table, and the same report flags the plan limits that will take a product down, the data rules it breaks, and services you are paying for twice - each with the fix that cuts it. It is aimed at developers, small teams and agencies picking their own services, explicitly not at enterprise FinOps. Prices are read from vendor pages at a point in time, so they drift.
Where this leads
The recurring tension in this batch is scope against trust. Moching takes over your keyboard and mouse, Bomfather sits in your kernel, Recast rewrites running code. Each one pairs that scope with a boundary - local storage, kernel-level enforcement, a separate verification VM - and the launches that get the boundary right are the ones you can actually leave running.