An agent that only works while your laptop is open and you are watching it is not really doing a job. This week's launches are about everywhere else: a server on a machine you already own, a Windows box where one agent drives the others, a desktop app that needs no connection at all, two different ways for agents to talk to each other, an API your agent can sign itself up for, and two tools about what the agent can see and reach once it arrives.
A server on a machine you already own
Mainplane is an open source agent ecosystem you host yourself. You run the install script on Linux or macOS, or a PowerShell one-liner on Windows, on any always-on machine, and the quickstart walks you through adding providers, workers and connectors. The vocabulary is the useful part: workers are the computers your agents control, and connectors are the ways you message them. The architecture puts secrets and controls, observability and the file server inside the server itself, with a network joining it to every worker. So the pitch is not a better chat window, it is that the old PC in the corner becomes the place your agents live, and you talk to them from wherever you happen to be. It needs a machine left running to be worth anything, and the written detail lives in separate docs behind the diagrams.
One agent driving the others, on Windows
Termexo is a Windows workbench built on the observation that most people now run more than one coding agent. It covers Claude Code, Codex, OpenCode, Antigravity and Grok Build, and the job it does is triage: you see which agent needs you, resume sessions, and reconnect from your phone. That last one is the detail that changes how it feels to use. You are not tied to the machine the agents are running on, so waiting for an agent does not mean staying at the desk. It is a Tauri 2 desktop app at version 0.11.0, open source, and also published to npm. It is Windows only, and it sits above your agents rather than replacing them, which makes it one more thing to keep updated.
The whole thing on your own computer
Hob takes the opposite route from both of those: no server, no orchestration, just a desktop app that runs open-weight models locally. It is a free, open source alternative to Claude and ChatGPT for Mac, Windows and Linux, and the thing it removes is the decision. Rather than asking you to choose weights, Hob picks the best frontier open-weight model that fits your computer and sets everything up. The README shows it reading a sales spreadsheet inside a project, charting revenue by region and summing up what stands out, as well as filling out PDFs. Because everything runs on your own machine it is private and works offline. The trade is the obvious one: you get open-weight models such as Gemma and gpt-oss, not frontier hosted ones, and quality is bounded by the hardware you own.
Agents in a room together
Tabichat is a shared chat layer that puts humans and several agent sessions into a single MQTT room. It was built because nothing on the market did the combination its author wanted: fully private, cheap and light to host as an MQTT broker using the IoT mosquitto protocol, able to talk to the model through the harness as well as in the room, and keeping all thinking visible in the harness whether the prompt arrived there or in the chat. A lead agent can be configured to coordinate the group and remotely renew other agents. Auto-renew fires at a configurable context percentage, but only when four conditions hold: the policy is armed, the seat has not already renewed this boot, its context reading is at or over the configured floor, and it has written a RESUME file to disk. It is mTLS capable. The constraint is real, though: pi is the only supported agent harness.
Kashkeshet is the minimal version of the same idea, and proud of it. It is an MCP server in Rust that lets your agents set up chat channels and talk to each other, and the premise is stated directly: modern agents do not need much hand-holding and are better served by a basic set of primitives they can use as they see fit. You install with a release script, run it as a background service that starts again at login, and connect your clients to localhost at 127.0.0.1:7420 using the shared access key it generates. A web page lets you watch the conversations. The README is unusually clear about the risk: security is minimal by design, it prevents accidental connections and nothing more, so this belongs on localhost or a network you control, such as a tailnet.
Models the agent provisions itself
Scaled Thought is an OpenAI-compatible API for open-weight models, and its onboarding is the interesting part. You paste one line into Claude Code, Codex, Cursor or any agent that can run commands, and the agent sets up models, keys and budgets itself. If you would rather do it by hand, you change the base URL and keep your existing OpenAI SDK code, calling models such as MiMo-V2.6-Flash and DeepSeek-V4.1-Flash. Spend is prepaid credits rather than a subscription, which makes the ceiling visible before you start. Two things to read before relying on it: there is no subscription option described at all, and the free usage offer is limited to new accounts in the US region and expires after 30 days.
What it sees when it gets there
XCrashlytics is a macOS CLI for Firebase Crashlytics and local Xcode Organizer reports, and it exists to close a specific gap. Your agent has the code, but a stack trace alone does not tell you how widespread a crash is or which release introduced it, and an agent cannot investigate evidence it has not seen. So it pulls crash messages, app stacks and impact reports out of Crashlytics and onto your terminal, where you and the agent can both read them, and on iOS it also reads reports Xcode has already downloaded. You start from a question, such as what is crashing checkout since the last release. It works with iOS, Android, macOS and other Crashlytics apps, is MIT licensed and needs macOS 15 or newer.
Heldkeys asks the question that should come before any of this: if a prompt injection hijacked the agent running here, which credentials could it use? It lists them for your terminal, the agent's sandbox or the CI job, without printing a single secret, and it also flags risky AI agent steps in GitHub workflows. The value is in how narrow it is. It does not try to be a scanner or a policy engine; it turns an abstract worry into a list you can read and shorten. Deciding what to remove is still yours, and it only sees the machine it runs on.
Put together, these eight describe a shift from an agent as a thing you watch to an agent as a thing you leave running, which is exactly why the credential question and the crash question show up in the same week.