For most of the last two years the answer to "where does a coding agent run" was simply: a terminal window. You opened one, watched it work, and opened another when you wanted a second attempt. That answer is breaking down. Agents now run for long stretches, fan out into subagents, edit files you are also editing, and cost real money while they do it. The tools arriving on Show HN this month are less about the model and more about the room you put the model in.
Below are eight recent entries that each make a different bet on that room. Some isolate the agent from your machine. Some let you run four at once and keep track of which one needs you. One is a markdown editor that assumes an agent will change the file underneath you.
Isolation first: runtimes that assume the agent will act
The first design decision is trust. If an agent is going to run shell commands, open ports and read files, do you let it loose on your host, or do you build it a room?
OpenShell takes the second position and makes it the product. Built in Rust and released under Apache 2.0, it is described by its authors as a safe, private runtime for fleets of autonomous AI agents. The word fleets matters: this is not a single-session wrapper but an execution environment meant to be handed to many agents at once. Version 0.1.x adds new isolation primitives and an expanded extension surface, which is the clearest signal yet that agent runtimes are converging on the same shape as container runtimes did a decade ago. It installs from PyPI, so the barrier to trying it is low, though the early version number means the interfaces are still settling.
MIKE reaches a similar place from a different direction. It is an LLM harness that puts a Kanban board in front of the work, and pairs it with gVisor sandbox isolation and extensions that run in sandboxed iframes or WebAssembly through wazero. Notably it also ships a Progressive Web App optimized for spotty networks, which tells you the intended user is not always sitting on a fast office connection. It talks to llama.cpp locally or to anything speaking the OpenAI Chat Completions API, so the sandbox and the model are independent choices.
The tradeoff is real. Both projects ask for Docker or Podman and a generated encryption key before you see anything useful. That is the price of treating the agent as untrusted.
Multiplexing: many agents, one screen
Once isolation is handled, the next complaint is attention. You can start four agents, but now you have four things to watch.
Corgi frames itself as a terminal multiplexer for humans and agents. It keeps watch over every agent you have running, marking each as working, needing you, finished or idle, and offers a single key to catch up on whatever changed. Beyond status it tracks where your tokens went with per-agent activity tracking, replays every file an agent touched in one scroll, and organizes workspaces into worktrees. Like tmux, your agents and dev servers keep running in the background when you leave. It installs with a curl one-liner and supports macOS, Linux and WSL, and it reuses the Claude Code, Codex, Pi and OpenCode accounts you already have rather than adding a vendor in the middle.
Craig CLI is more opinionated: it is a terminal workspace that takes a task all the way to a pull request. Claude Code, Codex and Cursor run in parallel, each task in its own Git worktree, and tasks, changes, pull requests and checks stay together in one keyboard-driven view. That worktree-per-task rule is the useful part, because it is what stops two agents from writing into the same tree. It installs from npm, needs Node 22 or newer, is MIT licensed and requires no Craig account at all.
Native workspaces on the desktop
Not everyone wants to live in a terminal. Two recent projects put the same coordination into a desktop application.
Twine is a native macOS app, written in Rust under the MIT license, that its authors call an Agentic Development Environment. It runs on macOS 26 or later and exists to coordinate coding agents inside one native window instead of a scattering of terminal tabs. Being native matters here: window management, notifications and file watching all behave differently than they do in a browser tab.
Halv is a desktop workspace with a sharper economic thesis. It runs Claude Code, Codex, Kimi and GLM side by side on the subscriptions you already pay for, and layers a local savings engine on top: command output filtering and context compression both happen on your own machine, and a Crux code index answers caller, reference and blast-radius questions from an index rather than by reading files. A live savings meter reports what each conversation saved. The paid Halv Cloud tier adds a router that sends each task to the model it needs plus automatic account rotation when one plan hits its ceiling. The desktop app and every local savings layer are Apache 2.0 and free forever.
Editing while an agent edits
The least discussed problem is collision. You are writing a document; an agent is adding links to the same document; most editors resolve this by clobbering one side or throwing up a file-changed-on-disk dialog.
EditMD is a small native macOS markdown editor built around exactly that case. You edit the rendered document in place and it saves plain markdown back, and it uses Y.js to merge your edits with the agent's in real time, the way a shared document does. The author's motivation was practical: with Claude writing much of the code, more time goes into editing docs, and the editor needed to survive a file being changed underneath it.
On the agent side of that same boundary, Keen Code is a context-aware terminal coding agent written in Go. It supports multiple providers, MCP servers, subagents and agent skills, and adds the knobs that matter once sessions get long: controllable tool output retention, hashline edits, compaction, a permission system and session management. It also routes tasks through a decision-model classifier with documented fail-open behaviour, which is a cheaper alternative to sending every request to a large model.
How to choose
Pick by the failure you actually have. If you do not trust the agent with your machine, start with OpenShell or MIKE. If you trust it but cannot keep track of it, Corgi or Craig CLI will help immediately. If your pain is spend, Halv's local filtering and index are the most direct lever of the eight. If your pain is a file being rewritten while you type in it, EditMD solves that one problem and nothing else.
The pattern across all eight is that the model is no longer the interesting part. The interesting part is the room, the meter and the lock on the door.
