Perspective ·
TRMNLY Is an Agentic Development Environment. Here's What That Means in Practice.
An agentic development environment needs a gate on what agents run, a way to run several at once, tools that reach the whole machine, and a way to follow the work when you walk away. This is how TRMNLY does each.

The last post listed what changes when agents write the code: the unit of work becomes a task, approval replaces undo, one cursor becomes many agents, and the agent needs the whole environment, not just the open file. TRMNLY was built around those needs. Here is how each one shows up in the app.
It starts from the terminal
TRMNLY is built on Ghostty. The middle of the window is a terminal, and the agent sits in a sidebar beside it.1 That's a deliberate starting point. Coding agents already do their work in a shell, and so do the CLIs many developers already pay for. TRMNLY doesn't move you into a new editor to get agents. It brings them to the place they already run.
The agent works on tasks, and shows every step
You give the agent a task in plain language, such as "run the tests and tell me which ones fail." It works in steps. Each step is a tool call, shown as a card: reading a file, running a command, reading the terminal.2 When one turn wants to change several files, a banner shows how many changes are pending, and you can uncheck any file before approving the rest.2
You don't have to use TRMNLY's own agent loop, either. If you already pay for Claude Code, Codex or Gemini CLI, you can pick it as the agent's model. The turn bills to the plan you already have, and TRMNLY never holds a key for it.3
Approval is the main control
Every tool is classed as READ or WRITE. The posture chip in the agent's header sets what runs without asking:4
- Approve, the default: reads run on their own, writes wait for you.
- Auto-accept: everything runs.
- Observe: read-only. Writes and commands are blocked.
Approving the same safe command fifty times is tedious, so any approval can become a standing rule from the card's Always… menu: for a tool, a path, a folder or a web host.4 Some guards ask even on Auto-accept. The first irreversible command in a session, such as rm -rf or git push --force, always prompts. On protected branches like main, force pushes and hard resets need two confirming clicks.4
Every call is logged in the Safety panel's audit log. ⌘. stops the current turn.2,4
Many agents, without collisions
Each project can have git worktrees, separate checkouts that share one repository. New Worktree… (⌃⌘N) creates one, and the agent works in whichever is active.5 So two agents can work on two branches without touching each other's files.
Fleet goes a step further. It runs up to 8 workers that pick ready cards from the project's task board, each in its own worktree.6,7 When a worker claims a card, other agents keep off it. The default landing policy is Open a PR: the branch is pushed, a pull request opened, and you merge.7 Fleet is experimental, and the lesson says to start it with one worker and a small budget.7
Tools for the whole machine, for any agent
The agent's tools reach well past the editor: your terminal panes, files, a built-in browser, SSH hosts, ports and containers.1,8 TRMNLY also runs an MCP server that hands the same tools to agents outside it. Claude Code, Codex, Cursor or any other MCP client can read your panes, run commands, drive the browser and work the task board, under the approval level you choose.8 An agent you start in another app gets the same environment as the one in the sidebar.
The work keeps going when you leave
Two pieces cover the time you aren't at your desk.
- The cloud workspace is a Linux box at trmnly.dev. It's billed by the minute against your credits while a terminal is attached, at $0.50 an hour on the base pack, with models on your own keys.9
- The iPhone app pairs to your account. You can read and type into your Mac's panes, and answer the agent's approval requests from a notification on the lock screen.10
What it isn't
TRMNLY isn't trying to replace your editor. It has a file tree with preview and inline editing, but its center is the terminal and the agents working in it.1 Keep the editor you like for the careful reading and the quick fix. Use TRMNLY for the part of the work you now hand to agents.
If you want to try it, the Learn TRMNLY course goes from installing the app to running agents across your projects, other machines and your phone, one short lesson at a time.
Sources
- 1.Learn TRMNLY, "A tour of the window." trmnly.dev/learn/tour-of-the-window
- 2.Learn TRMNLY, "Your first task." trmnly.dev/learn/your-first-task
- 3.Learn TRMNLY, "Claude Code, Codex and Gemini." trmnly.dev/learn/cli-agents
- 4.Learn TRMNLY, "Safety and approvals." trmnly.dev/learn/safety-and-approvals
- 5.Learn TRMNLY, "Worktrees." trmnly.dev/learn/worktrees
- 6.TRMNLY docs, "Fleet." trmnly.dev/docs/fleet
- 7.Learn TRMNLY, "Fleet." trmnly.dev/learn/fleet
- 8.Learn TRMNLY, "The MCP server." trmnly.dev/learn/mcp-server
- 9.Learn TRMNLY, "The cloud workspace." trmnly.dev/learn/cloud-workspace
- 10.Learn TRMNLY, "The iPhone remote." trmnly.dev/learn/iphone-remote