TRMNLY

Open a folder and say what you want. TRMNLY's agent reads the pane that failed, searches the code, proposes the fix and runs the command in that pane, once you approve it.

Most coding agents live one copy and paste away from the work. The error is in the terminal, the agent is in another window, and the developer is the clipboard between them. The agent can't see the dev server, the SSH session, the container or the port that's already taken, so every fix starts with explaining the machine to it.

TRMNLY puts the agent where the work happens. It's an agentic development environment: a native terminal with an agent built in that reaches the whole machine, and nothing runs without your say.

The Problem: Agents Can't See the Terminal

An IDE was built for a person typing. Working with agents means directing them, and today that looks like this.

• Copying an error out of the terminal into a chat, then pasting the answer back and hoping the command is right
• A separate agent per repo, per tool and per machine
• Trusting a whole session, or babysitting every keystroke
• No undo when an edit goes wrong
• Work that stops the moment you step away from the laptop

The Solution: An Agentic Development Environment

TRMNLY is built around agents rather than bolted onto them. The agents are first-class, the whole machine is their workspace, and approval is the interface between what an agent proposes and what runs.

The terminal underneath is Ghostty, unchanged in behaviour, so the shell, tmux, themes and muscle memory work exactly as before. Everything TRMNLY adds lives in sidebars you can hide.

Key Features

An agent that can see your panes

It reads what your terminals printed, runs commands in the pane you point at, edits files, searches code and drives a browser. Bring Claude, GPT, Gemini, anything on OpenRouter or a local model through Ollama, or run Claude Code, Codex and Gemini CLI inside the same sidebar.

The whole machine is the workspace

More than 300 tools reach every pane, SSH host, container, port and browser tab, plus a Kanban board the agent can pick cards from.

Approval is the interface

Commands can ask before they run, with rules per prefix, host and path. Deny rules see through sudo, env, subshells and quoting, every write is snapshotted, and a dry run stages a whole change so you can read the diff first.

Projects that remember

A project is a folder with its own tabs, environment, git worktrees, port aliases, board, notes and agent context. Parallel agents each get their own worktree.

An MCP server built in

TRMNLY re-exports its whole tool surface over MCP, so Claude Code, Codex or any MCP client can drive it with the same tools and the same approvals.

Your Mac, from anywhere

The iPhone app shows your Mac's terminals live, types into them and answers approvals from a notification or an Apple Watch. A cloud workspace keeps the work going after you close the lid.

Your Keys, Your Provider

On the Mac, model requests go straight from the machine to the provider on your own keys. TRMNLY never sits in that path or marks it up. The desktop app is $99 once per seat; cloud workspaces are $50 a month for 125 hours, or $0.50 an hour on credits.

Strategic Vision

TRMNLY is the developer layer of the AW3 ecosystem. It signs users in with 1Profile, AW3's identity product, alongside GitHub and email.

Its foundation enables:

• One sidebar that reaches every project, host and container
• Agents trusted to work alone, inside rules they can't talk around
• Work that carries on from a phone or a cloud workspace
• Any MCP-capable agent using the same tools behind the same approvals

Status and Stack

Status: available for macOS at trmnly.dev, with an iPhone app in public TestFlight beta, cloud workspaces in the browser, and Linux and Windows in progress
Platform: macOS 15 or later, iPhone, web and cloud
Stack: the Ghostty terminal core in Zig, a native Swift macOS app and iPhone app, an MCP server, and trmnly.dev on Next.js 16 with Neon Postgres and Drizzle