TRMNLY for Claude Code

Give Claude Code a terminal, browser, SSH and Docker on your Mac

TRMNLY is a native Mac terminal, built on Ghostty, in which every one of its agent's 420 tools is also an MCP tool — so Claude Code, Cursor or any other MCP client connected to it can drive your terminal panes, a real browser, git, SSH hosts, ports and Docker containers, behind approval rules you set.

Set it up

  1. Install TRMNLY and sign in. It runs on macOS 15 or later. Accounts are invite-only during early access, so request access first.
  2. Turn on the server. Press ⌘,, open AI Agent → External MCP clients and turn on Expose TRMNLY's tools to other apps. It stays on across restarts.
  3. Pick an approval tier from the Approvals menu. Approval for dangerous is the default; the tiers are below.
  4. Connect Claude Code. The Apps list shows the MCP clients on your Mac; press Connect on the Claude Code row. Or by hand:
    claude mcp add --scope user trmnly -- "$HOME/Library/Application Support/TRMNLY/mcp-bridge"
    Keep --scope user: without it the entry only applies to the folder you ran it in. The launcher is TRMNLY's own binary, so it needs no Node.
  5. Check it. Start claude, run /mcp and look for trmnly. Then ask for something only TRMNLY can answer, like “list my open TRMNLY panes”.

Tools arrive in Claude Code as mcp__trmnly__read_terminal, mcp__trmnly__browser_navigate and so on. TRMNLY also ships five Claude Code skills (~/.claude/skills/trmnly-*) that tell Claude Code when to reach for them. A session started in a project's folder works against that project — its commands run there and its cards land on that project's board — whichever window you're looking at.

Cursor, Claude Desktop, VS Code and Codex CLI are in the same Apps list. TRMNLY edits their JSON configs in place; Codex CLI's config is TOML, so it is copy-and-paste. For anything else, Copy stdio config and Copy HTTP config give a block to paste.

No Mac? If your account has a cloud workspace, Claude Code can drive that Linux box instead. Sign-in happens in the browser the first time it connects:

claude mcp add --transport http trmnly-cloud https://trmnly.dev/mcp

What Claude Code gets

The same 420 tools TRMNLY's own agent uses — the server re-exports all of them, not a subset. Grouped, with a few of each:

Terminal62 tools
Read any pane, run a command where you can see it, queue one for when the current command finishes, split panes and watch output for a pattern.read_terminalrun_commandqueue_commandsplit_panewatch_terminal
Browser146 tools
Drive TRMNLY's built-in browser, on system WebKit or the Chrome on your Mac: navigate, click, type, take screenshots, read the console and network log, run JavaScript.browser_navigatebrowser_clickbrowser_screenshotbrowser_consolebrowser_networkbrowser_eval_js
Files, git and code42 tools
Read, search and edit files, check status and diffs, open pull requests, ask the language server for diagnostics, and run tests or background jobs.read_filegrepgit_statusgit_diffgh_pr_createlsp_diagnosticsrun_testsrun_background_job
SSH and remote hosts6 tools
Run commands on the SSH hosts saved in TRMNLY, move files over SFTP and apply port-forward presets.ssh_execsftp_readport_forward_apply
Docker and containers13 tools
Bring a Compose project up, read container logs and run a command inside a container.docker_compose_updocker_container_logscontainer_exec
Ports10 tools
See what is listening on which port, reserve ports per project, and stop the process holding one.ports_listports_kill_listener
Projects, worktrees and the task board22 tools
List projects, create a git worktree to work in, add cards to the project's board and run several workers at once.worktree_createkanban_addfleet_start
The Mac itself60 tools
Take screenshots, launch apps, run AppleScript, and schedule prompts and runbooks. These are the tools most worth keeping behind an approval.screenshotapp_launchosa_execschedule_create

Every tool, with what each does and whether it counts as a read or a write: the agent tool catalog.

What needs your approval

Calls from Claude Code go through the same approval and audit system as TRMNLY's own agent. You pick how much an outside client may do without asking:

No approval
Every tool runs without asking.
Approval for writes
Read-only tools run; anything that changes something asks first.
Approval for dangerous
The default. Reads and writes run; dangerous tools — running JavaScript in the browser, executing commands, reading credentials — ask first.
Approval for everything
Every call asks first.

A call the tier gates waits for you to approve it — in TRMNLY, or on a paired iPhone — and where nobody can approve it, it is refused with an error that says why. Approving can grant the client that tool for the session; grants expire when idle and can be revoked under Client access, where a Trusted tick stops the asking for one client altogether.

Whatever tier you pick

  • Credentials passed as tool arguments are refused; use {{key:…}} placeholders from the Keys vault instead.
  • Secret-looking values are scrubbed from results unless you grant that client the read_secrets capability.
  • The destructive-command gate still applies: rm -rf, git push --force and the like are refused or need an explicit confirmation.
  • If the browser is in allowlist mode, outside clients are held to the same allowlist.
  • Every call is written to the audit log with the client's session id.
  • Each client can carry its own allow and deny patterns (browser_*, *_navigate), a rate limit and a cost budget. Deny wins over allow.

More in the MCP server docs and Agent safety, or the 12-minute lesson, The MCP server.

Common questions

How do I give Claude Code access to my Mac's terminal, browser, SSH and Docker?
TRMNLY is a native Mac terminal, built on Ghostty, in which every one of its agent's 420 tools is also an MCP tool — so Claude Code, Cursor or any other MCP client connected to it can drive your terminal panes, a real browser, git, SSH hosts, ports and Docker containers, behind approval rules you set.
What is the command to add TRMNLY to Claude Code?
Turn on Settings → AI Agent → External MCP clients → "Expose TRMNLY's tools to other apps", then press Connect on the Claude Code row, or run: claude mcp add --scope user trmnly -- "$HOME/Library/Application Support/TRMNLY/mcp-bridge". Without --scope user the entry only applies to the folder you ran it in.
Can Claude Code drive a real browser through TRMNLY?
Yes. The browser tools drive TRMNLY's built-in browser, running on system WebKit or the Chrome installed on your Mac: navigating, clicking, typing, screenshots, the console and network log, and running JavaScript in the page. Running JavaScript asks for approval under the default tier.
Is it safe to let Claude Code control my Mac this way?
You choose how much it can do. The default tier, Approval for dangerous, lets reads and ordinary writes run and asks before dangerous tools such as executing commands, running JavaScript in the browser or reading credentials. Credentials passed as arguments are refused, secret-looking values are scrubbed from results, destructive commands are gated, and every call is in the audit log.
Does this work with Cursor, Codex or Claude Desktop?
Yes. The External MCP clients settings list Claude Code, Claude Desktop, Cursor, VS Code and Codex CLI. TRMNLY writes the JSON configs in place; Codex CLI's config is TOML, so it is copy-and-paste. Any other MCP client can use the Copy stdio config or Copy HTTP config blocks.
Does Claude Code have to run inside TRMNLY?
No. The bridge Claude Code starts reports the folder it was started in, and TRMNLY runs tool calls against the project that folder belongs to. TRMNLY can also run Claude Code, Codex and Gemini CLI as the agent in its own sidebar.
What does it cost?
The MCP server is part of the Mac app, with no separate charge. A desktop seat is $99 with 12 months of updates. TRMNLY needs an account, and during early access accounts are approved by hand.
Is TRMNLY the same as TRMNL?
No. TRMNLY (trmnly.dev) is a Mac terminal and agentic development environment made by AW3 Technology. TRMNL is an unrelated e-ink display.

Try it with your own Claude Code

TRMNLY is in early access. Request access, and once your account is approved, download the Mac app and connect Claude Code in a minute. See also what an agentic development environment is and how TRMNLY compares.