All pages
Lesson 7.1 · Beyond the Mac · 12 min
The MCP server
Give Claude Code, Codex or any MCP client TRMNLY's tools.
After this lesson, Claude Code, Codex, Cursor or any other MCP client can call TRMNLY's tools. They can read your panes, run commands, drive the built-in browser and work the project board, under approval rules you set.
Time: 12 minutes · Works on: macOS app, cloud workspace
What the MCP server is
MCP is the Model Context Protocol, the standard way coding agents load outside tools. TRMNLY's own agent has several hundred tools. The MCP server hands that same set to agents running outside TRMNLY. A Claude Code session in one of your panes can then read another pane, queue a command, or add a card to the board, just as the sidebar agent can.
The calls go through the same approval and audit system as the sidebar agent. You decide how much an outside client may do without asking.
Turn it on
- Press
⌘,to open Settings, then choose the AI Agent tab. - Find the External MCP clients section. You can also type "MCP" into the search bar at the top of Settings.
- Turn on Expose TRMNLY's tools to other apps. The setting stays on across restarts.
- Pick a level from the Approvals menu:
- No approval runs every tool without asking.
- Approval for writes runs read-only tools and asks before anything that changes something.
- Approval for dangerous is the default. It runs reads and writes, and asks before dangerous tools such as running JavaScript in the browser, executing commands, or reading credentials.
- Approval for everything asks before every call.
Tip: Start on Approval for dangerous. Move to a stricter level if you are trying a client you don't know yet.
The Plugins tab in the right sidebar has the same switch and a Manage clients… button that opens this section.
Connect Claude Code
The Apps list in the same section shows the MCP clients TRMNLY found on your Mac: Claude Code, Claude Desktop, Cursor, VS Code and Codex CLI. Each row reads that client's own config file, so the status shows what the client will actually load.
- Find the Claude Code row and press Connect. TRMNLY registers itself for every project on the Mac.
- Check the row now reads Connected · all projects.
- Open a new pane and start
claude. Run/mcpinside Claude Code and look fortrmnlyin the server list. - Ask Claude Code something only TRMNLY can answer, such as "list my open TRMNLY panes".
Some rows may show a different status:
- Connected · … only means an older entry that works in one folder. Press Fix to register it for every project.
- Connected · needs node means an entry from before TRMNLY had its own launcher. Press Update.
- Add by hand means TRMNLY can't write this client's config. Codex CLI is like this because its config is TOML.
For any client not in the list, use Copy stdio config or Copy HTTP config at the bottom of the section and paste the block into that client's settings. The HTTP option needs Also listen on localhost HTTP / SSE turned on.
Which project a client works in
A tool call that doesn't name a project runs against the project the client was started in. TRMNLY matches the client's working folder against your project and worktree folders. A Claude Code session you start in a project's folder files cards on that project's board and runs commands there, whatever window you're looking at.
To pin a client to one project regardless of where it starts, pick the project from the menu on its row under Client access. Leave that menu on Follow client folder unless you need a pin.
Trust and limits per client
Every client that has connected gets a row under Client access.
- Tick Trusted to stop being asked about that client at all. Untick it to go back to the approval level above.
- Open the ⋯ menu on the row and choose Edit tool access… to set allow and deny patterns such as
browser_*. You can also set a rate limit here. Deny patterns win over allow patterns. - Approvals you grant "for this session" are listed under Grants, where you can revoke them.
Some protections apply whatever level you pick. Credentials passed as arguments are refused, so use {{key:…}} placeholders from the Keys vault. Secret-looking values are scrubbed from results. The destructive-command gate still applies. Every call is recorded in the audit log with the client's name.
The cloud workspace over MCP
If your account has cloud workspace access, an MCP client can drive the Linux box at trmnly.dev with no Mac involved:
claude mcp add --transport http trmnly-cloud https://trmnly.dev/mcp
codex mcp add trmnly-cloud --url https://trmnly.dev/mcp
The name is trmnly-cloud so it doesn't hide the Mac's trmnly server. The first time the client connects, your browser opens a trmnly.dev page asking whether to connect it. Press Allow. Connecting costs nothing. The first tool call that needs the box starts it, and the minutes bill against your credits. The cloud workspace covers the box itself.
Check yourself
- The Claude Code row in External MCP clients reads Connected · all projects.
- From Claude Code in a project folder, ask it to add a card to the TRMNLY board. The card appears in that project's Tasks tab.
- Set the level to Approval for everything, ask Claude Code to read a pane, and see the approval prompt appear in TRMNLY.
Go further
- MCP server: every setting, the binding rules, and the full cloud endpoint reference.
- Agent safety: the audit log and the destructive-command gate that also apply to outside clients.
- Claude Code, Codex and Gemini: running those CLIs inside TRMNLY's sidebar instead.