← All posts

Integrations ·

Claude Code in Your Terminal Just Got More Than 400 New Tools

TRMNLY's MCP server hands every tool its own agent has to Claude Code, Codex and any other MCP client, behind the same approval gates. One command connects it.

A glass sphere joined by glowing cyan lines to a ring of small glass tool tiles

In The Pragmatic Engineer's 2026 survey, Claude Code was the most-loved AI tool at 46%, far ahead of Cursor at 19%.1 It runs in a terminal, and if that terminal is TRMNLY, Claude Code can use everything TRMNLY's own agent can.

One tool list, two agents

TRMNLY's in-app agent has more than 400 tools: the terminal, the browser sidebar, files, git, SSH, ports, the task board, automation.2 The MCP server gives that same set to outside clients over the Model Context Protocol,2 the open standard that lets AI apps connect to tools and data.3 Every tool the agent has is a tool Claude Code can call, and the calls run the same code the sidebar runs.2

So a Claude Code session can read another pane's output, drive the in-app browser, queue a command, or file a card on the project's board.

The same gates

An outside client doesn't get a back door. You pick an approval tier in Settings: no approval, approval for writes, approval for dangerous tools (the default), or approval for everything.2 Since 0.1.96, calls from Claude Code or any other MCP client also go through your command rules, the destructive-command prompt and the branch guard.4

Secret-looking values are scrubbed from results, the browser allowlist still applies, and every call is written to the audit log with the client's session id.2

It knows which project it's in

A Claude Code session started in a project folder works on that project, whatever window you're looking at. A worktree folder binds the worktree too.5

Reachable from every pane

A local socket comes up when TRMNLY launches, whether or not the external-clients setting is on.6 Every pane gets a per-launch $TRMNLY_SOCKET_TOKEN, so the trmnly command in a pane keeps working, while other programs on the Mac are refused the risky tools: the key vault, spawning a shell, typing into a pane.7 Running trmnly kanban add from a project folder puts the card on that project's board.2

How to connect

Turn on Settings → AI Agent → External MCP clients. The Apps list shows the clients on your Mac, with a Connect button for each.2 Or do it by hand:

claude mcp add --scope user trmnly -- "$HOME/Library/Application Support/TRMNLY/mcp-bridge"

The bridge runs TRMNLY's own binary, so you don't need Node.2,8 Codex's config is TOML, so for Codex you copy the setup from Settings and paste it in.2 Any other client can take the copied stdio config. Gemini CLI, for one, can launch a stdio MCP server from its settings.9

TRMNLY also installs five skill files under ~/.claude/skills/trmnly-* that tell Claude Code when to use these tools.2 If you'd rather not run another agent at all, you can pick Claude Code, Codex or Gemini CLI as the sidebar's own model.10

Sources

  1. 1.The Pragmatic Engineer, "AI Tooling for Software Engineers in 2026" (March 3, 2026). newsletter.pragmaticengineer.com
  2. 2.TRMNLY docs, "MCP server." trmnly.dev/docs/mcp
  3. 3.Model Context Protocol, "Specification." modelcontextprotocol.io
  4. 4.TRMNLY changelog, 0.1.96, "An approval means what the card said" (October 6, 2026).
  5. 5.TRMNLY changelog, 0.1.94, "An MCP client's folder decides which project it works on" (October 2, 2026).
  6. 6.TRMNLY changelog, 0.1.65, "The tools are always reachable from a pane" (August 23, 2026).
  7. 7.TRMNLY changelog, 0.1.66, "The always-on local MCP socket is gated by a per-launch token" (August 23, 2026).
  8. 8.TRMNLY changelog, 0.1.91, "Claude Code in the sidebar can ask before it runs a command" (September 27, 2026).
  9. 9.Gemini CLI docs, "MCP servers with Gemini CLI." github.com/google-gemini/gemini-cli
  10. 10.TRMNLY changelog, 0.1.43, "Claude Code, Codex, and Gemini CLI as agent providers" (August 4, 2026).