All pages

Agentic development foundations · Lesson 1.3 · 10 min

Pick the right agent

The built-in agent, Claude Code, Codex, Gemini or a local model, and when each fits.

After this lesson you can choose between TRMNLY's built-in agent on a hosted model, the same agent on a local model, a coding CLI such as Claude Code, Codex or Gemini, and the headless trmnly-code CLI, and you know what each choice changes about cost, privacy and control.

Time: 10 minutes · Works on: macOS app.

The choice is about the run, not loyalty

There's no single best agent. The right one depends on the task in front of you: how much it needs to know about your machine, how much context it needs, whether the code may leave your Mac, and who pays. It's normal to use two or three in one day, and TRMNLY is built so you can.

Ask four questions before you start a run.

  1. Does the task need the machine, or only the repository? Reading your terminal panes, driving the built-in browser, reaching saved SSH hosts and ports: that's TRMNLY's tool surface. Pure code changes inside a repo can be done by any capable coding agent.
  2. How much context does it need? A bug that spans many files needs a model with a large context window. A focused fix in one file doesn't.
  3. May the code leave this Mac? If not, the answer is a local model.
  4. Who pays? Your own API key, a subscription you already have, or your own hardware.

The options

Agent Where you talk to it Best for Who pays
Built-in agent on a hosted model The Agent tab Tasks that need the terminal, browser, board or other TRMNLY surfaces Your API key, sent straight to the provider
Built-in agent on a local model The Agent tab Private or offline work, small focused tasks Your hardware
Claude Code, Codex or Gemini CLI as the agent's model The Agent tab Using a subscription you already pay for, inside the sidebar Your CLI subscription
Claude Code, Codex or Gemini CLI in a pane A terminal pane Long coding sessions in the CLI you already know Your CLI subscription
trmnly-code Any shell Scripted or batch runs, PR reviews from the command line Local by default

The built-in agent on a hosted model

This is the agent with the most reach. It has TRMNLY's whole tool catalog: terminal, files, git, browser, Kanban, worktrees, connections, ports and more. Every tool is classed as read or write, write tools wait for approval under the default posture, and every call is recorded in the audit log. Choose it when the task touches more than code. "Run the dev server, open the page and tell me why the form doesn't submit" is a built-in agent task.

It runs on Anthropic, OpenAI, Google Gemini or OpenRouter, with your own key. On the Mac those requests go straight from your Mac to the provider and don't spend TRMNLY credits.

The built-in agent on a local model

A model running through Ollama, or another OpenAI-compatible server such as LM Studio, keeps everything on your Mac and costs nothing per turn. The trade-off is room. A local model gets a core set of 15 tools instead of the full catalog, because every tool description rides in every request and the full set is larger than most local context windows. Choose it for private code, offline work, and well-scoped tasks: "add a docstring to every exported function in this file", not "refactor the auth layer".

A coding CLI as the agent's model

Pick Claude Code (CLI), Codex (CLI) or Gemini CLI in the agent's model menu and TRMNLY runs that CLI behind the sidebar. The turn bills to the plan you already pay for, and TRMNLY never holds a key for it.

Know the trade-off: the CLI runs its own tools. Its edits and shell commands don't pass through TRMNLY's approval cards. You set its limits separately, with the gear menu's CLI agent permission level.

A coding CLI in a pane

Run claude, codex or gemini in a tab like any other program. This is the CLI exactly as you know it, with TRMNLY around it: the pane can come back in the same Claude Code conversation after a restart, ⌃⇧⌘S types a screenshot's path straight into it, and once you connect it over MCP it can call TRMNLY's tools to read other panes, drive the browser or work the Kanban board.

trmnly-code

trmnly-code is TRMNLY's agent as a command-line program, with the same tools, approval rules and audit log. With no flags it runs on a local Qwen coder model through Ollama. Choose it when you want an agent in a script: one prompt per line from a file with trmnly-code batch prompts.txt, or a pull request review with trmnly-code review 128.

A worked example

Say you have three jobs this afternoon.

  • A login page renders blank in the browser. That needs the dev server's output, the page and its console, so it goes to the built-in agent on a hosted model.
  • A config parser in a client's private repository needs a new field. The code must stay on your machine, and the change is small, so it goes to the built-in agent on a local model.
  • A three-hour refactor across a service. You live in Claude Code for long sessions, so you run claude in a pane, connected over MCP so it can still read your test pane and file cards on the board.

Do it in TRMNLY

Set up two agents and give each the task that suits it.

  1. Connect a hosted model. Open Settings with ⌘,, choose API Keys, paste a key into a row under Agent Providers and press Save.
  2. Connect a local model. In the right sidebar choose More → Local LLM. Press Start Ollama if it isn't running, pull a model such as qwen3-coder:30b under Pull a model, then use the row's menu to Set as agent default. Test it in Try a model at the bottom of the tab.
  3. Switch per conversation. In the Agent tab, open the model menu in the header. Your CLI subscription (no API credits) is listed first, then a section per provider. Picking a model changes only this conversation, so two tabs can run two different models side by side.
  4. Try a CLI as the model. With the CLI installed and signed in, pick Claude Code (CLI), Codex (CLI) or Gemini CLI in the same menu. Then open the gear menu's CLI agent section and confirm the permission level. Accept edits (default) lets it edit files; Read-only refuses every write.
  5. Try a CLI in a pane. Open Settings → AI Agent, find External MCP clients, turn on Expose TRMNLY's tools to other apps, and press Connect beside Claude Code. For Codex CLI, TRMNLY doesn't write its config file; it copies the snippet to your clipboard for you to paste. Start a new claude session in a pane and ask it to read your other terminal pane.
  6. Try trmnly-code. In Settings → AI Agent, find trmnly-code CLI and press Install. In a shell, run trmnly-code "list the TODO comments in this repo".
  7. Give each agent one task from your own list that suits it, using the four questions above.

Check yourself

  • You can say which agent you'd pick for a task that needs the browser, and why.
  • You know a local model gets 15 core tools, and what kind of task fits that.
  • You know that a CLI's own edits don't pass through TRMNLY's approval cards, and where its permission level is set.
  • Two tabs in the same project ran on two different models.

Go further