Choosing Your LLM & Coding Harness

Agentum makes two independent AI choices, and it’s worth keeping them straight:

  • The LLM provider is the model your agents talk with — it powers chat, summaries, digests, and the intelligence inside every agent.
  • The coding harness is the tool that writes agent code when you use Create, Improve, or Sidekick — Claude Code or Codex, running on your Mac.

They’re configured side by side in System → Settings → LLM Provider, and they don’t have to match: you can chat through Gemini and build with Codex, or run everything on the free default.

The LLM provider

Your options

Provider What you need Notes
Agentum AI (default) Nothing Free to start, no key, no signup. Runs on a text model Agentum funds and provides to trial and licensed users, with a $10 included usage budget.
Claude (Anthropic) Anthropic API key (sk-ant-…) Get one at console.anthropic.com.
Gemini (Google) Gemini API key (AIzaSy…) Get one at aistudio.google.com. Also enables voice transcription.
DeepSeek DeepSeek API key (sk-…) Get one at platform.deepseek.com.
MiniMax API key and GroupId From minimaxi.com.
None — Disables LLM features. Agents that don’t use AI keep working.

Agentum AI is the zero-setup path. There’s nothing to paste: your license itself is the credential, and Agentum’s servers meter and forward requests. If you already pay for an API key elsewhere — or want a specific model — bring your own key instead.

It’s free to start, not unlimited. Agentum pays the model bill, so each account gets a $10 included AI budget covering everything Agentum funds — chatting with Agentum, onboarding, building agents, the thinking inside the agents you run, and Agentum Code sessions. The app’s Account card always shows how much is left. If you use it up, connect your own key on this same screen and carry on — or contact support to have the budget raised; everything else in Agentum keeps working either way.

Picking a model

Each bring-your-own provider has a model field with suggested picks as one-click chips, or type any model name your provider supports. Leave it empty to use Agentum’s built-in default for that provider. The status line at the top of the card shows what’s live — e.g. Connected · Claude · claude-sonnet-4-6.

Switching providers

  1. Open System → Settings → LLM Provider.
  2. Pick the provider from the dropdown. Only the fields that provider needs are shown.
  3. Paste the key (there’s a reveal toggle if you want to double-check it), then Save.
  4. Click Restart agent when the banner asks — the running agents pick up the new provider on restart.

Switching is non-destructive: your other providers’ keys stay saved, so you can flip back and forth freely.

One caveat: the agent-building flows (Create Agent’s structured planning calls) need a provider that supports structured tool calls — that’s Agentum AI, Claude, Gemini, or DeepSeek. If you run MiniMax as your chat provider, Agentum tells you when a build needs one of the others.

The coding harness

When Agentum builds or edits agent code, it doesn’t do it with the chat provider — it hands the work to a real agentic coding CLI running on your Mac, with the agent’s source folder and a prepared task. You choose which CLI:

  Claude Code (default) Codex
Made by Anthropic OpenAI
CLI binary claude codex
Sign-in Claude subscription or Anthropic API key ChatGPT account
Install curl -fsSL https://claude.ai/install.sh \| bash curl -fsSL https://chatgpt.com/codex/install.sh \| sh
Docs docs.claude.com developers.openai.com/codex/cli
Desktop-app hand-off Yes (Claude desktop app) No

Select yours under Coding assistant at the bottom of the LLM Provider card. The setting applies to newly started sessions — a coding session that’s already open keeps whatever it launched with.

Detection and one-click install

You don’t have to pre-install anything. Agentum finds the CLI by probing your shell’s PATH plus the common install locations (~/.local/bin, Homebrew, npm, Volta, Bun…). If it’s missing when a session starts:

  • The inline session (embedded terminal) explains what’s missing, asks for consent, runs the official installer right there, and drops you into the CLI when it finishes.
  • The Terminal hand-off prints the exact install one-liner so you can run it yourself.

Where the harness shows up

Three surfaces use the harness — all covered in depth in Improving Agents:

  • Create Agent — the “draft it, then perfect it” hand-off after Agentum generates your first working version.
  • Improve — the per-agent tab that turns “add a snooze button” into a prepared coding session.
  • Sidekick — the companion rail, which can host a live inline coding session next to the agent.

Inline vs. Terminal vs. Desktop app

Every hand-off offers up to three places to run the session:

  • Inline — an embedded terminal inside Agentum itself. No window juggling; the session can move between the Improve tab, the Sidekick rail, and a full-screen view without restarting. This is the recommended default, and the only surface with the install-on-missing flow.
  • Terminal — opens macOS Terminal (or iTerm, auto-detected) with the CLI already launched in the agent’s folder and your task loaded. Good if you like your own terminal setup. Optional toggles: skip the CLI’s permission prompts, and auto-close the window when the session ends.
  • Desktop app — opens the task in the Claude desktop app via deep link. Claude Code only — the option is hidden when Codex is selected.

There’s also a fourth, tool-agnostic path: “Use another AI” copies a self-contained prompt to your clipboard for ChatGPT, Cursor, or anything else.

Frequently asked

Is Agentum AI really free? Yes — we pay the model provider, and you’ll never get a bill from us or from them for it. Every account includes a $10 usage budget, and the app’s Account card shows exactly how much is left. If you’re a heavy user and want headroom you control, connect your own key.

What if I use up the budget? Connect your own key in System → Settings → LLM Provider and Agentum picks up where it left off — your settings, agents, and everything else are unaffected. Or contact support to have your budget increased.

Does the harness cost extra? Agentum doesn’t charge for it, but the CLI uses your own Anthropic/OpenAI account — a Claude subscription, an API key, or a ChatGPT plan, billed by them.

Can I use different tools for different agents? The harness setting is global, but it only affects how a session launches — you can switch it between sessions any time.

Is my chat provider used for coding at all? For the big code-writing sessions, no — that’s the harness. Agentum’s own orchestration (planning your agent, reviewing what changed) uses the chat-side providers.

Next steps