Operator sells the supervision layer, never the intelligence. Every model call is made with your key or on your silicon, and the bill goes to the provider at their price with nothing added.

There is no Operator proxy, no Operator gateway, and no Operator API key anywhere in the product.

Adding a key

On the Mac, open Settings and paste a key for any provider you use. Anthropic, OpenAI, Google, xAI and several others are supported.

Where keys live:

• Stored in your Mac’s Keychain.
• Never uploaded, never synced to your phone, never sent to a relay.
• Handed to a job only at the moment it starts, as an environment variable inside that one sandbox.
• A job already running keeps the copy it was given. Deleting a key stops the next job, not the current one.

If you would rather enter a key from your phone, you can. It gets sealed to your Mac on the way and is not readable in between.

Choosing an agent

An agent is the program that does the work. Operator drives several, and they are genuinely different tools rather than skins on one thing.

Agent	Good at
Claude Code	The reference runtime. Best all round, and the one Tag adopts from your terminal.
Codex CLI	OpenAI’s, strong on code, blocks on its own approval calls.
Gemini CLI	Long context, Google models.
Grok	xAI’s agent.
Pi	Small, multi provider, runs anywhere.
OpenCode	Open source, pairs well with open weights models.

Pick per job on the launch sheet, or set a default per bot. If you have a key for it, you can use it.

Choosing a model

Within an agent you can pin a specific model. The picker is curated per agent, so you see the models that agent can actually use rather than a list of every model that exists.

There is also an Auto mode that picks for you based on the job.

Running a local model

If you install Ollama, Operator can run jobs entirely on your Mac with no key and no bill. It suggests models that your specific machine can actually run well, rather than letting you pick one that will swap itself to death.

Local models are slower and less capable than the frontier ones. For summarising, drafting, and classification they are genuinely fine, and nothing leaves the machine.

What this means for cost

A job’s cost is the tokens the agent used, at your provider’s rate. Operator adds nothing to it.

Two habits that matter more than model choice: give a job a clear instruction so it does not explore, and use a smaller model for the jobs that do not need a bigger one.