Execution modes

Where agents run decides where your code is processed and which model they use. You can start on one mode and move to another.

Trailboss-hosted

Agents run on Trailboss runners, in isolated environments created for each job and torn down after it. They use your own model provider key.

  • Your code is checked out into the job's environment for the length of the job, then the environment is destroyed.
  • You pay for the hours agents work on our runners. Tokens are billed by your model provider; Trailboss never marks them up.
  • Your plan sets how many agents work at once and how long a job may run.

In your AWS account

A Trailboss worker runs in your AWS account, and agents run there, under roles you control. Use Claude on Amazon Bedrock in that account, or your own model key.

  • Code is checked out and run in your account.
  • With Bedrock, your data stays in your account: code, prompts and responses don't leave it.
  • With your own model key instead, prompts go to that provider.
  • The worker only accepts jobs signed by Trailboss, for your workspace, and reports back status and summaries, with secrets removed before anything leaves.

On your machine

The CLI can run jobs on your own computer against a local model (any OpenAI-compatible endpoint, such as Ollama). It is a good fit for trying things out and for work that should never leave your laptop. Jobs for your machine are only ever claimed by your machine.

The same rules everywhere

In every mode, a job is checked against your workspace's plan, budget and policies before it is queued, and every action is recorded in the audit log. Changes reach production only through releases, which pass the quality gates and the approvals your policies require for their risk.