Coding agents burn tokens fast
A coding agent makes many model calls — read the file, write a patch, run the test, fix the test. Without a budget, one tricky task can run up a bill. With one, it stops at a number you set.
Work from your terminal
$ remoteagent chat "why is this build failing?"
$ remoteagent exec
$ remoteagent auditKeys never reach the repo
Provider keys sit in the vault, so there is nothing to accidentally commit. The agent calls the model through a proxy and only ever holds a revocable device token.
The sandbox keeps the repo safe
Commands are argv-based and allowlisted, and deletes go to the trash. A web page cannot talk the agent into running rm -rf on your project.
A sensible setup for coding
- Give the agent a per-run token budget.
- Allowlist the commands your toolchain needs.
- Keep the workspace scoped to the repo.
- Rate answers so the good runs are exportable for fine-tuning.