Runs on CRUX or on your machine: where a CRUX agent works

CRUX team · · Updated · 5 min read
Two places an agent can run
Every agent in a CRUX workflow has a place where it runs, and the dashboard shows it next to the agent: Runs on CRUX or Runs on your machine. The two differ in where the agent works, whose model does the thinking, and what the agent can see.
| Runs on CRUX | Runs on your machine | |
|---|---|---|
| Where it works | On CRUX's servers, as a chat in the dashboard | In your own coding agent: Claude Code, Codex, Cursor |
| Whose model | The project's model key (Anthropic, OpenAI, Google or another provider) | Your coding agent's own model, on the plan you already pay for |
| What it sees | The tracker, and, when allowed, the code and wiki of the repository bound to the project, read-only | Your checkout: your branch, your files, your tests |
| What it changes | Only the task: comments, artifacts, the tracker actions its job needs | Code too: it edits files, runs tests, opens pull requests |
Runs on CRUX
An agent on CRUX works in the browser. Anyone on the team can use it, a PM, QA, a designer or a client, without installing anything or having a coding agent of their own.
It thinks with the model the project set up. The project adds its own model key (Anthropic, OpenAI, Google or another provider), and the session's sidebar shows the model it runs on. You pay the provider directly, at their price: CRUX doesn't resell tokens.
It reads the tracker and, when the agent is allowed to, the code and the wiki of the repository bound to the project. It can open files, search them and read their history, but it can't change them, and it has no shell. When the project has no repository yet, the agent still starts, and the chat says what it goes without.
What it writes goes to the task: a question, a plan, test cases, an estimate. It drafts in the chat and posts after your "ok". (More on these agents: AI agents for the rest of the team: PMs, QA and clients.)
Runs on your machine
An agent on your machine is a prompt for your own coding agent. CRUX builds the prompt from the workflow: the agent's instructions, the task, the artifacts of the steps before it, and the skills to load. You paste it into Claude Code or another coding agent connected to CRUX (how to connect one), and the work happens there.
So it uses your coding agent's own model, on your plan, and it sees what you see: the branch you are on, your local files and your tests. It can write code, run the test suite and open a pull request, which an agent on CRUX never does. It reads and writes the task through CRUX, as you.
How to pick
Whoever sets up the workflow picks where each agent runs, in the agent's settings. A few rules of thumb:
- A step that writes code runs on your machine. It needs a checkout, a build and tests. Such an agent runs only there, so its step offers one way to run.
- Text work runs on CRUX. Interviews, requirements, test cases, estimates, status reports. People who don't write code use these agents in the browser, and every session is paid from the same project key, whoever starts it.
- An agent on CRUX can also run on your machine. Its step offers both, so a developer who prefers their own coding agent, or wants the agent to read their local branch, can run the same step there.
Start session or Copy prompt & wait
When you open a pipeline step in the dashboard, it shows where it can run:
- Run on CRUX → Start session. The session starts on CRUX's servers with the model in the sidebar, and you talk to the agent right there.
- Run on my machine → Copy prompt & wait. The button copies the step's prompt and marks the step as waiting for your agent; it then reads Waiting for your agent…. Paste the prompt into Claude Code. When your agent picks the step up, the session becomes active in the dashboard, and the result lands on the same task.
The button's menu has two more ways to copy: the prompt without the desktop-only instructions, and a slash command. A chat agent on your machine has no step to wait on, so it offers plain Copy prompt. From Claude Code itself, the crux plugin starts a step with /crux:crux-run.
A person starts every run
Wherever an agent runs, it never starts by itself. A task entering a status starts nothing, and no agent in CRUX runs in the background. A chat begins when someone opens it, and a pipeline step begins when a person clicks Start session or pastes its prompt. When a step finishes, the next one waits for you. When an agent waits for your approval, it shows up in Needs you on your home page. (The tour of these words: How CRUX works.)
What it costs, and where you see it
- On CRUX, every turn is recorded with its tokens and its cost at the provider's price. Spend shows them on the project's Workspace, on the organization page and on your home page, by agent or by person. A member sees their own spend; the project's admin, or an org owner or admin, sees everyone's. Each chat shows its own cost too.
- On your machine, the run is paid through your coding agent's plan. CRUX doesn't see those tokens, so Spend doesn't show them.
In short
- Runs on CRUX: in the browser, on the project's model key, reading the bound repository without changing it. It suits text work and people who don't code.
- Runs on your machine: in your coding agent, on its own model, in your checkout. This is where code gets written.
- Start session runs a step on CRUX; Copy prompt & wait hands it to your agent.
- A person starts every run, and Spend shows what the runs on CRUX cost.
Try this on your own project. CRUX is free for one person, no card.
Start freeNext: Planned tasks, run in parallel: waves of agent sessions in git worktrees
All articles