# Roscoe

**Predictable workflows from unpredictable AI.**

On its own, AI does the same task a little differently every time. That's fine
when you're chatting, and a problem when you need the same job done the same way
over and over. Roscoe fixes that. You wrap your AI steps in a workflow, and the
workflow runs the task the same way every time, on your own machine.

A workflow is a set of steps with clear rules for what happens next. Say you
handle customer refunds: read the request, check it against your policy, and for
the close calls, ask a few AIs to vote before you approve. In Roscoe that's a
workflow you can run again and again, and it behaves the same way each time even
though the AI in the middle doesn't. The AI still does the thinking; the workflow
keeps it on rails.

You don't have to build it by hand. Describe what you want to Claude and it
writes the workflow for you (see [Claude integration](/docs/running/mcp)), or
draw it on a graph in the [visual editor](/docs/running/web-ui). Either way it's
saved as a plain text file you can read, share, and roll back.

## What's in the box

Roscoe is built to make AI dependable enough to trust with real decisions, so it
gives you more than a way to chain steps together.

**It helps AI decide well.** An AI judge weighs a decision and sends the run the
right way. Consensus asks several AIs the same question and goes with the answer
they agree on, for the calls that really matter. Round robin pits options against
each other head-to-head and keeps the strongest, with a scorecard showing how it
won.

**It catches problems early.** Roscoe checks a workflow before it runs and flags
dead ends, missing branches, and loops with no exit. A dry run lets you give each
step a stand-in output, so you can test the whole path before running it for
real.

**It keeps you in control.** Every loop has a hard cap and stops itself if it
stalls. A human checkpoint pauses the run for a person to approve or choose.
Every run is recorded step by step, ready to replay. And a spending cap in
dollars or tokens stops a run once it goes over.

**It scales up.** Drop a whole tested workflow in as a single step, and build big
systems from small, proven parts. Give each step a different model: a fast one
where you can, a strong one where it counts.

Already have an AI skill, a prompt, or a checklist? Point Roscoe at it and it
turns that into a workflow for you.

## Runs on the Claude plan you already have

Connect Roscoe to Claude and your AI steps run on your existing Claude
subscription: no separate API key to set up, and no extra bill to reconcile.
Without a subscription, Roscoe runs on its own with your own API key. See
[Claude integration](/docs/running/mcp).

## Three ways to use it

- **Desktop app**, for everyone: a Mac and Windows app with the dashboard built
  in and no terminal needed. See [the desktop app](/docs/getting-started/desktop).
- **Command line**, for developers: `roscoe run <workflow>` runs a workflow and
  streams the trace as it happens. See [running from the CLI](/docs/running/cli).
- **From Claude**: ask Claude to build and run workflows for you, in plain
  language. See [Claude integration](/docs/running/mcp).

## Where to next

- [Install Roscoe](/docs/getting-started/install) — get the app or the CLI.
- [Write your first workflow](/docs/getting-started/first-workflow) — build one
  end to end.
- [Core concepts](/docs/getting-started/concepts) — the handful of terms used
  throughout.
- [Troubleshooting common errors](/docs/troubleshooting/common-errors) — when
  something goes sideways.