Cube: human-friendly, self-improving AI workflows
> Your agent builds, runs, and improves the workflow. Step in occasionally to approve key decisions or inspect the process.
Cube brings AI judgment, program execution, and human review into one workflow. Your agent builds and runs the process, then uses your feedback and execution results to work out how to improve it.
What Cube provides
- Agents interpret goals, make judgments, and build or revise workflows.
- Programs execute repeatable steps with defined inputs and outputs.
- Human review points bring key decisions and their context to you.
- Sheets show structured records and results; docs hold context, drafts, and evidence; buttons run defined actions.
- A desk is the persistent workspace that brings these readable, actionable components together.
You can inspect the workflow behind the work when you need to understand or change how it runs. The agent supplies the intelligence and works with Cube through Cube's own command line, run inside Cube's engine — not through the agent's own tool-calling loop reaching out to Cube.
How feedback improves a workflow
For example, a sales workflow qualifies leads and prepares outreach. You reject an agency lead and explain, "Focus on product companies." Your agent can revise the qualification step before the next batch. Reply outcomes and failed executions provide further evidence for changes to targeting, follow-up, or execution.
The change is to the underlying workflow, so later runs can use it. Its effect still has to be checked against the results; a saved correction alone does not establish that the workflow improved.
Example workflows
These illustrate how Cube can be used; they are not links to prebuilt templates.
- Sales pipeline: qualify prospects, prepare outreach for review, and revise the workflow using decisions and reply outcomes.
- Research: collect sources, compare findings, and update the research process as new evidence arrives.
- Job search: shortlist roles, prepare applications, and refine the search from your preferences and results.
Choosing a resident
Cube runs the agent, not the other way around: open a desk and choose a Claude or Codex account. Cube signs it in and starts that resident inside its own engine as a process it manages — there is no separate agent environment to configure, and no MCP server to add anywhere. Read [how a resident works](https://www.usecube.io/mcp.md) for what that process can do and where it runs.
A resident's availability follows Cube's own deployment and the signed-in account's provider quota; this documentation does not promise unrestricted subscription usage.
Start
Open [Cube](https://www.usecube.io/new) and choose a resident. The new-desk screen is initially empty; the first message or Add cubes creates the working desk.
Reading this introduction does not authorize creating a desk, signing in an account, or changing work. Follow the user's actual request when taking actions.
[Homepage](https://www.usecube.io/) · [Agent documentation index](https://www.usecube.io/llms.txt)