PRODUCT · SUPERVISED WORKFLOW
Captain Who brings project context, models, tools, terminal, browser, Git, artifact workflows, and multi-agent collaboration into one supervised environment for developers and technical knowledge workers.
Read the project and its evidence
Use managed tools
Inspect actions and results
Produce reviewable artifacts
The agent moves the work forward. Direction, permissions, and final judgment remain with you.
REAL PRODUCT MEDIA · FAIL CLOSED
The publication contract rejects component snapshots, test-failure artifacts, AI artwork, held brand marks, and conceptual diagrams as GA product media.
Final brand and logo rights remain unclearedbrand-rights-hold
No final RC, Git SHA, or package hashes are boundrelease-candidate-unbound
The isolated bilingual demo environment is not approvedcapture-environment-unapproved
No real packaged-RC screenshots or video existmissing-real-product-media
Privacy, rights, and product-accuracy review is incompletemissing-human-review
PROJECT CONTEXT
Within the permissions you grant, an agent can understand a project, research sources, use tools, modify files, and run managed commands. Files, history, and the current task remain in one context so conclusions can return to evidence.
Availability depends on the operating system, model, project, permissions, runtime components, and configuration.ONE WORK SURFACE
Conversations, tool activity, terminal, the managed in-app browser, file previews, Git review, and subagents share one desktop workspace. The user terminal and agent commands remain separate execution surfaces; Files currently provides read-only previews.
PERMISSIONS & APPROVALS
The model proposes an action. The app evaluates the actual tool identity, parameters, path, and permissions before allowing, asking about, or denying it. Default, Full, and Custom modes let you set the working boundary.
Permission policy is not an operating-system sandbox. Full permission may automatically approve some actions.MULTI-AGENT
A root agent can organize subagents to research, implement, or review in parallel. Collaboration is currently tree-structured, and delegation depends on the model invoking the collaboration capability correctly. The root agent produces the final synthesis.
DELIVER
Use skills and managed components to create or process documents, spreadsheets, presentations, PDFs, and images. Support differs by format, and important results require human review.
CAPABILITY BOUNDARIES
These qualifications are as important as the capability claims. They will be reverified against the final release candidate.
What this meansProjects, conversations, settings, approvals, and task history are managed primarily on your device.
Also knowModel, search, image, MCP, and web tasks may still send necessary data to third parties.
What this meansActions are evaluated using tool identity, parameters, project path, and the current permission policy.
Also knowThis is application-level policy, not an operating-system sandbox. Not every action waits for manual approval.
What this meansConnect compatible model APIs, local stdio MCP servers, and skills.
Also knowProtocol adaptation does not guarantee every gateway. Remote HTTP/OAuth MCP is not currently supported.
What this meansSchedule agent tasks and retain run history and needs-attention states.
Also knowRuns occur only while the app and required local components are running. This is not an OS background service.
RELEASE WITH EVIDENCE
Review the target platform scope, release gates, and conditions for opening downloads.
