Share agent output without screenshots or Loom

How people do this today

The usual handoff begins with a local preview. A coding agent builds a dashboard or walkthrough, the developer opens localhost, captures a few screenshots or records a Loom video, and sends the images with an explanation. That is enough to prove pixels existed at one moment. It is not enough for a teammate who needs to filter data, follow links, test mobile layout or return after the next revision. Teams sometimes keep the local server running and share their screen live, but the viewer still depends on the developer's laptop and calendar. The better unit of communication is the artifact itself: the files, behavior and current data behind one address.

Where it breaks

• The evidence freezes immediately. A screenshot cannot show interactions, loading states, responsive behavior or the record behind a chart. Once the agent changes the output, the image and the working version disagree.
• The route belongs to a laptop. A localhost URL works only on the machine running the process. A screen recording hides that dependency, so a reviewer discovers the real access problem only when asking to try the tool.
• Feedback loses its target. Comments arrive in chat with phrases such as the second card or timestamps from a video. The agent cannot reliably connect those notes to the current artifact, and reviewers cannot see whether a later state resolves them.
• Access becomes all or nothing. Screen sharing exposes the whole desktop, while an image can be forwarded without control. Neither creates a deliberate public, password-protected or invite-only boundary around the artifact.
• Useful data is omitted. A database-backed tool may be reduced to a picture because reviewers cannot reach its local service. They approve the appearance without testing the actual workflow.

The replacement workflow

1. Download and sign in. Download the Mac app from rungrain.com, then run grain login. Grain is Mac first; Windows has a waitlist. Keep the coding-agent plan you already pay for.
2. Initialize the current project. From the folder containing the agent's work, run grain init. This creates the workspace and writes agent configuration to .mcp.json for Claude Code, .codex/config.toml for Codex and .cursor/mcp.json for Cursor. Pass --no-agent-config only when you intentionally want to skip those files.
3. Give the agent an artifact-shaped prompt. For example: Turn the release notes and screenshots in this folder into an interactive launch walkthrough. Add a page for each change, mobile navigation and a comment-ready review view. Do not depend on local-only paths. The agent writes the artifact files in the initialized workspace.
4. Publish the files. Run grain push. This moves the current workspace state into the publishing flow. If the artifact is an app, it can include Postgres, logins and scheduled jobs rather than flattening the result into an image.
5. Mark the review state. Run grain checkpoint -m "launch walkthrough ready for review". A named checkpoint gives reviewers a clear state and makes the next update intentional.
6. Choose access and receive the URL. Run grain share --public for anyone with the link, or grain share --password <pw> for password access. Invite-only sharing is also available. The share URL has the form https://<slug>.share.rungrain.com.
7. Send the link with one request. Ask reviewers to open the artifact, test the relevant path and leave comments on the page. The agent can read those comments, update the files, and repeat push and checkpoint without creating a new trail of screenshots.

What it costs

Grain's free tier includes the Mac app, CLI, unlimited published artifacts, ten sandboxes that sleep when idle, one always-on workspace and a self-hostable core. No card is required, and Grain does not sell or mark up agent tokens. You continue paying the provider of Claude Code, Codex, Cursor or the agent you already chose. Pro costs $65 per user per month and adds more always-on workspaces, team seats, shared spaces and custom domains. Enterprise is priced by contact for deployment in the customer's cloud with SSO, audit, policy controls and direct support. The useful comparison is not minutes of video versus pages of output. It is whether reviewers need to watch an explanation or can operate the artifact itself.

Limits

• Grain is not a recording tool. Keep a video when narration, cursor movement or an unreproducible desktop sequence is itself the evidence.
• Grain is not a code editor or coding agent. Your agent creates and revises the artifact; Grain publishes and keeps it.
• Public links are reachable by anyone who has the URL. Use a password or invited access when the content should not be public.
• Windows is not available yet. The desktop product is Mac first, with a Windows waitlist; the self-hostable core runs on Linux.
• A shared artifact does not replace production deployment for a service with its own CI, infrastructure policy or operational requirements.

Frequently asked questions