Pincushion for Antigravity.
Visual feedback that lands in your IDE. Stakeholders drop pins on your live app; your Antigravity agent reads them via MCP — selector, DOM snippet, screenshot, viewport, thread, project context, and acceptance criteria — and ships the change from inside your editor.
Why this pairing works
Antigravity is agent-first by design — you hand the agent a task and review its work, rather than driving every edit yourself. That's exactly the handoff shape Pincushion produces. A pin isn't a comment for you to interpret; it's a complete work packet the agent can execute end to end. Feedback from someone looking at the live product becomes a task the agent picks up with full context already attached.
The flow
- Stakeholder drops a pin on any element of your live app via the Chrome extension or the one-line script-tag widget
- You open Antigravity and run the
/pinsor/implementworkflow (Pincushion ships Antigravity workflows in.agent/workflows/) - The agent receives each pin's full work packet via MCP — no copy-pasting screenshots, no translating feedback into prompts
- The agent finds the source code, makes the change, calls
fix_and_resolvewith branch + PR URL - Deploy hook auto-resolves the pin on production. Stakeholder sees "Resolved in PR #142, deployed at 3:24pm"
Install (60 seconds)
Add Pincushion to ~/.antigravity/mcp.json (or via Antigravity's MCP settings):
{
"mcpServers": {
"pincushion": {
"command": "npx",
"args": ["pincushion-mcp", "--project-dir", ".", "--cloud-sync"]
}
}
}
Run npx pincushion-mcp login once to link your license key — after that, pins dropped on your deployed site sync to the agent automatically. If you can't run a local server, the hosted endpoint at https://pincushion.io/api/mcp exposes the cloud-side tools (authenticate with Authorization: Bearer <license key> or ?token=).
Slash commands available after install
/pins— show all open pins on this project/implement— run the full approved-pin loop (claim → fix → resolve)/resolve <id>— claim and close one pin/critique <url>— Pincushion AI drops ≤3 UI/copy/a11y pins on the page/feedback-summary— project overview
Your agent already does the implementing. Pincushion makes sure the feedback arrives as something it can implement — selector, DOM, screenshot, viewport, thread.
Feedback as tasks, not conversations
Antigravity's model is task-and-artifact: you queue work, agents execute, you review the plans and diffs they produce. Screenshot-based feedback fights that model — someone has to sit between the reviewer and the agent, converting prose into task specs. Pins remove the intermediary. Each approved pin is already a bounded, verifiable task: this element, this page, this viewport, this requested change, this definition of done. A batch of pins from a review session becomes a batch of agent tasks with no specification step, and because claim_pin marks work in progress, parallel agents won't trip over the same pin.
The trail survives the run, too. Each resolved pin carries its commit SHA, branch, and PR URL; the deploy hook stamps the production URL when the fix ships; and Pincushion AI can re-check the page post-deploy and write a verified/regressed verdict back to the pin (Pro). When an agent did the work, that receipt chain is what makes the review trustworthy.
What's in the work packet
- CSS selector — the agent greps the repo with it and lands on the right component
- DOM snippet — disambiguates repeated components and reveals page state
- Screenshot + viewport — what the reviewer saw, at the size they saw it
- Thread + acceptance criteria — what "done" means, in the stakeholder's words
- Project context + likely files — so the first file opened is usually the right one
Pricing
Reviewers and stakeholders are always free — unlimited, no accounts to buy. The developer free tier includes one project, unlimited manual pins, and 25 AI actions a month, and every signup starts a 14-day Pro trial. Pro is $19/mo: unlimited automation, realtime push to your IDE, unlimited projects, and access controls.
Also using other editors? Pincushion works the same way in Cursor, Claude Code, Windsurf, and VS Code. New to the idea? Read what a visual feedback MCP server actually does.