Pincushion

Pincushion MCP Server

The implementation-context layer for AI-native development. Stakeholders drop visual pins on any page of your live app; your AI coding agent reads each pin through MCP and ships the fix — in Claude Code, Cursor, VS Code, Windsurf, or any MCP client.

What makes a Pincushion pin different

A pin isn't a feedback item — it's an agent work packet. Each one carries everything an agent needs to implement the change without a back-and-forth:

• URL + element selector — exactly what, exactly where
• Screenshot + viewport + DOM snippet — the visual and structural context
• Thread + project context — the conversation and the codebase it lives in
• Likely files + acceptance criteria — where to look, and how to know it's done

The loop closes itself: a stakeholder pins it → your agent reads it via MCP and fixes it in your IDE → the resolve records the commit, branch, and PR → an optional post-deploy critique verifies the fix actually landed.

This server is also how Pincushion AI runs design/copy/a11y critiques on a live page and writes the pins straight back onto it.

Hosted quick start

Use Node.js 22.22+ within Node 22, or Node 24, and run this in your app's repository:

npx pincushion-mcp setup

The guided wizard signs you in to Pincushion, asks you to review the project URLs, and explicitly choose an editor. It writes supported project MCP configuration when safe; Manual, Skip, or a conflicting existing configuration can leave it unchanged. Restart the editor and ask your agent to call get_project_context and get_actionable_pins; configuration written alone does not verify a connection. If you already have a critique report, run its npx pincushion-mcp claim <report-token> command from the owning app's repository instead, then verify a real pin read.

Pincushion hosts cloud sync. The normal setup does not require your own Supabase project, database table, or backend API key. Sign-in supplies the existing account's cloud credentials automatically; npx pincushion-mcp login is the recovery command if sign-in is missing or expired. Do not paste credentials into issue reports.

Setup connects feedback access; it does not generate an AI critique. Critique execution requires a supported coding-agent workflow and real page captures. A public report lets reviewers add pins without an extension. Install the optional Chrome extension only when you want to place feedback directly on other web pages.

Use the maintained MCP setup guide for exact Claude Code, Cursor, VS Code, and Codex commands, and the public documentation for current plan limits and troubleshooting. The optional legacy self-host connection flags below are not prerequisites for hosted setup.

CLI Flags

npx pincushion-mcp [flags]

Flag	Description	Default
--project-dir PATH	Root directory containing .feedback/	Current working directory
--sync-url URL	Optional legacy self-host adapter endpoint; not hosted setup	None
--api-key KEY	Optional legacy self-host adapter credential; not hosted setup	None
--license-key KEY	Explicit account credential override; normal hosted setup uses sign-in	Signed-in account when available
--rest	Enable REST API mode	Disabled (uses MCP/stdio)
--port PORT	Port for REST API server	3456

Examples

Local project:

npx pincushion-mcp --project-dir /path/to/project

Optional legacy self-host adapter (not hosted Pincushion setup):

npx pincushion-mcp \
  --project-dir /path/to/project \
  --sync-url https://abcd1234.supabase.co/api \
  --api-key sb_project_key_abc123...

REST API server:

npx pincushion-mcp --rest --port 8080

---

Tools

get_annotations

Retrieve annotations from .feedback/. Filter by page, component, or status.

Parameters:

• pageUrl (string, optional) — Filter by page URL (partial match)
• componentName (string, optional) — Filter by LWC component name
• status (string, optional) — Filter by open, in-progress, or resolved

Example:

await mcp.callTool('get_annotations', {
  componentName: 'wmlHomePage',
  status: 'open'
});

search_annotations

Full-text search across all annotations, comments, selectors, and tags.

Parameters:

• query (string, required) — Search term

Example:

await mcp.callTool('search_annotations', {
  query: 'button label'
});

get_feedback_summary

High-level rollup of all feedback: counts by status, priority, page, and component.

Example:

await mcp.callTool('get_feedback_summary', {});

get_component_feedback

Get all feedback for a specific LWC component with a plain-language summary.

Parameters:

• componentName (string, required) — LWC component name

Example:

await mcp.callTool('get_component_feedback', {
  componentName: 'wmlHomePage'
});

resolve_annotation

Mark an annotation as resolved after fixing the issue.

Parameters:

• annotationId (string, required) — Annotation ID
• comment (string, optional) — Resolution message
• resolvedBy (string, optional) — Name to attribute resolution (default: "AI Agent")

Example:

await mcp.callTool('resolve_annotation', {
  annotationId: 'ann_abc123',
  comment: 'Updated button label in line 42 of wmlHomePage.js'
});

add_agent_reply

Add a reply to an annotation thread (e.g., ask clarifying questions).

Parameters:

• annotationId (string, required) — Annotation ID
• body (string, required) — Reply message
• author (string, optional) — Author name (default: "AI Agent")

Example:

await mcp.callTool('add_agent_reply', {
  annotationId: 'ann_abc123',
  body: 'Is this button in the main navigation or sidebar?'
});

fix_and_resolve

Combine fixing code and marking an annotation as resolved in one call. Optionally records commit / branch / PR metadata so the dashboard can backlink to what shipped.

Parameters:

• annotationId (string, required) — Annotation ID
• fixDescription (string, required) — Description of the fix
• filePath (string, optional) — File where fix was applied
• lineNumber (number, optional) — Line number of the fix
• commitSha (string, optional) — Commit SHA that landed the change
• branchName (string, optional) — Branch the commit was made on
• prUrl (string, optional) — Pull request URL (GitHub/GitLab/Bitbucket; shape-validated)

Example:

await mcp.callTool('fix_and_resolve', {
  annotationId: 'ann_abc123',
  fixDescription: 'Updated button label to match design spec',
  filePath: 'src/components/wmlHomePage.js',
  lineNumber: 42,
  commitSha: 'abc123def456',
  branchName: 'pincushion/checkout-fix',
  prUrl: 'https://github.com/acme/app/pull/142'
});

get_implementation_packet

Fetch a single implementation packet for one page URL — selector list, full pin payloads, suggested branch name, and traceability config. Use when an agent wants to batch-fix one page in a single branch.

await mcp.callTool('get_implementation_packet', { pageUrl: '/checkout' });

assign_pin_to_agent

Dispatch a pin straight to your local coding agent. Promotes the pin to ready if not already, marks pending_implementation, and writes a .feedback/.agent-queue/<id>.json trigger file that agent-loop.mjs picks up and shells out to Cursor / Claude Code / Codex.

await mcp.callTool('assign_pin_to_agent', { annotationId: 'ann_abc123' });

link_pin_deploy

Attach a deploy URL to a resolved pin. Typically called by the deploy-hook edge function once production includes the fix, but available manually too.

await mcp.callTool('link_pin_deploy', {
  annotationId: 'ann_abc123',
  deployUrl: 'https://acme-app.vercel.app'
});

record_pin_verification

Write Pincushion AI's post-deploy verdict back to the pin. Called by the critic agent after /critique-latest-deploy runs against a fresh deploy.

await mcp.callTool('record_pin_verification', {
  annotationId: 'ann_abc123',
  status: 'verified',  // or 'regressed' or 'inconclusive'
  notes: 'Button matches the primary token. No regression on adjacent CTAs.'
});

get_time_to_fix_metrics

Pro/Team feature — Free callers get sample size + upgrade hint. Median + p25/p75 of pin-to-resolve duration, with a 5-pin minimum so the metric is never noise.

await mcp.callTool('get_time_to_fix_metrics', { scope: 'project', projectId: 'pc_proj_abc' });
// → { sampleSize, thresholdMet, median, p25, p75, medianHuman, ... }

get_setup_instructions (NEW)

Get setup and configuration instructions for all supported agents.

Example:

await mcp.callTool('get_setup_instructions', {});

Slack and Microsoft Teams integrations

Pincushion can notify Slack or Microsoft Teams through project-scoped incoming webhooks. The defaults are intentionally quiet and Figma-inspired: notify when a pin is ready for implementation, when someone is @mentioned, and when a collaborator adds follow-up on work already being handled. Every newly dropped pin and every resolution are opt-in events.

Recommended use cases:

• Developer channel: pin_ready and follow_up
• Design or PM channel: mention and optionally resolved
• Launch or QA channel: pageUrlPatterns plus pin_ready, follow_up, and resolved
• Temporary incident channel: enable a focused subscription, then pause it after the ship window

Example:

await mcp.callTool('configure_collaboration_integration', {
  projectId: 'my-project',
  provider: 'slack',
  webhookUrl: 'https://hooks.slack.com/services/...',
  targetLabel: '#product-feedback',
  events: ['pin_ready', 'mention', 'follow_up'],
  pageUrlPatterns: ['staging.example.com/checkout'],
  sendTest: true
});

For Slack, use create_slack_install_link when the hosted Slack app secrets are configured. It returns an Add-to-Slack URL; after approval, Slack returns the incoming webhook and Pincushion stores it automatically.

Use list_collaboration_integrations to audit configured destinations, remove_collaboration_integration to disconnect one, and preview_collaboration_notification to see the payload shape before adding a real webhook. Webhook URLs are stored server-side and returned only as masked values.

---

Auto-Agent Loop (Optional)

For agents that don't watch the file system (Claude Code, Cursor, generic), agent-loop.mjs polls .feedback/.agent-queue/ and dispatches new pins to the configured agent automatically.

# from inside the pincushion-mcp directory
npm run agent-loop -- --project-dir /path/to/your/project

# or directly
node agent-loop.mjs --project-dir /path/to/your/project [--agent claude-code|cursor|generic] [--interval 3000]

The bridge (server.js) writes one trigger file per approved pin into .feedback/.agent-queue/. The loop reads them, builds a prompt with the pin's thread + element selector, and shells out to the chosen agent. The agent uses MCP tools (claim_pin → fix → fix_and_resolve) and the queue file is removed when the pin closes.

detectAgent() auto-detects claude or cursor on the PATH; falls back to generic (writes the prompt to .feedback/.agent-prompt and stdout). Run with --interval 3000 to control poll cadence.

---

Local File Structure

The server reads annotations from .feedback/ in your project:

.feedback/
├── annotations/
│   ├── example-com-login.json
│   ├── example-com-dashboard.json
│   └── ...
└── index.json

Each annotation file contains:

{
  "pageUrl": "https://example.com/login",
  "pageTitle": "Login",
  "annotations": [
    {
      "id": "ann_abc123",
      "status": "open",
      "priority": "high",
      "tags": ["design", "accessibility"],
      "createdAt": "2026-03-19T10:30:00Z",
      "element": {
        "lwcComponent": "wmlLoginForm",
        "selector": ".login-button",
        "textContent": "Sign In"
      },
      "thread": [
        {
          "author": "Design Team",
          "timestamp": "2026-03-19T10:30:00Z",
          "body": "Button label should say 'Sign In' not 'Login'",
          "type": "comment"
        }
      ]
    }
  ]
}

---

Optional legacy self-host connection

The --sync-url and --api-key flags select a separate legacy sync adapter. They are not needed to use Pincushion's hosted service. The older instructions to create a Supabase project and annotations table are not a supported hosted onboarding path; simply creating those resources does not establish a compatible backend. Use this adapter only with an already compatible operator-managed endpoint. See hosted setup for the default path and current pricing for account plans.

---

Troubleshooting

"Module not found" error

Make sure you have a supported Node.js runtime: Node 22.22+ within Node 22, or Node 24:

node --version

Install dependencies:

npm install @modelcontextprotocol/sdk

Annotations not appearing

Check that .feedback/ exists in your project directory:

ls -la .feedback/

If it doesn't exist, create it and add some test annotations, or the extension will create it when you pin your first feedback.

Hosted feedback not appearing

Run npx pincushion-mcp login if sign-in is missing or expired, then rerun setup in the intended repository. Confirm the acknowledged project and registered page URLs, restart the selected editor, and ask it to read get_project_context and get_actionable_pins. Preserve existing project identity; do not create a replacement project or backend to work around an access failure. See the maintained troubleshooting guide.

Agent cannot find the server

Rerun npx pincushion-mcp setup from the intended repository, preserving the acknowledged project identity and existing configuration. Follow the maintained editor setup guide if setup leaves a conflicting configuration unchanged. Restart the selected editor and verify a real project/pin read. Avoid replacing a project-scoped configuration with an unscoped global entry.

---

Development

Clone the repository and install dependencies:

git clone https://github.com/jcooley8/pincushion-plugin.git
cd pincushion-plugin
npm install

Run the server:

npm start

Or with test data:

npm start -- --project-dir ./test-feedback

---

License

MIT License. See LICENSE file for details.

---

Support

• Issues: GitHub Issues
• Docs: Pincushion Documentation

---

Changelog

v1.0.0 (March 2026)

• Initial release
• Support for Cursor, Claude Desktop, Claude Code, VS Code, Windsurf, Antigravity
• Local .feedback/ file support
• Supabase remote sync
• REST API wrapper for non-MCP clients
• New tools: fix_and_resolve, get_setup_instructions