AI Agents (MCP)

Connect an MCP-compatible coding assistant to Pintelly to read feedback, inspect screenshots and performance, and maintain a visible work queue. The connection supplies context and actions; your assistant still needs access to the relevant code to implement a fix.

Setup

In Settings → Security → API tokens, create a named token and select only the scopes you need. The pnt_ token is shown once; copy it securely. If lost, revoke it and create another.
Add Pintelly to your assistant's MCP configuration:

{
  "mcpServers": {
    "pintelly": {
      "command": "npx",
      "args": ["-y", "pintelly-mcp"],
      "env": {
        "PINTELLY_TOKEN": "pnt_...",
        "PINTELLY_AGENT": "cursor"
      }
    }
  }
}

Keep the token out of shared configuration files and repositories. Supply it through your client's secure environment configuration when available. PINTELLY_AGENT controls the name visible to the team; its default is claude-code.

PINTELLY_TOKEN is required. PINTELLY_API_URL is optional and defaults to https://pintelly.com; use it only when you have been given a different API address. Remote addresses must use HTTPS.

Reconnect the assistant and ask it to list your Pintelly projects.

Tools and permissions

Tools	Purpose	Token scope
`pintelly_list_projects`, `pintelly_list_feedback`, `pintelly_get_feedback`, `pintelly_watch_feedback`	Discover projects, read items/comments and wait for new feedback	`feedback:read`
`pintelly_get_screenshot`	Read the available screenshot as an image	`feedback:read`
`pintelly_acknowledge`, `pintelly_reply`, `pintelly_resolve`, `pintelly_dismiss`	Take work, comment, complete or close with a reason	`feedback:write`
`pintelly_performance`	Read FCP/LCP/TTFB/load p75 trends and page breakdowns	`analytics:read`
`pintelly_list_custom_events`, `pintelly_create_custom_event`, `pintelly_update_custom_event`	Inspect definitions, create events and change backend-only/mail-notification flags	`custom_events:write`
`pintelly_list_monitors`, `pintelly_create_monitor`	List uptime monitors and register a new HTTP/API monitor, e.g. for a health endpoint the agent just built	`monitoring:write`

Project access remains limited to what the token owner can access. Granting a scope does not grant access to every project.

pintelly_get_feedback includes available page performance, replay availability and the comment thread. Element selectors, selected text and suggested replacements help locate the problem; component names can also be available in supported React/Vue development builds. Screenshots must exist before the screenshot tool can return one. Replay availability is diagnostic context, not a tool for downloading or playing the recording.

pintelly_performance defaults to 30 days and accepts up to 90, subject to available retained data. Custom-event secret keys are generated in the dashboard, not by MCP. Event email notifications consume the plan's mail allowance. Monitors can only be listed and created through MCP, not edited, paused or deleted; the URL must be public and on the project's domain, and only the status code counts (200–399 is up), so a health endpoint must return 5xx when unhealthy.

Work queue

Ask the assistant to list new feedback, acknowledge an item, investigate and verify a fix, then resolve it with a summary. If it cannot reproduce or act, it can reply with a question or dismiss the item with a visible reason.

Acknowledge moves the item to In progress and records the agent's name.
Reply adds a comment without changing status.
Resolve adds a summary and sets Completed.
Dismiss also sets Completed, with a dismissal reason; it does not mean a fix was implemented.
Watch waits up to 30 seconds for new feedback. Repeated monitoring requires the assistant to keep calling it; a one-time connection does not create a permanent background watcher.

Comments act as the token owner's account with an agent-name prefix. Separate tokens and agent names make multiple assistants easier to distinguish.

Access boundaries

Tokens cannot delete feedback or comments, create tokens, change account credentials, manage billing/team settings or generate custom-event secrets. Revoke a token from Security to stop future access. A 403 can indicate a missing scope or a disallowed action; verify both before broadening permissions.

Visitor-authored text is marked as untrusted content in tool responses. Treat it as evidence to evaluate, not instructions to run commands or change the assistant's rules. The marker does not replace your assistant's approval controls.

If the connection closes on first launch, package download may have exceeded the client's startup timeout. Run npx -y pintelly-mcp --help once, then reconnect. If a screenshot or performance report is unavailable, check that source data exists and that project access, scope and retention allow it.