AI assistants (MCP)

Set it up

1. In the dashboard, open Account → AI assistants and press Create token. Name it after the assistant or computer it is for, and choose Read only or Can edit triggers. Read only is the default.
2. Copy the token. It starts with pxk_ and is shown once.
3. Add the server to your client, as below. The address is https://api.pixelary.app/mcp.
4. Ask: “List my Pixelary sites.”

Claude Code

claude mcp add --transport http pixelary https://api.pixelary.app/mcp \
  --header "Authorization: Bearer pxk_your_token"

Run /mcp in Claude Code to check the connection. To share the setup with a project without sharing the token, put this in .mcp.json and set PIXELARY_TOKEN in your shell:

{
  "mcpServers": {
    "pixelary": {
      "type": "http",
      "url": "https://api.pixelary.app/mcp",
      "headers": { "Authorization": "Bearer ${PIXELARY_TOKEN}" }
    }
  }
}

Cursor

Put this in ~/.cursor/mcp.json, or in .cursor/mcp.json for one project, and set PIXELARY_TOKEN in your environment:

{
  "mcpServers": {
    "pixelary": {
      "url": "https://api.pixelary.app/mcp",
      "headers": { "Authorization": "Bearer ${env:PIXELARY_TOKEN}" }
    }
  }
}

VS Code

Put this in .vscode/mcp.json. VS Code asks for the token once and hides what you type:

{
  "inputs": [
    { "type": "promptString", "id": "pixelary-token", "description": "Pixelary token", "password": true }
  ],
  "servers": {
    "pixelary": {
      "type": "http",
      "url": "https://api.pixelary.app/mcp",
      "headers": { "Authorization": "Bearer ${input:pixelary-token}" }
    }
  }
}

Windsurf

Put this in mcp_config.json (~/.config/devin/mcp_config.json on macOS and Linux) and set PIXELARY_TOKEN in your environment:

{
  "mcpServers": {
    "pixelary": {
      "serverUrl": "https://api.pixelary.app/mcp",
      "headers": { "Authorization": "Bearer ${env:PIXELARY_TOKEN}" }
    }
  }
}

Claude apps and other clients

Connectors in the Claude apps (web, desktop, mobile) sign in with OAuth. Anthropic also has a Request headers option for adding a custom connector with a fixed token, in beta for some organizations. If you see it, enter Authorization and Bearer pxk_your_token. In a Team or Enterprise organization, everyone who uses that connector shares the credential and acts as you, so use a read-only token there, or wait for OAuth.

Any other client that speaks Streamable HTTP and lets you set an Authorization header should work. We have listed only the clients whose setup we checked in their own docs.

Tools

Every tool works on the sites you can open in the dashboard, with the role you have there.

Clients that show prompts also get audit_my_tracking and weekly_report, two ready-made requests.

No tool can reach access tokens, billing, the team, domains, consent settings or data export, delete anything, or show other people’s email addresses. Those stay in the dashboard.

What to ask

• “Audit the tracking on brightside.studio and tell me what to fix first.”
• “How many purchases did we send to Meta this week, compared with last week?”
• “Which events failed in the last day, and what did the platform say?”
• “Is any pixel still in test mode?”
• “Add a Lead trigger for the contact form on /contact.”
• “Switch off the Newsletter trigger.”

How changes work

A write tool works in two steps, on purpose. The first call checks the change with the same rules as the dashboard, changes nothing and returns a preview:

{
  "status": "preview",
  "applied": false,
  "preview": "Create a form trigger on Brightside: sends Lead to every pixel when {\"selector\":\"form#contact\"} matches. It starts switched on, with no republish needed.",
  "confirm_id": "pxc_…",
  "expires_at": "2026-11-02T10:10:00.000Z",
  "next_step": "Nothing has changed yet. Show this preview to the user…"
}

The assistant shows you the preview. Only after you agree does it call the tool again, with the same arguments plus confirm_id. The change is saved at once and reaches your site without a republish.

A confirm_id works once, for 10 minutes, with the token that asked for it, on that site and for exactly that change. A different payload, or a trigger that changed since the preview, is refused and nothing changes. The change shows on the Triggers screen like any other.

Security

• A token acts as you. It has your role on each site: Owner, Editor or Viewer. If you lose access to a site, the token does too, at the next call.
• Treat it like a password. Keep it in an environment variable or your client’s secret prompt, not in a file you commit. Pixelary stores only a hash, so a lost token cannot be shown again: revoke it and create a new one.
• Revoke any time in Account → AI assistants. It stops working at once. Tokens do not expire on their own; the list shows when each was last used, so you can spot the ones nobody needs. Signing out of the dashboard does not revoke them.
• What your assistant sees goes to its provider. Stats, event names, page paths and trigger settings are sent to the AI service you connected, under your agreement with it. The event log holds no contact details or IP addresses, and addresses in page paths are masked.
• Log text is data, not instructions. Event names, page paths and platform messages come from website visitors and ad platforms. Pixelary cleans and caps them and marks them as untrusted, and the tools tell assistants not to follow them. A careful client still matters: this is why changes need a preview.
• Browsers are kept out. Requests from a web page on another site are refused with 403, as the MCP specification asks.

Limits

Troubleshooting

For developers

The endpoint speaks Streamable HTTP with one JSON response per POST: no sessions and no event stream, so GET and DELETE answer 405. It supports protocol revision 2026-07-28 and, for clients that open with initialize, 2025-11-25, 2025-06-18 and 2025-03-26. It offers tools and prompts only.