Pedra MCP Server

Pedra MCP Server

Official Model Context Protocol server for the Pedra API — use Pedra's AI real-estate photo editing (virtual staging, renovation, room emptying, enhancement, sky replacement, object removal/blur, property videos, and hosted 360° virtual tours) directly from Claude, ChatGPT, Cursor, and any other MCP client.

It exposes one tool per API endpoint. Each photo and video tool is a single blocking call that returns the final asset URL(s) — there are no job IDs to poll. Virtual tours build in the background: pedra_create_virtual_tour returns a tourId, and pedra_get_virtual_tour is polled until it's ready.

Quick start

The server reads your Pedra API key (Settings in your Pedra account) from the PEDRA_API_KEY environment variable. No account yet? Start it without a key — see No Pedra account yet?.

The server runs over stdio and is published to npm, so most clients just run it with npx — no global install needed.

Claude Desktop (one-click)

Download the latest pedra-mcp.mcpb from Releases and double-click it (or drag it into Claude Desktop → Settings → Extensions). Claude installs the bundled server and prompts for your PEDRA_API_KEY — no JSON editing.

Claude Desktop (manual)

Or add this to your claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "pedra": {
      "command": "npx",
      "args": ["-y", "@pedra-ai/mcp"],
      "env": { "PEDRA_API_KEY": "your-api-key" }
    }
  }
}

Cursor

In ~/.cursor/mcp.json (or .cursor/mcp.json in a project):

{
  "mcpServers": {
    "pedra": {
      "command": "npx",
      "args": ["-y", "@pedra-ai/mcp"],
      "env": { "PEDRA_API_KEY": "your-api-key" }
    }
  }
}

Any MCP client

Run the binary directly with the key in the environment:

PEDRA_API_KEY=your-api-key npx -y @pedra-ai/mcp

Or install it:

npm install -g @pedra-ai/mcp
PEDRA_API_KEY=your-api-key pedra-mcp

Smithery

You can also install and configure Pedra automatically via Smithery:

npx -y @smithery/cli install @pedra-ai/mcp --client claude

(swap claude for cursor, windsurf, etc.) Smithery prompts for your PEDRA_API_KEY and writes the client config for you.

No Pedra account yet?

Start the server without PEDRA_API_KEY, for example in Claude Code:

claude mcp add pedra -- npx -y @pedra-ai/mcp

Then just ask for something ("stage this photo with Pedra"). Without a key the server adds two tools, and the assistant does the rest:

1. pedra_request_access with your email. Pedra emails you a link (valid 30 minutes): a new account chooses a password there, an existing one clicks Allow. Nothing is created until you click, and you can decline.
2. pedra_check_access until you've confirmed. Once approved, every Pedra tool works for the rest of the session. The key is kept in memory only, never written to disk; the tool result shows it along with how to keep it: add it as PEDRA_API_KEY to the server's env (e.g. claude mcp add pedra -e PEDRA_API_KEY=<key> -- npx -y @pedra-ai/mcp, or paste it into the extension's settings in Claude Desktop).

Until then, the other tools answer "No API key yet: call pedra_request_access with the user's email". Inbox confirmation is required, disposable email domains are refused, and requests are rate-limited. New accounts start with Pedra's free trial credits.

(In ChatGPT and claude.ai, use Pedra's hosted connector instead: connecting it runs a sign-in where new users can create an account.)

Tools

The nine image-editing tools (enhance through blur) also take propertyId, to save the result into that property's gallery, and name, to save the original photo there under that name in the same call. The result then includes source (imageId, name) of the original. preserveAspectRatio returns the result at the input's exact size.

Most image tools take an imageUrl plus a few optional parameters; see each tool's input schema in your MCP client. The imageUrl (and maskUrl, and each create_video frame) accepts any of:

• a public https:// URL,
• a data: URI, or
• an absolute path to a local image file — the server reads it off disk and inlines it as base64 for you, so you can point a tool at a file you just dragged in without hosting it first (.jpg, .jpeg, .png, .webp, .gif, .bmp, .tif/.tiff, .heic/.heif, .avif; up to 40 MB).

Note: an image pasted into the chat is not a file path, so it can't be forwarded to the tool — drag in a file, or save the paste and pass its path. Photos on the user's phone go through pedra_create_upload_link: they upload on a no-login page (up to 100 files, 24 h), then pedra_list_property_images returns their URLs for editing, pedra_create_video or pedra_create_virtual_tour.

Uploads are limited per day: 30 images on the free plan, 500 on paid plans (reset at midnight UTC), counting every image stored from outside the app (pedra_add_images_to_property, pedra_add_local_panoramas, URL scenes in tours, upload-link files); over it the call fails with HTTP 429, upload_limit and nothing is stored. Upload links: 5 a day free, 50 paid (upload_link_limit).

Example prompts once connected:

"Use Pedra to virtually stage https://example.com/empty-living-room.jpg as a minimalist living room."

"Virtually stage /Users/me/Desktop/empty-living-room.jpg as a minimalist living room."

"How many Pedra credits do I have left?"

Virtual tours

POST your 360° photos, get back a hosted, linked, shareable virtual tour: AI names the rooms and places the door-to-door navigation points. The tour tools have the same names, descriptions and input schemas as Pedra's hosted MCP server (https://app.pedra.ai/mcp), so an agent behaves the same on either.

A typical flow:

1. Get the 360° photos into a property.
   • Files on this computer → pedra_add_local_panoramas with the paths in walking order. This tool exists only in this local server: it reads the files off disk and sends them base64-encoded, 10 per call under the API's 50 MB request limit. Files over ~33 MB can't go this way.
   • Photos on someone's phone (or big files) → pedra_create_upload_link with type: "360", and hand over the uploadUrl (no login, valid 24 h).
   • Photos already online → skip this step and pass the URLs as scenes.
2. Build it: pedra_create_virtual_tour with the propertyId (uses every 360° photo in upload order) or with scenes. It returns a tourId straight away.
3. Wait: poll pedra_get_virtual_tour until status is "ready" (about 10 s per room) — then share tourUrl or embed embedCode. A "failed" build says why and costs nothing.
4. Adjust with pedra_update_virtual_tour (free, instant) or append rooms with pedra_add_virtual_tour_scenes.

Linking costs max(3, ceil(rooms/3)) credits for "sequential" (default; each room to the next), 5–160 for "smart" (AI works out which rooms connect), and nothing for "none". Deleting a tour is API-only (not exposed as a tool).

"Make a Pedra virtual tour of Calle Mayor 12 from /Users/me/360/entrance.jpg, /Users/me/360/living.jpg and /Users/me/360/kitchen.jpg, in that order, and give me the link."

How it works

This server is a thin wrapper over @pedra-ai/sdk, which encodes the API's contract details:

• Synchronous by design (except virtual tours). Every photo and video endpoint blocks and returns the final URL(s) in the response body. Even pedra_create_video polls server-side and returns the finished videoUrl inline (it can take up to ~10 minutes; the API keeps the connection alive with a heartbeat).
• Errors are tool errors. The API's 4xx responses (insufficient credits, bad image, …) come back as MCP tool errors with a readable message, not crashes.

Privacy Policy

This server sends the image/video URLs and parameters you pass to the Pedra API to perform the requested edit, authenticated with your PEDRA_API_KEY. Without a key, pedra_request_access sends the email address you give it to Pedra so Pedra can email you a confirmation link; a key obtained that way is held in memory for the session only. It stores no data itself. Data collection, usage, storage, retention, third-party sharing, and contact information are covered by Pedra's privacy policy: https://pedra.ai/privacy.

License

MIT © Pedra