Claude, ChatGPT and Copilot¶

Sight has an MCP server, so AI assistants can work with your checks for you. MCP (the Model Context Protocol) is the standard way assistants connect to tools. Ask in your own words, for example:

• "What does our site-completion template check?"
• "Make the front door check need two doors, and show me what would change."
• "Why did job TASK-123 fail?"
• "Which jobs failed at site 42 this week?"

The assistant uses Sight's tools with your company's API key. It can do exactly what that key is allowed, and nothing more.

Sign in with Sight: Claude and ChatGPT¶

In the Claude and ChatGPT apps you sign in instead of pasting a key.

1. Add a custom connector (Claude) or app (ChatGPT) with the server URL above. Your assistant opens Sight's sign-in: sign in as you do on sight.raku.so.
2. The first time, open AI assistants in the sidebar on sight.raku.so (sight.raku.so/assistant) and choose the company and model your assistants work with, and what they may do: send photos and read results, write checks, or both. Your assistant gives you this link if you ask it something first.
3. Ask away.

You need to be an owner or admin of the company. Choosing makes a company key named "AI assistant: you@…", listed at Company → API keys. Nobody ever sees the key itself. Revoking it, or disconnecting on the same page, ends your assistants' access at once. So does leaving the company or no longer being an owner or admin there: Sight checks on every request.

On a free account there is no company to choose, so you connect your Try Sight project instead. Your assistants then work through your test key, with Sight's sample hard hat check and within the test key's daily limit: they can send photos and read the results, but not write checks. Making a new test key on your dashboard keeps them connected, and disconnecting leaves your test key working in your own code.

What your assistant thinks and writes runs on your Claude or ChatGPT plan. The photos it checks count toward your Sight plan, as they would from your own software.

With a key: Claude Code, Copilot, Cursor and your own agents¶

Make the key at Company → API keys, allowed:

• Photo checks, to read templates and results and to check photos;
• Check templates, to write templates, preview them and make them live.

Give it only what the assistant needs.

Claude Code¶

claude mcp add --transport http sight https://sightapi.raku.so/mcp \
  --header "Authorization: Bearer $SIGHT_API_KEY"

GitHub Copilot in VS Code¶

Add .vscode/mcp.json. VS Code asks for the key once and keeps it out of the file:

{
  "inputs": [
    { "type": "promptString", "id": "sight-key", "description": "Sight API key", "password": true }
  ],
  "servers": {
    "sight": {
      "type": "http",
      "url": "https://sightapi.raku.so/mcp",
      "headers": { "Authorization": "Bearer ${input:sight-key}" }
    }
  }
}

Microsoft Copilot Studio¶

In your agent, add a tool: Model Context Protocol, with the server URL above. For authentication choose API key, sent as a header named X-Api-Key, with your Sight key as its value.

Other assistants and your own agents¶

Any MCP client that can send a header works the same way: Cursor, Codex, and the OpenAI and Anthropic APIs' MCP tools.

The tools¶

Sending photos from your phone or computer¶

In Claude, Gemini and Copilot, a photo you attach to the chat does not reach Sight. Ask your assistant to check photos and it gives you a photo link instead. Open it on your phone (it offers the camera and your photo library) or your computer, take or choose the photos, and each one is checked as it arrives, with its verdict on the page. Then tell your assistant you are done, and it reads the job's results.

A photo link works for one job and one check, for an hour, and for the number of photos your assistant asked for (10 unless it says otherwise, at most 20). It acts as your assistant's own key, so revoking the key stops its links too, and each photo counts toward your plan as any photo does. Anyone with the link can send photos to that job until it ends, so keep it to yourself.

In ChatGPT you can also attach a photo to the chat: ChatGPT passes it to check_photo directly.

What keeps it safe¶

• Only what the key may do. A key allowed only Photo checks cannot change a template; the assistant is told why, in the same words as the API. Signed in, your assistant may do only what you allowed when you connected it.
• Nothing goes live on its own. make_live is marked as a change that needs your approval, so your assistant asks you first. The assistant is also told to show you the preview and wait for your yes. A draft replacing a live version must be previewed first.
• Every change is a new version. Make an older one live again to roll back. Jobs already started keep the version they started with.
• On the record. Every change appears in your company's activity, marked as made through MCP, with the key's name.
• The same checks as everywhere. A template an assistant writes goes through the same gate as the console and the API, plan limits included.
• Photo links are short-lived. One job, one check, an hour, a set number of photos, as the assistant's key. Sight keeps only a fingerprint of each link, never the link itself.