MCP Server

Connect external AI tools directly to Viberglass. Supported tools include Claude Desktop, Cursor, LM Studio, or your favourite local coding harness as well as any other MCP-compatible client. Your AI assistant can list projects, create tickets, trigger agent runs, review plans, and approve work without leaving the tool you already use.

What you can do

Once connected, your AI tool can:

• Browse projects and clankers: discover what's available before creating work
• Create tickets: file a bug or feature request in any project
• List and search tickets: filter by status, severity, or free-text search
• Trigger agent runs: kick off research, planning, or execution on any ticket
• Review plans: read the research or planning document and leave inline comments
• Approve or revoke: advance the workflow or send the agent back to revise

Think of it as giving your AI assistant full access to the Viberglass ticket lifecycle.

Prerequisites

You need a Viberglass API token. Create one in the platform under Secrets → API Tokens. Copy the token value. Note that you won't be able to see it again.

If you don't see the Secrets page, ask your admin for an API token.

Connecting your AI tool

MCP uses a standard connection format. The exact setup varies by client, but the information you need is always the same:

Replace your-viberglass-instance with the domain where Viberglass is running (for example, app.viberglass.io). Replace <your-api-token> with the token you created.

Claude Desktop

Open your Claude Desktop configuration file and add the Viberglass server:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "viberglass": {
      "url": "https://your-viberglass-instance/api/mcp",
      "headers": {
        "Authorization": "Bearer <your-api-token>"
      }
    }
  }
}

Restart Claude Desktop after saving the file.

Cursor

Open Settings → MCP and add a new server with the URL and authorization header above. Cursor discovers available tools automatically.

Other MCP clients

Any client that supports the Model Context Protocol with Streamable HTTP transport should work. Point it at the same URL and include the Authorization header with your API token.

Using the tools

Once connected, the tools appear in your AI assistant alongside its other capabilities. There's no special syntax, you can just ask in natural language.

Examples:

• "List all open tickets in the backend project"
• "Create a ticket titled 'Fix login redirect on mobile' in the web app project"
• "Run the research phase on ticket abc-123 using the Claude agent"
• "Show me the review state for ticket abc-123"
• "Approve the planning document for ticket abc-123"
• "Add a comment on line 42 of the planning document: 'This should use the async version'"

Your AI assistant translates these into the right tool calls. You never interact with the raw API directly.

Workflow tips

Start with discovery. Ask the assistant to list projects and clankers first, so it knows the IDs to reference when creating tickets or triggering runs.

Review before executing. After the agent finishes research or planning, read the document through your AI tool and leave comments if something needs to change. Approve only when you're satisfied.

One ticket at a time. Triggering multiple runs on the same ticket simultaneously can cause conflicts. Let each phase finish before starting the next.

Use search. Instead of remembering ticket IDs, ask the assistant to search by title or filter by status. It's faster than browsing.

Revoke to revise. If you approved a plan prematurely, revoke the approval. The agent can then pick up your comments and revise.

Limitations

• MCP access is read/write for tickets and reviews. Managing clankers, secrets, integrations, and claws still requires the web UI or Slack.
• There is no file upload through MCP. Attach screenshots and logs through the Chrome extension or the web UI.
• Each MCP request is independent, there's no persistent session. Your AI tool handles context on its side.