Set up MCP

MCP (Model Context Protocol) allows compatible coding agents to use Traceback as a tool. After setup, an agent can discover available projects, read test definitions, create ephemeral QA packs, run existing tests, and verify whether a fix actually changed the product behavior.

How it works

When an agent connects to Traceback via MCP, it gains access to a set of tools:

  • list_projects — discover available workspaces
  • list_tests — find existing test definitions
  • run_test — trigger a run and wait for results
  • create_test — author a new test from a prompt
  • get_issue — read issue details and proposed fix

The agent calls these tools mid-session — for example, after making a code change, it can call run_test to verify the fix actually worked.

Prerequisites

  • A Traceback account with at least one workspace
  • A compatible MCP client (Claude Code, Cursor, Windsurf, Codex CLI, or similar)

Step 1: Generate an access token

In Traceback, go to Settings → Access tokens and create a new token. Give it a descriptive name like cursor-mcp.

Warning

Store your token securely. It grants full access to your Traceback workspace. Don’t commit it to source control.

Step 2: Configure your MCP client

Copy the Traceback MCP configuration and paste it into your client’s MCP config file.

Claude Code (~/.claude/claude_desktop_config.json):

{
  "mcpServers": {
    "traceback": {
      "command": "npx",
      "args": ["-y", "@traceback/mcp-server"],
      "env": {
        "TRACEBACK_TOKEN": "your-token-here"
      }
    }
  }
}

Cursor (.cursor/mcp.json in your project root):

{
  "traceback": {
    "command": "npx",
    "args": ["-y", "@traceback/mcp-server"],
    "env": {
      "TRACEBACK_TOKEN": "your-token-here"
    }
  }
}

Step 3: Test the connection

Restart your client and ask the agent to list your Traceback projects:

What Traceback projects do I have access to?

The agent should respond with your workspace names. If it doesn’t, check that your token is correct and that the MCP server started without errors.

Tip

The first successful connection should be a capability test — ask the agent to list projects and run one existing test before handing it any autonomy.

Next steps

PreviousOverviewNextMCP permissions