Connecting MCP servers

Azure DevOps and Jira, for both Claude Code and Copilot — different file, different key.

Requirements are read through the connected MCP server for your provider — never raw REST. init already seeded both config files with placeholders (unless you passed --no-mcp); this page is about filling them in.

Claude Code and Copilot use different files and different top-level keys for MCP config. .mcp.json is not read by Copilot; .vscode/mcp.json is not read by Claude Code. Both need to exist and stay in sync if you use both hosts.

Host	File	Top-level key
Claude Code	.mcp.json (repo root)	mcpServers
GitHub Copilot (VS Code)	.vscode/mcp.json	servers

Azure DevOps

{
  "mcpServers": {
    "azure-devops": {
      "command": "npx",
      "args": ["-y", "@azure-devops/mcp", "<your-ado-org>"]
    }
  }
}

Replace <your-ado-org> with your organization name (the part before .visualstudio.com or after dev.azure.com/). Authentication is via az login (Azure CLI) — make sure you’ve logged in before your first /shipflow run.

Jira / Atlassian

{
  "mcpServers": {
    "atlassian": {
      "type": "sse",
      "url": "https://mcp.atlassian.com/v1/sse"
    }
  }
}

This is Atlassian’s remote MCP — it handles OAuth in-browser on first use, no local install needed.

Copilot: same servers, different wrapper

.vscode/mcp.json uses the exact same server definitions, just under servers instead of mcpServers:

{
  "servers": {
    "azure-devops": {
      "command": "npx",
      "args": ["-y", "@azure-devops/mcp", "<your-ado-org>"]
    },
    "atlassian": {
      "type": "sse",
      "url": "https://mcp.atlassian.com/v1/sse"
    }
  }
}

Only enable what you need

Set only the server(s) that match your shipflow.config.md → requirements.provider. Enabling both is fine if your team uses both providers across different repos, but each /shipflow run only calls the one matching that repo’s configured provider.

Verifying the connection

If the required MCP server isn’t connected, feature-context will say so plainly rather than falling back to raw REST or fabricating the work item’s contents — that’s by design, not a bug to work around. If a run stops with an MCP-not-connected message, check your host’s MCP status (in Claude Code, /mcp; in Copilot, the MCP servers panel) before re-running.