Skip to content
Scope

Using MCP servers, skills & extensions

Three optional capabilities extend what an agent can do during a run:

  • MCP servers — Model Context Protocol servers that expose tools the agent can call.
  • Skills — packaged Copilot agent skills, pinned to specific commits.
  • Extensions — VS Code extensions installed for the duration of a run. VS Code Copilot only.

All three are configured on a profile, not per-run.

When to use what

Section titled “When to use what”
  • MCP servers when you want the agent to access an external system during the run — a knowledge base, a custom tool, a service.
  • Skills when you want to give a Copilot-driven agent specific expertise (e.g. an “Azure” skill for Azure-related tasks).
  • Extensions when you want to measure how a VS Code extension affects coding outcomes (e.g. linters, language servers, AI helpers). Requires the VS Code Copilot coding agent.

MCP servers

Section titled “MCP servers”

MCP servers are referenced by slug. The slug is the name Scope uses to identify a registered MCP server.

Add MCP servers to a profile:

{
  "mcpServers": ["filesystem", "github"]
}

The agent gains access to whatever tools that MCP server exposes. The behavior of those tools is part of what you’re benchmarking — the same MCP server may be used effectively by one agent and ignored by another.

To discover the slugs registered in your Scope deployment, see the MCP servers page in the Portal.

Skills

Section titled “Skills”

Skills are Copilot agent skills, identified by a path inside a Git repository:

github/<owner>/<repo>/<skill-path>

When you save a profile with a skill referenced by path, Scope resolves it to a commit hash and stores the pinned form in the profile version:

github/vercel-labs/agent-skills/azure@deadbeef

If you provide a pre-pinned reference, it passes through unchanged.

This pinning is the reason profile versions are immutable: the same profile version run a year from now will use the exact same skill code.

{
  "skillRevisions": [
    "github/vercel-labs/agent-skills/azure",
    "github/vercel-labs/agent-skills/database@abc1234"
  ]
}

Skills only apply to Copilot-based agents.

VS Code extensions

Section titled “VS Code extensions”

Extensions are identified by their VS Code Marketplace ID, optionally pinned to a version:

ms-python.python
ms-python.python@2024.8.1

Unpinned IDs are resolved to the latest stable version at profile version-creation time and stored as <id>@<version>.

{
  "extensions": [
    "ms-python.python",
    "esbenp.prettier-vscode@10.4.0"
  ]
}

Extensions only apply to the VS Code Copilot coding agent. The GitHub Copilot CLI and Claude Code CLI agents reject extensions with HTTP 400 — see Defining profiles → worker-specific constraints.

Combining them

Section titled “Combining them”

A single profile can use any mix of MCP servers, skills, and (where supported) extensions:

{
  "name": "VS Code + Python + Azure",
  "model": "gpt-4o",
  "mcpServers": ["filesystem", "github"],
  "skillRevisions": ["github/vercel-labs/agent-skills/azure"],
  "extensions": ["ms-python.python"]
}

Tips

Section titled “Tips”
  • Vary one thing at a time. When measuring the impact of a skill or extension, keep everything else identical between profiles. Otherwise you can’t attribute differences cleanly.
  • Pin everything for benchmarks you’ll re-run. Skill paths without commits and extension IDs without versions resolve to “latest” — fine for exploration, dangerous for long-lived comparisons.
  • Don’t over-stack tools. Adding three MCP servers and five skills to chase a higher pass rate makes the resulting profile harder to reason about. Smaller profiles compare better.

See also

Section titled “See also”