# cli-vs MCP server

> A TypeScript harness that compares Claude Code performance across CLI skills, MCP servers, and baseline tool surfaces.

- Repository: https://github.com/chief-builder/cli-vs-mcp
- Page: https://hostedmcp.org/servers/chief-builder/cli-vs-mcp
- Transport: —
- Stars: 2
- License: MIT

## Environment variables

| Name | Required | Secret | Description |
|---|---|---|---|
| `GITHUB_AGENT_TOKEN` | no | no |  |
| `GITHUB_AGENT_TOKEN_RW` | no | no |  |
| `GITHUB_CONTROLLER_TOKEN` | no | no |  |
| `GITHUB_SANDBOX_OWNER` | no | no |  |

## FAQ

### What does this harness measure?

It measures token cost, time cost, success scores, and whether tool calls stayed within the intended surface.

### How are trials isolated to prevent data leakage?

Trials use paired seeds for state, memory-based fixture servers, and strict file-tool confinement rules.

### Can I use this to compare CLI skills against MCP servers?

Yes, it compares a baseline with no tools, a skill arm using a CLI, and an MCP arm using an MCP server.
