Skip to content

MCP_SETUP.md

Latest commit

 

History

History
111 lines (85 loc) · 2.65 KB

MCP_SETUP.md

File metadata and controls

111 lines (85 loc) · 2.65 KB
Raw

MCP server setup

Quale exposes its 3 harness-proven commands as MCP tools — typed functions any MCP-compatible agent can call directly, no shell overhead.

When to use MCP vs shell vs skill

Method Best for Tradeoffs
MCP server Agents with MCP support Typed function calls, no shell parsing, lower token overhead
Shell commands Agents without MCP, or when you want explicit control Works everywhere, but requires parsing JSON output
Skill file OpenCode users who want automatic invocation Auto-loaded, but requires agent to remember to call

All three methods call the same underlying engine. The 75% accuracy and 0.0 extra edits metrics apply to all three.

How it works

Run quale --mcp to start the MCP server. It reads JSON-RPC requests from stdin and writes responses to stdout (stdio transport — zero network, zero dependencies).

Discovered via tools/list on startup. Agent sees 3 tools:

Tool Input Returns
edit_context {file: str, task?: str, path?: str} Risk, verification candidates, scope guard
verify_packet {file: str, diff?: str, path?: str} Verification + entangled candidates
orient {path?: str} Module map, landmarks, languages

OpenCode

Add to ~/.config/opencode/opencode.json:

{
  "mcp": {
    "quale": {
      "type": "local",
      "command": ["quale", "--mcp"],
      "description": "Structural codebase analysis. 75% test-file accuracy."
    }
  }
}

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "quale": {
      "command": "quale",
      "args": ["--mcp"]
    }
  }
}

Claude Code

claude mcp add quale -- pip install quale && quale --mcp

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "quale": {
      "type": "local",
      "command": ["quale", "--mcp"]
    }
  }
}

VS Code (Cline / Continue)

Add to your MCP config:

{
  "mcpServers": {
    "quale": {
      "command": "quale",
      "args": ["--mcp"]
    }
  }
}

Measured effect

12-trial comparison (Mistral-7B, 4 conditions):

Condition Correct test file Method
Baseline (no skill, no MCP) 2/3 Agent guesses by naming convention
Skill only 3/3 quale ec via shell
MCP only 2/3 edit_context tool
MCP + skill 3/3 Typed tool preferred, skill as fallback

MCP + skill together is the strongest configuration. The agent prefers the typed edit_context tool over shell commands, but the skill ensures correct behavior even when MCP isn't available.