@dzhechkov/mcp-server-tools

An MCP server that exposes the DZ cross-platform harness as Model Context Protocol tools, so any MCP client can drive the harness without the CLI.

Tools

| Tool | Purpose | |---|---| | skill_list | List the skills in the configured skills directory | | skill_get | Read one skill's SKILL.md and parsed frontmatter | | skill_compile | Compile a skill for a target platform (claude / codex / opencode / hermes) | | harness_verify | Compile + structurally verify a skill for a platform |

claim_check_text — vet prose before you write it

Runs claim-check over raw text (not a file), so an agent can check a paragraph it is about to write. Mirrors dz claim-check, but voluntary and text-only.

• Input: { text: string, failOn?: 'high' | 'medium' | 'none' }.
• Fail-closed: empty, whitespace-only, or invisible-character-only text is an error, not a pass — a claim-check called with nothing has vetted nothing.
• Findings are reported, never thrown. failOn sets a gated flag; it never rescues the empty error.
• A clean result means "no untagged quantitative claim found", not "your text is correct".

From an agent, in plain language:

"Before I write this paragraph, claim-check it: 'our model hits 99% accuracy'."

Configuration

The server reads skills from DZ_SKILLS_DIR (an environment variable), defaulting to .claude/skills relative to the working directory.

Usage — register with Claude Code

Add to your project's .mcp.json:

{
  "mcpServers": {
    "dz-harness": {
      "command": "dz-mcp-server-tools",
      "env": { "DZ_SKILLS_DIR": ".claude/skills" }
    }
  }
}

Transport is stdio. The server is also importable — createDzMcpServer() returns a configured McpServer for embedding or testing.

Status

0.1.0 — alpha, part of the extended-a-migration feature (Phase 5).