dataforge-mcp

dataforge-mcp exposes DataForge's shipped CSV profiling, detection, repair, verification, and transaction-revert paths as Model Context Protocol tools.

cd dataforge-mcp
python -m pip install -e ".[dev]"
dataforge-mcp serve --allowed-root /path/to/csv/workspace

For local development from this repository:

cd dataforge-mcp
python -m pip install -e ".[dev]"
dataforge-mcp serve --allowed-root ..

The default transport is stdio, which is what local desktop MCP clients expect. For local Streamable HTTP experiments:

dataforge-mcp serve --transport streamable-http --host 127.0.0.1 --port 8000

dry_run is the safe default. To allow file mutation through MCP, start the server with an explicit allowed root and --enable-apply:

dataforge-mcp serve --allowed-root /path/to/csv/workspace --enable-apply

Tools

• dataforge_profile(path: str) - summarize CSV shape plus detected issues.
• dataforge_detect_errors(path: str) - return detected issues only.
• dataforge_verify_fix(fix_spec: dict) - run one candidate fix through stale value checks, safety, and verification.
• dataforge_apply_repairs(path: str, mode: "dry_run" | "apply") - propose verified repairs and optionally write a reversible transaction.
• dataforge_revert(txn_id: str) - restore a transaction's original bytes.

Client Configuration

Use the same server command for Claude Desktop, Cursor, Windsurf, or any local MCP client that supports stdio servers:

{
  "mcpServers": {
    "dataforge": {
      "command": "dataforge-mcp",
      "args": ["serve", "--allowed-root", "/path/to/csv/workspace"]
    }
  }
}

If your client cannot resolve the console script, replace command with the absolute path returned by your shell:

which dataforge-mcp

On Windows PowerShell:

Get-Command dataforge-mcp

Before describing a build as agent-ready, run an MCP Inspector smoke check against a fixture directory and confirm the profile, detect, verify, dry-run apply, and disabled-apply paths:

npx @modelcontextprotocol/inspector dataforge-mcp serve --allowed-root /path/to/csv/workspace

Safety Model

apply mode uses DataForge's detector -> repairer -> SafetyFilter -> SMTVerifier -> transaction-log path. The tool writes the transaction journal and source snapshot before mutating the CSV, and dataforge_revert restores the snapshot only when the current file still matches the recorded post-state hash.

The MCP server does not enable live LLM repair fallback by default. It does not send CSV contents to any external model provider. It also rejects CSV and schema paths outside the configured allowed roots, and apply mode is disabled unless the server is started with --enable-apply or DATAFORGE_MCP_ENABLE_APPLY=1.

Release

The package is intended to release independently from the nested dataforge-mcp/ source directory as the dataforge_07_mcp distribution, but it is not published yet. After PyPI Trusted Publishing is configured, the workflow will build on tags matching:

dataforge-mcp-v*

The package depends on dataforge_07 and the official Python mcp SDK; it does not vendor DataForge or add MCP dependencies to the core package.