MCP server and CLI

The code name threshold stays on the package, the command and the MCP server. This page is generated from docs/MCP.md.

Cranley’s tools work from any MCP client (Claude Code, Claude Desktop, Cursor, a LangChain agent) and from a terminal. The server reads the same registry as the web app, so a tool added under server/threshold/tools/ shows up in both without further changes.

It runs on your machine with your data folder ($THRESHOLD_DATA_DIR, default <data-dir>) and keys ($THRESHOLD_DATA_DIR/.env). Results are for the person running it: lists stop at 25 rows, findings at 50, and no tool returns panorama or floor plan images.

What it exposes

Kind	Name	Writes
Tools	every built tool in the registry (77 read-only today, e.g. garage_fit, door_widths, fire_preplan), inputs taken from its manifest	no
Tools	search_buildings, building_facts, workspace_read, list_findings	no
Tools, with --write only	agent_turn, workspace_create, workspace_ops, room_labeller	yes, marked WRITES and readOnlyHint: false
Resources	threshold://scan/{sid}/summary, threshold://workspace/{id}, threshold://measures
Prompts	wheelchair_check, fire_preplan, garage_check

A tool returns the ToolResult (Contract 3) as structured content, plus a few lines of text built from the result’s own fields: the verdict, the metrics with their confidence, failing checks, findings and assumptions.

Claude Code

claude mcp add threshold -- uv run --directory <repo>/server python -m threshold.mcp
# with agent turns and workspace writes:
claude mcp add threshold-write -- uv run --directory <repo>/server python -m threshold.mcp --write

Then ask: “Will an F-150 fit in the garage of scan ?” Claude calls garage_fit and answers from its result.

When the API is running, the same read-only server is at http://127.0.0.1:<port>/mcp over streamable HTTP:

claude mcp add --transport http threshold-http http://127.0.0.1:8787/mcp

The HTTP endpoint answers only requests addressed to localhost or 127.0.0.1 (MCP’s DNS-rebinding guard), so it is not reachable through the tailnet or a funnel. From another machine, run the stdio server over ssh as below.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows) and restart Claude Desktop.

On the machine that has the repo and the data folder:

{
  "mcpServers": {
    "threshold": {
      "command": "uv",
      "args": ["run", "--directory", "<repo>/server", "python", "-m", "threshold.mcp"]
    }
  }
}

From a laptop, with the server on another machine (<server-host>) reachable by ssh:

{
  "mcpServers": {
    "threshold": {
      "command": "ssh",
      "args": ["<server-host>", "cd <repo>/server && uv run python -m threshold.mcp"]
    }
  }
}

Cursor

.cursor/mcp.json in a project, or ~/.cursor/mcp.json for every project:

{
  "mcpServers": {
    "threshold": {
      "command": "uv",
      "args": ["run", "--directory", "<repo>/server", "python", "-m", "threshold.mcp"]
    }
  }
}

A LangChain deep agent

server/examples/mcp_deepagent.py loads the tools with langchain-mcp-adapters and hands them to create_deep_agent:

cd server && uv run python examples/mcp_deepagent.py "will an F-150 fit in the garage of scan <scan-id>"

The threshold command

uv sync in server/ installs it. Run it as uv run threshold ..., or activate server/.venv. uv run python -m threshold.cli does the same.

threshold tools                                         # every tool, [writes] marks the ones that write
threshold run garage_fit --scan <scan-id> -i vehicle=ford_f150
threshold run door_widths --scan <scan-id> -i min_width_m=0.81
threshold search greenwood --city Houston
threshold search --where 'garage_width>=6' --kind house --limit 5
threshold building <scan-id>                          # index row, plan summary, listing claims
threshold ask "will an F-150 fit in this garage" --scan <scan-id>   # agent turn; progress on stderr

--json prints the full result instead of text. -i key=value reads the value as JSON when it parses (numbers, true, lists) and as a string otherwise. threshold ask writes a workspace, the same kind the web app opens.

Testing

cd server && uv run pytest tests/mcp
npx @modelcontextprotocol/inspector --cli uv run --directory server python -m threshold.mcp --method tools/list