citadel-predict-mcp

Model Context Protocol (MCP) server for Citadel Predict pre-execution AI agent cost estimation.

citadel-predict-mcp connects your hosted Citadel Predict API directly into Claude Desktop and Claude Code via a local standard I/O (stdio) MCP server.

With this server configured, Claude can natively estimate token usage ranges ($low, expected, high$) and flag out-of-distribution risks for autonomous workflows before running them—without requiring manual CLI execution.

Features

• ⚡ Native Claude Tool Calling: Claude automatically decides when to call estimate_agent_cost when planning or dispatching tasks.
• 🔒 Zero Network Exposure: Runs strictly as a local stdio subprocess spawned by Claude Desktop / Claude Code.
• 🎯 Pre-Execution Guardrails: Predicts token consumption bounds before multi-step tools or reasoning loops execute.
• 🛡️ User-Friendly Error Handling: Catches authentication, rate-limiting, and validation issues, presenting clear, actionable suggestions to Claude rather than raw stack traces.

Installation

Install the package via pip:

pip install citadel-predict-mcp

(For local development from the repository root: pip install -e packages/citadel-predict-mcp)

Verifying installation:

citadel-predict-mcp --help

Quickstart: One-Command Auto-Configuration (Recommended)

To automatically configure both Claude Desktop and Claude Code on Windows, macOS, or Linux with zero manual JSON editing:

# Run from repository root with your active Python environment:
python scripts/configure_mcp.py

# Or pass parameters non-interactively:
python scripts/configure_mcp.py --api-key cp_live_your_key_here --api-url http://localhost:8000

The script automatically detects your active virtual environment, locates the platform-specific executable (citadel-predict-mcp.exe on Windows or citadel-predict-mcp on macOS/Linux), and cleanly merges the server definition into both claude_desktop_config.json and .claude/settings.json while preserving all existing configurations.

Manual Configuration (Alternative)

If you prefer to configure manually:

1. Claude Desktop (claude_desktop_config.json)

Add citadel-predict under mcpServers:

{
  "mcpServers": {
    "citadel-predict": {
      "command": "citadel-predict-mcp",
      "env": {
        "CITADEL_API_KEY": "cp_live_your_api_key_here",
        "CITADEL_API_URL": "http://localhost:8000"
      }
    }
  }
}

2. Claude Code (.claude/settings.json)

claude mcp add citadel-predict -- citadel-predict-mcp

Available Tools

estimate_agent_cost

Description:

Estimate token cost and usage range for an AI agent task BEFORE running it. Use this when the user is about to execute, dispatch, or run a multi-step agent task and cost/budget matters.

Input Parameters:

• task_text (string, required): The natural language description of the agent task (1 to 4000 characters).
• tools (array of strings, optional): List of tool names available to the agent (e.g. ["web_search", "draft_document"]).
• num_tools (integer, optional): Tool count if specific tool names are not listed.
• model_id (string, optional, default: "claude-sonnet"): Model calibration profile to evaluate against.

Sample Return Payload:

{
  "success": true,
  "model_id": "claude-sonnet",
  "expected_tokens": 3200,
  "low_tokens": 1500,
  "high_tokens": 5800,
  "out_of_distribution": false,
  "ood_reasons": [],
  "confidence": "normal",
  "driving_factors": ["task_length", "tools_count"],
  "summary": "Expected: 3,200 tokens (Range: 1,500 – 5,800)"
}

Manual Test Checklist for Testers

Follow this 5-minute checklist to verify your MCP setup:

1. Installation:
   • Run citadel-predict-mcp --help in your terminal to verify the command is accessible on your PATH.
2. Configuration:
   • Add the JSON entry to claude_desktop_config.json with your active Citadel API key.
3. Restart:
   • Fully quit and reopen Claude Desktop.
4. Invocation Test:
   • Send the following prompt in a new Claude Desktop chat:

     "I am planning to have an agent research competitor pricing across 5 company sites and compile a markdown report. Estimate the token cost and budget range before we start."

5. Verification:
   • Confirm Claude invokes the estimate_agent_cost tool (indicated by a tool-call widget in the conversation).
   • Confirm Claude receives the token ranges (expected_tokens, low_tokens, high_tokens) and presents a natural language summary with the budget estimate to you.

License

MIT