HCM-LLM MCP Server

A FastAPI-based Model Context Protocol (MCP) server for Highway Capacity Manual (HCM) analysis and transportation engineering calculations. So far, this server provides comprehensive two-lane highway analysis following HCM Chapter 15 methodology.

Features

• Semantic search over HCM documentation

• Complete HCM Chapter 15 (two-lane highway) and Chapter 12 (basic freeway) analysis

• Full HCM chapter coverage, chapters 10 through 28, behind three capability tools (hcm_analyze, hcm_describe, hcm_validate) over 33 methods, each taking the library's own example-case JSON and each validated against its published example problem

• Input validation gateway against HCM/AASHTO constraints (via transportations-validator)

• Full-corpus validation (300+ rules across HCM/AASHTO/MUTCD/HSM/ADA/...) with citations, terrain/context-gated rules, and clarification requests — runs in-process, no database

• Knowledge-graph reasoning: abductive design repair (Two-Lane & Basic Freeway), defeasible code reconciliation, inverse design, and forward/backward chaining — every repair candidate re-executed through the verified library

• YAML-based function registry for easy extensibility

• Function calling interface with 15+ transportation analysis functions

• MCP server compatibility for integration with AI assistants (supporting Claude)

• RESTful API endpoints for direct access

• Dynamic endpoint generation based on registry

• Comprehensive test suite and validation tools

Related MCP server: MCP WebAnalyzer

Connect to Remote MCP Server

This server can be used as a backend for AI code agents like Claude Desktop, allowing them to perform complex transportation analyses and access HCM documentation dynamically.

To enable this functionality, add the server to your AI assistant's configuration as an MCP server.

For Claude Desktop Users

From user setting, you can find Connectors tab and click Add custom connector.

Then add https://api.hcm-calculator.com/mcp to you Claude configuration.

For GitHub Copilot on VSCode Users

You can also use this server with GitHub Copilot by configuring it as a custom MCP server.

To do this, type Ctrl+p and select MCP: Open User Configuration and modify the following to your mcp.json:

{
	"servers": {
		"hcm-mcp": {
			"url": "https://api.hcm-calculator.com/mcp"
		}
	}
}

Connect to Local MCP Server

You can also run this server locally for development or testing purposes.

uv venv

# Windows
.venv\Scripts\activate
# Linux
source .venv/bin/activate

uv pip install .

Then running the server.

# Setup the database.
python hcm_mcp_server/scripts/import_hcm_docs.py

# Start the server.
python mcp_server_fastapi.py

For Claude Desktop Users

Open Claude Desktop and add the server as a custom MCP server with the URL http://localhost:8000/mcp.

Add to your Claude Desktop configuration (claude_desktop_config.json):

Note: Seems like this json settings are not working these days (https://github.com/anthropics/claude-code/issues/4188), and it did not work in my desktop environment, either.

{
  "mcpServers": {
    "hcm-mcp-local": {
      "url": "http://127.0.0.1:8000/mcp"
    }
  }
}

For GitHub Copilot on VSCode Users

Same thing as above, you can use this server with GitHub Copilot by configuring it as a custom MCP server.

To do this, type Ctrl+p and select MCP: Open User Configuration and modify the following to your mcp.json:

{
	"servers": {
		"hcm-mcp-local": {
			"url": "http://127.0.0.1:8000/mcp"
		}
	}
}

Then you can use the function calling interface directly in your code editor.

Project Structure

hcm-mcp-server/
├── mcp_server_fastapi.py        # Main FastAPI application
├── functions_registry.yaml      # Function registry configuration
├── hcm_mcp_server/
│   ├── example_prompts/                  
│   │   ├── *.txt                # Example prompts for function calling
│   │   └── *.json               # Example json files for web validation
│   ├── core/                    # Core application modules
│   │   ├── dependencies.py      # Dependency injection and utilities
│   │   ├── registry.py          # Function registry implementation
│   │   ├── models.py            # Pydantic data models
│   │   └── endpoints.py         # Dynamic endpoint creation
│   ├── functions/                  
│   │   ├── chapter15.py         # Chapter 15: Two-Lane Highways
│   │   └── research.py          # Research and documentation
│   └── scripts/                    
│       ├── import_hcm_docs.py   # Import HCM documentation and setup ChromaDB
│       └── validate_registry.py # Registry validation
├── data/                           
│   └──  hcm_files/               # HCM documentation files
└── chroma_db/                    # ChromaDB storage

Configuration

Environment Variables

Create a .env file based on .env.example. Copy and paste the following content, or cp .env.example .env:

CHROMA_DB_PATH=./chroma_db
HOST=127.0.0.1
PORT=8000
ALLOWED_ORIGINS=http://localhost:3000,http://localhost:3001
LOG_LEVEL=INFO
DB_MODE=local
PUBLIC_SUPABASE_URL=https://
PUBLIC_SUPABASE_API=your-anon-key / service-role-key

Function Registry

Functions are defined in functions_registry.yaml:

functions:
  chapter15:
    identify_vertical_class:
      module: "functions.chapter15"
      function: "identify_vertical_class_function"
      description: "Identify vertical alignment class range"
      category: "transportation"
      chapter: 15
      step: 1
      parameters:
        type: "object"
        properties:
          segment_index:
            type: "integer"
          highway_data:
            type: "object"
        required: ["segment_index", "highway_data"]

Ablation arms (restricted MCP surfaces)

For the Table 5 / Figure 7 2x2 ablation, the same app can be launched exposing only a subset of tools, so a model can be evaluated under each condition in isolation:

python mcp_server_fastapi.py     # ct  : full system (all tools), port 8000
python mcp_server_kg_only.py     # kg  : 7 reasoning/validation tools only, port 8001 (no Chroma needed)
python mcp_server_rag_only.py    # rag : query_hcm only, port 8002

Both launchers are thin wrappers that set two env vars before importing the app:

• HCM_MCP_INCLUDE_OPS — comma-separated operation ids the MCP surface exposes (unset = all). Filtering uses FastApiMCP(include_operations=...).

• HCM_ENABLE_RAG — set to false to skip loading the embedding model + vector store (the kg-only arm needs neither).

Point each VS Code / Claude Desktop MCP client at the port for the arm under test (e.g. http://localhost:8001 for kg-only) so the model sees only that arm's tools. The base arm is simply no MCP server attached.

API Usage

Complete Highway Analysis

curl -X POST "http://localhost:8000/analysis/chapter15/complete" \
  -H "Content-Type: application/json" \
  -d '{
    "segments": [{
      "passing_type": 0,
      "length": 2.0,
      "grade": 2.0,
      "spl": 50.0,
      "volume": 760.0,
      "volume_op": 1500.0,
      "phf": 0.95,
      "phv": 5.0
    }],
    "lane_width": 12.0,
    "shoulder_width": 6.0,
    "apd": 5.0
  }'

Function Calling Interface

curl -X POST "http://localhost:8000/tools/call" \
  -H "Content-Type: application/json" \
  -d '{
    "function": {
      "name": "chapter15_determine_free_flow_speed",
      "arguments": {
        "segment_index": 0,
        "highway_data": {
          "segments": [{"passing_type": 0, "length": 2.0, "grade": 2.0, "spl": 50.0}],
          "lane_width": 12.0,
          "shoulder_width": 6.0
        }
      }
    }
  }'

List Available Functions

# List all functions
curl -X POST "http://localhost:8000/tools/list"

# Filter by category
curl -X POST "http://localhost:8000/tools/list" \
  -H "Content-Type: application/json" \
  -d '{"category": "transportation"}'

# Filter by chapter
curl -X POST "http://localhost:8000/tools/list" \
  -H "Content-Type: application/json" \
  -d '{"chapter": 15}'

Query HCM Documentation

curl -X POST "http://localhost:8000/tools/query-hcm" \
  -H "Content-Type: application/json" \
  -d '{
    "question": "What factors affect free flow speed in two-lane highways?",
    "top_k": 5
  }'

Available Functions

Chapter 15 Functions

• chapter15_identify_vertical_class - Step 1: Identify vertical alignment class range

• chapter15_determine_demand_flow - Step 2: Calculate demand flow rates and capacity

• chapter15_determine_vertical_alignment - Step 3: Determine vertical alignment classification

• chapter15_determine_free_flow_speed - Step 4: Calculate free flow speed

• chapter15_estimate_average_speed - Step 5: Estimate average travel speed

• chapter15_estimate_percent_followers - Step 6: Estimate percentage of following vehicles

• chapter15_determine_follower_density_pl - Step 8a: Follower density for passing lanes

• chapter15_determine_follower_density_pc_pz - Step 8b: Follower density for PC/PZ segments

• chapter15_determine_segment_los - Step 9: Calculate segment Level of Service

• chapter15_determine_facility_los - Step 10: Calculate facility Level of Service

• chapter15_complete_analysis - Complete HCM Chapter 15 procedure

Chapter 12 Functions (Basic Freeway Segments)

A different equation family than Chapter 15 — the lane width -> FFS -> capacity/speed -> density -> LOS chain. Requires transportations-library>=0.1.12.

• chapter12_determine_free_flow_speed - Step 2: Estimate and adjust free-flow speed

• chapter12_estimate_capacity - Step 3: Base and adjusted capacity (pc/h/ln)

• chapter12_estimate_demand_volume - Step 4: Per-lane flow rate v_p

• chapter12_calculate_speed - Step 5a: Space mean speed via the speed-flow curve

• chapter12_estimate_density - Step 5b: Density D = v_p / S

• chapter12_determine_segment_los - Step 6: Segment Level of Service

• chapter12_complete_analysis - Complete HCM Chapter 12 basic-freeway procedure

HCM Analysis Capabilities (full chapter coverage)

Every HCM method the compute library implements, chapters 10 through 28, behind three capability tools. The method is an argument, not a tool: thirty-three near-identical schemas would cost every caller context and blunt tool selection, and the ten published tools above are already capability-shaped.

• hcm_analyze — {method, config}. Runs one method. The tool description carries the method catalog, one compact line each, and method is an enum of the thirty-three ids.

• hcm_describe — {method?}. With a method: its input schema sketch, its result-field meanings and the example-problem fixture that validates it. Without one: the catalog, every method id with its chapter and a one-line summary. Call this first; it is how a caller learns a method's shape without reading the Rust bindings.

• hcm_validate — {method, config}. Parses and checks a config without running the analysis, returning the library's own validation errors or ok. Iterating on a config costs a parse rather than a full analysis. Twenty-four of the thirty-three methods have a real validation step behind the constructor (serde deserialisation, constructor range checks, and for Chapter 15 the Exhibit 15-8 parameter ranges via tl.validate_input); the other nine are single JSON entry points in the library where parsing and computation are one call, and those say so in the response rather than running the analysis and calling it a validation.

The input is always the compute library's own example-case (fixture) JSON, passed as config — not a second flattened schema invented for the MCP layer. An example case from transportations-library/tests/ExampleCases/hcm/ can be handed over unmodified. Requires transportations-library>=0.3.7.

Each method ships its worked example under hcm_mcp_server/data/examples/<method>.json, and tests/test_methods.py drives every one of them through hcm_analyze against the published values of that example problem, at the tolerances the compute library's own test suite asserts.

Each method also keeps a method-shaped REST route at /analysis/hcm/<method-with-hyphens> for direct API callers. Routes are not MCP tools, so these cost a caller's context nothing.

Domain refusals carry the library's own message. An un-digitised mixed-flow grade, an off-domain specific-upgrade PCE and a malformed config all come back as {"success": false, "error": "..."} in the library's words, because those messages say what the published HCM data covers.

These tools are not in the default MCP surface. mcp_server_fastapi.py is the ct ablation arm, and the ten tools it advertises by default are the surface the published experiment ran against (see tests/test_frozen_surface.py). Set HCM_MCP_FULL_COVERAGE=true to append the three capability tools to the MCP mount. Without it they are still reachable over REST and through /tools/call.

Validation Functions

• validation_validate_design_full - Validate a design against the full rule corpus (300+ rules: HCM, AASHTO, MUTCD, HSM, ADA, OpenDRIVE, ...) with citations, terrain/jurisdiction-gated rules, and clarification requests when an input is missing or its context is ambiguous. Runs in-process over the bundled seed corpus — no database. (The Chapter 15/12 tools use a lighter semantic-firewall gateway; this is the complete engine.) Requires transportations-validator>=0.2.0 + sqlalchemy.

Research Functions

• query_hcm - Query HCM documentation database

Reasoning Functions

The X-KG reasoning layer reasons over the knowledge graph and the verified executable substrate. Repair and inverse-design re-execute every candidate through transportations-library before returning it, so results are proved compliant rather than asserted. No database is required.

• reasoning_propagate_change - Forward-chain: downstream parameters affected by a changed input

• reasoning_diagnose_failure - Backward-chain: upstream causes of a failing parameter

• reasoning_repair_design - Abductive repair: minimal compliant fix for a Two-Lane Highway (HCM Ch.15)

• reasoning_repair_freeway - Abductive repair: minimal compliant fix for a Basic Freeway (HCM Ch.12)

• reasoning_reconcile_codes - Defeasible adjudication of conflicting code provisions, with an argument trace

• reasoning_inverse_design - Goal-directed synthesis: feasible geometries reaching a target LOS

Dependencies: the reasoning functions require transportations-validator>=0.2.0 and transportations-library>=0.1.12 (the latter for the BasicFreeways binding used by reasoning_repair_freeway). Both are on PyPI, so a normal pip install (or uv sync) resolves them.

API Endpoints

Hit the API endpoints directory to perform analyses or query HCM documentation.

Note: /docs for detail api endpoints description is under construction and will be available soon.

Core Endpoints

• POST /tools/call - Execute any registered function

• POST /tools/list - List available functions with filtering

• GET /mcp/discovery - MCP capability discovery

Per-method HCM analysis

POST /analysis/hcm/analyze                # {method, config}
POST /analysis/hcm/describe               # {method?} - catalog, or one method's schema + worked example
POST /analysis/hcm/validate               # {method, config} - parse and check, without running

POST /analysis/hcm/<method-with-hyphens>  # method-shaped convenience route, e.g. /analysis/hcm/analyze-roundabout

The per-method routes take {"config": { ... }} in that method's example-case schema. See HCM Analysis Capabilities above for the full list.

Chapter 15 Analysis

• POST /analysis/chapter15/complete - Complete HCM analysis

• POST /analysis/chapter15/segment - Single segment analysis

Research

• POST /tools/query-hcm - Query HCM database

• POST /research/search_hcm_by_chapter - Search HCM content by specific chapter

• GET /research/get_hcm_section - Get specific HCM section content

• POST /research/summarize_hcm_content - Summarize HCM content for a topic

Reasoning & Validation

Dedicated endpoints (and therefore first-class MCP tools) for the X-KG reasoning layer and full-corpus validation. Each resolves its implementation from the registry, so the surface stays in sync with function_registry.yaml.

• POST /reason/propagate-change - Forward-chain downstream impacts

• POST /reason/diagnose-failure - Backward-chain upstream causes

• POST /reason/repair-design - Minimal compliant fix (Two-Lane Highway, HCM Ch.15)

• POST /reason/repair-freeway - Minimal compliant fix (Basic Freeway, HCM Ch.12)

• POST /reason/reconcile-codes - Defeasible multi-jurisdiction adjudication

• POST /reason/inverse-design - Goal-directed geometry synthesis

• POST /validate/design-full - Validate against the full rule corpus with citations + clarifications

Utility

• GET /health - Health check

• GET /registry/info - Registry information

• POST /registry/reload - Reload function registry

Data Models

Highway Segment

{
  "passing_type": 0,      # 0=PC, 1=PZ, 2=PL
  "length": 2.0,          # miles
  "grade": 2.0,           # percent
  "spl": 50.0,            # speed limit (mph)
  "volume": 760.0,        # vehicles/hour
  "volume_op": 1500.0,    # opposing volume
  "phf": 0.95,            # peak hour factor
  "phv": 5.0              # percent heavy vehicles
}

Highway Facility

{
  "segments": [...],      # list of segments
  "lane_width": 12.0,     # feet
  "shoulder_width": 6.0,  # feet
  "apd": 5.0,             # access points/mile
  "pmhvfl": 0.02,         # percent HV in fast lane
  "l_de": 0.0             # effective passing distance
}

Adding New HCM Chapters

1. Create Function Module

Create functions/chapter16.py:

def new_analysis_function(data: Dict[str, Any]) -> Dict[str, Any]:
    """Implementation for new analysis."""
    try:
        # Your implementation here
        return {"success": True, "result": "analysis_result"}
    except Exception as e:
        return {"success": False, "error": str(e)}

2. Update Registry

Add to functions_registry.yaml:

functions:
  chapter16:
    new_analysis:
      module: "functions.chapter16"
      function: "new_analysis_function"
      description: "New analysis function"
      category: "transportation"
      chapter: 16
      parameters:
        type: "object"
        properties:
          input_param:
            type: "string"
        required: ["input_param"]

3. Restart Server

The registry will automatically load the new functions.

Development

Running Tests

Note: Test will be added soon.

pytest tests/

Validating Registry

Note: Not used yet.

python scripts/validate_registry.py

Setting Up Development Database

python scripts/import_hcm_docs.py

Customization

Custom Analysis Models

Extend models in core/models.py:

class CustomAnalysisInput(BaseModel):
    parameter1: float = Field(description="Custom parameter")
    parameter2: str = Field(description="Another parameter")

Custom Functions

1. Implement function in appropriate module

2. Add to functions_registry.yaml

3. Restart server or call /registry/reload

Support

This project is beta version and mainly for research purpose for now. It is widely appreciated for any contributions or feedback!

For issues and questions:

• Open an issue on GitHub

• Check the API documentation at /docs

• Review function registry at /registry/info

• Validate setup with utility scripts