MCP 服务器

Geneva Forecasting MCP

Geneva MCP brings production-grade forecasting directly into AI assistants and coding agents. Connect any MCP-compatible client to the Geneva Forecasting Engine and run rigorous time series forecasts through natural conversation.

README

Geneva MCP

Time-Series Forecasting & Demand Planning for AI Agents Powered by the <a href="https://portal.roadmap-tech.com">Geneva Forecasting Engine</a> — by <a href="https://portal.roadmap-tech.com">RoadMap Technologies</a>

Geneva MCP brings production-grade statistical forecasting directly into AI assistants and coding agents. Connect any MCP-compatible client to the Geneva Forecasting Engine and run rigorous time series forecasts through natural conversation — no code, no pipelines, no exports required.

"Forecast the next 12 months of this sales data with 95% prediction intervals."

The agent calls the Geneva Engine behind the scenes, receives a numerical analysis and a forecast chart, and delivers a complete, statistically grounded interpretation — all in a single conversation turn.

Why Geneva MCP?

Free to Get Started — No Credit Card Required

The Geneva API is free with 1,000 forecasts per month. Sign up at the Developer Portal, generate your API key, and start forecasting in minutes. No trial periods, no paywalls, no surprises.

Forecast in Claude Chat

With Geneva MCP connected, Claude becomes a full forecasting analyst. Paste your data directly into the conversation, ask a question in plain language, and Claude will forecast, analyze, and visualize the results — model selection, accuracy metrics, trend commentary, and a chart, all delivered in a single response.

<img src="https://d3deh4arfmquak.cloudfront.net/claude-geneva-mcp.png" alt="Geneva MCP in Claude Desktop — natural language forecasting with chart" width="720"> Claude with Geneva MCP — describe your data, get a complete forecast with analysis and chart in one turn

Forecast from Excel with Claude Cowork

Claude Cowork users can forecast directly from spreadsheets. Drop your .xlsx or .csv file into the conversation, describe the column you want to forecast, and Geneva handles the rest — complete with prediction intervals and a chart. No data pipelines, no Python scripts, no manual exports.

From spreadsheet to forecast in one sentence.

<img src="https://d3deh4arfmquak.cloudfront.net/geneva-claude-excel.png" alt="Geneva MCP with Claude Cowork — forecasting directly from an Excel spreadsheet" width="720"> Claude Cowork with Geneva MCP — attach an Excel or CSV file and forecast any column in plain language

Fast & Purpose-Built

Geneva is a dedicated forecasting engine, not a general-purpose language model. Forecasts return in milliseconds with structured output and a chart — every time, with consistent and reproducible results. Geneva returns both a full numerical analysis and a chart image ready for the agent to embed directly in the conversation.

<img src="https://d3deh4arfmquak.cloudfront.net/geneva_forecast.png" alt="Geneva Forecast — time series chart with prediction intervals" width="720"> Chart generated by Geneva MCP — historical data, forecast line, and confidence interval bands at your chosen level

Requirements

Requirement	Details
Python	3.10 or later
API Key	Free — sign up at portal.roadmap-tech.com
Agent	Claude Desktop or any MCP-compatible client

Security: Your API key is injected as a process environment variable in your agent's MCP configuration. It is never exposed in the chat, the agent context window, or transmitted to any third party.

Installation

Option 1 — uvx (Recommended)

uvx runs Geneva MCP in an isolated environment with no permanent installation — it pulls the latest version from PyPI automatically each time. This is the cleanest approach and what the configuration examples below use.

First, install uv (a fast Python package manager) if you don't have it:

Platform	Command
Mac	`brew install uv`
Windows	`powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 \| iex"`
Linux / Mac (no Homebrew)	`curl -LsSf https://astral.sh/uv/install.sh \| sh`

Then run the server directly with:

uvx geneva-mcp

uv is open source and maintained by Astral — the same team behind the popular ruff linter. It is the package manager recommended by Anthropic for MCP server distribution.

Option 2 — pip

If you already have Python 3.10+ and prefer a traditional install, pip works without any additional tooling:

pip install geneva-mcp

Then use geneva-mcp as the command in your agent config instead of uvx geneva-mcp.

When to use pip vs uvx: uvx is self-contained and version-isolated — ideal for running MCP servers where you don't want package conflicts. pip installs into your active Python environment and is the simpler choice if you're already managing a Python setup.

Supported Agents

Agent	Status	Notes
Claude Desktop	✅ Fully Supported	Full multimodal output (text + chart)
Claude Cowork	✅ Fully Supported	Forecast directly from Excel & CSV files
Any MCP-compatible client	✅ Supported	Via stdio transport

Quick Start

Step 1 — Get Your Free API Key

Visit the Geneva Developer Portal to create a free account and generate your API key. The free tier includes 1,000 forecasts per month with no credit card required.

Step 2 — Configure Claude Desktop

Add the Geneva MCP server to your Claude Desktop configuration. Replace gva_your_key_here with your API key.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "geneva": {
      "command": "uvx",
      "args": ["geneva-mcp"],
      "env": {
        "GENEVA_API_URL": "https://api.roadmap-tech.com",
        "GENEVA_API_KEY": "gva_your_key_here"
      }
    }
  }
}

Restart Claude Desktop — the forecast tool will appear automatically in the tool list.

If you installed via pip

Point command at the geneva-mcp binary inside your virtual environment:

{
  "mcpServers": {
    "geneva": {
      "command": "{path_to_venv}/venv/bin/geneva-mcp",
      "env": {
        "GENEVA_API_URL": "https://api.roadmap-tech.com",
        "GENEVA_API_KEY": "gva_your_key_here"
      }
    }
  }
}

Replace {path_to_venv} with the absolute path to the directory containing your virtual environment. You can find the binary path by running which geneva-mcp with the venv activated.

Claude Cowork

Follow the same configuration as Claude Desktop above. Once configured, you can attach an .xlsx or .csv file to the conversation and instruct Claude to forecast any column — Geneva will handle model selection, computation, and visualization automatically.

Step 3 — Start Forecasting

Ask your agent to forecast any time series in plain language:

"Here are my last 36 months of revenue: [120, 135, 128, ...]. Forecast the next 18 months with 95% confidence intervals."

The agent receives structured forecast data and an embedded chart visualization, then provides a complete analysis covering model selection rationale, accuracy metrics, trend direction, seasonal patterns, and prediction intervals.

How It Works

┌──────────────────────┐     stdio (JSON-RPC)     ┌──────────────────────┐
│  Claude Desktop      │ ◄──────────────────────► │  Geneva MCP Server   │
│                      │                          │  (local process)     │
│                      │                          │                      │
└──────────────────────┘                          └────────┬─────────────┘
                                                           │ HTTPS
                                                           ▼
                                                  ┌──────────────────────┐
                                                  │  Geneva Engine API   │
                                                  │  api.roadmap-tech.com│
                                                  └──────────────────────┘

The MCP server runs locally as a lightweight Python process. It communicates with the AI agent over stdio using the MCP protocol and calls the Geneva Forecasting Engine API over HTTPS. The server is fully stateless and produces no local side effects.

Tool Reference

`forecast`

Run a single time series forecast through the Geneva Forecasting Engine.

Note: Each tool call processes one time series. For batch forecasting across multiple series, use the Geneva Python SDK directly.

Parameter	Type	Required	Default	Description
`data`	`list[float]`	✅	—	Time series observations (3–10,000 data points)
`horizon`	`int`	—	auto	Number of future periods to forecast
`wave_periods`	`list[int]`	—	`[12]`	Seasonal cycle lengths (`12`=monthly, `4`=quarterly, `52`=weekly, `7`=daily)
`confidence_level`	`float`	—	`0.95`	Prediction interval confidence level (0.0–1.0)
`method`	`int`	—	auto	Force a specific method (0–9). Omit to let the Expert System select automatically
`seasonal_transform`	`int`	—	`0`	Seasonal transform: `0`=None, `1`=Seasonal, `2`=MPT
`smoothing`	`bool`	—	`false`	Enable median smoothing for noisy data
`max_periods_factor`	`float`	—	engine default	Fit window cap factor — increase for long series
`holdout_ratio`	`float`	—	engine default	Fraction of data reserved for model evaluation

Returns: A structured multimodal response containing:

Text summary — Selected model name, accuracy metrics (MAD, MAPE, RMSE), forecast values with prediction intervals, in markdown format ready for agent interpretation
Chart image — A forecast chart with historical data, forecast line, and shaded prediction interval bands

Forecasting Methods

The Geneva Expert System automatically evaluates all 10 methods and selects the best fit for your data. You can also specify a method manually using the method parameter.

ID	Method	Best For
0	Linear Regression	Steady linear growth or decline
1–5	Non-Linear Regression (5 curve types)	Exponential, s-curve, power trends
6	Simple Exponential Smoothing	Stable, level series
7	Double Exponential Smoothing (Holt)	Trending series without seasonality
8	Holt-Winters (Seasonal)	Trending + seasonal series
9	Croston (Intermittent Demand)	Sporadic, zero-inflated series

Troubleshooting

`GENEVA_API_KEY environment variable is not set`

The env block is missing from your agent's MCP configuration, or the API key value is empty. Add a valid key from the Geneva Developer Portal.

`Server transport closed unexpectedly`

The Python process crashed before the MCP handshake completed. Verify your installation by running the server manually:

GENEVA_API_KEY=your_key GENEVA_API_URL=https://api.roadmap-tech.com geneva-mcp

Chart rendering issues

The server uses a headless matplotlib backend (Agg) that requires no display server. If chart generation fails, reinstall matplotlib:

pip install matplotlib --force-reinstall

Related Resources

Resource	Description
Geneva Developer Portal	Free API keys, usage dashboard, and account management
Geneva Python SDK	Direct API client with batch forecasting support
Geneva API Documentation	Full API reference, parameter guides, and examples