ModelRisk MCP

<!-- mcp-name: io.github.vosesoftware/modelrisk-mcp -->

An open Model Context Protocol server for [Vose Software's ModelRisk](https://www.vosesoftware.com).

Use it with Claude Desktop, Claude Code, Cursor, Zed, or any MCP-compliant client to read, build, fit, and run Monte Carlo risk models in Excel from a conversation.

ModelRisk MCP is an open MCP server on the standard Anthropic Model Context Protocol. No proprietary layer, no lock-in. The 1417-entry function catalogue, the Vose methodology principles, and the audit rule set are all included in the package — and editable.

Stable: `0.3.2` — the server starts ModelRisk itself when no Excel is running (auto-launch + add-in activation); programmatic simulation via run_simulation (XLL command surface, no fragile COM dispatch); .vmrs results read via the official ModelRisk SDK; activation ships bundled so no environment configuration is required. 50 tools across reading, building, fitting, simulation, scenario-sweep, restore, charting, audit, a quantitative analysis & decision layer (distribution calculator, tail risk, fit-and-rank, fast aggregates, correlation matrix, distribution comparison, backtest, uncertainty decomposition), and workflow surfaces.

Purpose

Monte Carlo risk modelling is the right way to reason about an uncertain future — and almost nobody does it.

The mathematics has been settled for decades: when your inputs are uncertain, you don't reason with a single "best guess" number, you reason with the distribution of outcomes. A point estimate ("Q3 revenue will be $4.2M") hides exactly the thing a decision-maker needs to know — how wrong could it be, how bad is the downside, which assumption is driving the risk. ModelRisk has given Excel users a rigorous, validated engine for this for years.

The barrier was never the maths. It was the friction:

• The discipline carries a real skill curve — 1,400+ functions, the right distribution family for each input, the methodology traps (fit without parameter uncertainty, correlations ignored, risk events modelled as p × impact) that quietly produce confident-but-wrong answers.
• Setting a model up by hand is slow and fiddly — wrapping inputs, naming outputs, wiring copulas and time-series, formatting the report.
• So under deadline pressure, most teams fall back to a single-point spreadsheet and a gut-feel range. The rigorous tool sits unused.

This server removes that friction. Large language models can now drive domain tools through the open Model Context Protocol. ModelRisk MCP puts the whole ModelRisk surface — build, fit, simulate, audit, interpret — behind a conversation. You describe the problem in plain language; Claude proposes the right distributions, wires the structure, runs the simulation, reads the tail, and writes the report. The expertise moves into the loop; the friction drops to a sentence.

It is built on five deliberate principles:

1. Open, not locked-in. A server on the standard Anthropic MCP — works with Claude Desktop, Claude Code, Claude for Excel, Cursor, Zed, and any compliant client. MIT-licensed. The function catalogue, methodology, and audit rules ship in the package and are editable. No proprietary connector, no vendor cage.
2. Methodology-grounded, not just mechanically capable. The server is opinionated about correct practice. Fits default to uncertainty=TRUE; risk events use the bimodal VoseRiskEvent; the 13-rule audit encodes the mistakes Vose practitioners have seen across decades of consulting. It won't just do what you ask — it'll steer you toward what's right.
3. Local-only, no telemetry. Everything runs on your machine against your Excel and your ModelRisk install. No data leaves your computer; the activation key is bundled and offline.
4. Excel stays the model. No re-platforming, no shadow tooling. The workbook is the model — versionable in Git, openable by anyone with Excel + ModelRisk, reproducible by re-running one tool call.
5. Safe by design. Every write previews first (dry_run=True), lands in Excel's undo stack, and is logged. The server can modify your workbook — and does so under nine layered safeguards (see Safety by design).

The goal is simple: make defensible, quantitative risk analysis something you reach for by default — because it's now no harder than asking.

What this does

This server turns Claude (or any MCP client) into a methodology-aware co-pilot for ModelRisk. It can:

• Build new Monte Carlo models from a description — insert distributions, fit families to data, build aggregates, copulas, time-series, risk events.
• Run simulations from the conversation. run_simulation triggers the same XLL command the ribbon "Simulate" button uses, blocks until the run finishes, saves a .vmrs next to the workbook, and auto-pins it as the results source.
• Read model structure and per-iteration results — inputs, outputs, percentiles, correlation matrices, tornado rankings — directly from .vmrs files via ModelRisk's official SDK (MRService.dll). No COM dispatch fragility.
• Audit a workbook against Vose's methodology rules and propose fixes.
• Interpret results into structured executive summaries with contingency analysis.

Every formula written to Excel is validated against the ModelRisk function catalogue first — there's no path to a hallucinated function name reaching your workbook.

See the [user manual](docs/user-manual.md) for a walkthrough of the nine things you can do, a realistic end-to-end example, and what the server explicitly does and doesn't do. New to Monte Carlo or to the ModelRisk MCP toolchain? Start with the [15-minute quick-start tutorial](docs/quick-start.md), then pick a [walk-through scenario](docs/scenarios.md) matching your problem (budgets, data fitting, loss aggregation, correlated inputs, stress tests, model audits); unfamiliar with a term, see the [glossary](docs/glossary.md).

Feature comparison

Install

Prerequisites

• Windows 10 or 11, 64-bit
• Excel 2019 or newer with the ModelRisk add-in installed and loaded
• One of:
• Python 3.11+ (recommended via uv)
• Or the standalone modelrisk-mcp.exe from the latest release — no Python knowledge required

Activation (results reading): None required — MRService.dll (the SDK that reads .vmrs files) is activated automatically by a bundled offline key. Two env overrides exist for non-standard setups (both only affect reading .vmrs results; simulations run regardless):

• MRSERVICE_DLL_PATH — full path to an MRService.dll if it isn't in the standard ModelRisk folder, or if your ModelRisk's MRService is a version the bundled key doesn't cover. Point it at a copy whose version the key accepts (e.g. one under …\Vose Software\Tamara\MRLibrary\MRService.dll); its sibling DLLs resolve from that folder, so point at it in place.
• MRSERVICE_ACTIVATION_KEY — your own activation key, to override the bundled one for any MRService version.

Recommended: auto-wire from PyPI

pip install modelrisk-mcp
modelrisk-mcp install      # adds the server to Claude Desktop / Claude Code config

Then restart Claude Desktop. modelrisk-mcp install edits claude_desktop_config.json for you (backing it up first) and works on every current Claude version. (Excel + a licensed ModelRisk add-in are still required — that's the product itself.)

One-click .mcpb (Claude Desktop Extension)

Download `modelrisk-mcp.mcpb` from the latest release, open it (Claude Desktop → Settings → Extensions → Install Extension…), then restart Claude. The bundle ships the server as a standalone exe — no Python, no config edit.

⚠️ Known issue (Claude Desktop, latest Windows MSIX builds, e.g. 1.12603.x): the Extensions installer can silently do nothing when you pick a .mcpb — no error, no install. This is a Claude Desktop installer bug (it fails before logging), not a problem with the bundle (it validates with mcpb and installs fine once the client is fixed). Until Anthropic patches it, use the `pip install` + `modelrisk-mcp install` path above — it bypasses the Extensions UI entirely.

From PyPI

pip install modelrisk-mcp

From source

git clone https://github.com/vosesoftware/modelrisk-mcp
cd modelrisk-mcp
uv sync
uv run python -m modelrisk_mcp     # speaks MCP over stdio

Standalone .exe

Download modelrisk-mcp.exe from Releases, drop it anywhere on disk, and point Claude Desktop at it. See docs/claude-desktop.md.

Wire into Claude Desktop

Three options, simplest first.

One-command auto-wire (recommended)

pip install modelrisk-mcp
modelrisk-mcp install

modelrisk-mcp install detects every installed MCP client (Claude Desktop, Claude Code), backs up its existing config, and adds the modelrisk server entry — preserving any other servers you already have configured. Output looks like:

  + Claude Desktop   added    C:\Users\you\AppData\Roaming\Claude\claude_desktop_config.json
      Registered 'modelrisk' -> {'command': 'C:\\...\\Scripts\\modelrisk-mcp.exe'}
      backup: ...claude_desktop_config.json.bak.20260521-153000

Restart Claude Desktop / Claude Code to pick up the new server.

To undo: modelrisk-mcp uninstall. To register a second instance with a different name (e.g. dev and prod side-by-side): modelrisk-mcp install --name=modelrisk-dev.

Zero-install via uvx (if you already use uv)

If you have uv installed, you can skip the pip install step entirely. Just add to %APPDATA%\Claude\claude_desktop_config.json directly:

{
  "mcpServers": {
    "modelrisk": {
      "command": "uvx",
      "args": ["modelrisk-mcp"]
    }
  }
}

uvx downloads modelrisk-mcp into an ephemeral cache on first run and updates automatically when new versions hit PyPI.

Manual JSON edit (if you must)

Open %APPDATA%\Claude\claude_desktop_config.json and add the entry by hand:

{
  "mcpServers": {
    "modelrisk": {
      "command": "C:/path/to/modelrisk-mcp.exe"
    }
  }
}

Use the absolute path to the .exe you downloaded from the latest release, or "command": "python", "args": ["-m", "modelrisk_mcp"] if you pip installed.

After any of the three, restart Claude Desktop so it spawns the MCP server subprocess. The ModelRisk tools appear under the connections icon. Full guide: docs/claude-desktop.md. Claude Code setup: docs/claude-code.md.

Wire into Claude for Excel (HTTP transport)

Claude for Excel runs inside an Office.js sandbox and can't spawn subprocesses, so it talks to MCP servers over HTTP. Start the server in HTTP mode:

$env:MODELRISK_MCP_TOKEN = [Guid]::NewGuid().ToString("N") * 2
modelrisk-mcp --transport=streamable-http --port=8000 --token=$env:MODELRISK_MCP_TOKEN

Then in Claude for Excel: Settings → Connectors → Add MCP server, URL http://127.0.0.1:8000/mcp, paste the token. Full guide: docs/claude-for-excel.md.

Why this is interesting: Claude for Excel's sandbox can't reach Excel's COM surface or the ModelRisk ribbon on its own. ModelRisk MCP runs outside the sandbox and bridges that gap — Claude for Excel can do things via this server it structurally can't do otherwise (run simulations, dispatch ModelRisk COM, write distributions through the safety pipeline).

First conversation

Open a workbook in Excel that has at least one Vose function — even a single =VoseNormal(0,1). Then in Claude:

Summarise the active workbook's risk model — inputs, outputs, distributions.

Or jump straight into building:

/build-risk-model

This walks through 9 steps, from identifying outputs through running the simulation and interpreting results. See the slash-command catalogue for the other workflows.

Safety by design

The server can both read and modify your workbook — that's the central differentiator. We make that safe with nine layered mechanisms (spec §11):

1. `dry_run=True` is the default on every building tool. Claude must explicitly pass dry_run=False to commit. Previewing comes free; a forgotten flag becomes a preview, never an overwrite.
2. Every write lands in Excel's native undo stack. Ctrl+Z works exactly as you'd expect.
3. Bulk-write guard. Tools writing >50 cells in one call require explicit confirm_bulk=True. Time-series and copula tools — which write contiguous ranges by design — are exempt.
4. No automatic saves. The server never calls Workbook.Save(). You control Ctrl+S.
5. No overwriting non-Vose formulas. A formula-tokenised detector (not a substring check) refuses to overwrite a cell whose existing formula uses non-Vose functions. The one tool explicitly allowed to do this is replace_constant_with_distribution, by design.
6. Audit log of every write in %LOCALAPPDATA%\VoseSoftware\modelrisk-mcp\writes.log — timestamp, cell, before/after formulas, before value. JSONL, append-only.
7. Read-only mode. Launch with --read-only to disable every building/simulation tool.
8. Single-writer mutex. Two MCP server instances can't drive the same Excel concurrently — the second instance raises ConcurrentWriterError on any building tool call.
9. Restore from audit log. The restore_cell tool reads writes.log and rewrites the pre-write formula — even after Excel's undo stack has been cleared.

What's inside

• 52 tools — 12 reading, 14 building (incl. create_aggregate for FFT / Panjer / MC frequency-severity methods), 5 simulation (incl. run_simulation, run_scenarios, get_samples, restore_cell, restore_deterministic_state), 8 analysis (compute_distribution, fit_and_rank_distributions, get_tail_risk, fit_tail, compute_correlation_matrix, compare_distributions, backtest_output, decompose_uncertainty), 10 workflow / reporting (incl. plan_risk_model, audit_model, diagnose_workbook, create_tornado_chart, create_histogram_chart, create_cdf_chart, build_drivers_report, build_executive_report, generate_executive_summary, save_workbook_as), 3 VMRS (read_vmrs, set_active_vmrs, list_vmrs_variables)
• 5 resources — modelrisk://functions, modelrisk://distributions, modelrisk://methodology, modelrisk://workbook/current, modelrisk://audit-rules
• 5 slash-command prompts — /build-risk-model, /audit-model, /interpret-results, /add-uncertainty, /import-legacy-model
• 1417-entry function catalogue extracted directly from the ModelRisk IDL + XLL header
• 17 audit rules — 13 Monte-Carlo-methodology (VOSE-001 … VOSE-013) + 4 spreadsheet-integrity (SS-001 … SS-004) — with editable severity in data/audit_rules.yaml; add your own with docs/authoring-audit-rules.md
• Methodology-grounded distribution selection guide in data/distributions.yaml

Methodology

The server is opinionated about Monte Carlo methodology — fetch modelrisk://methodology from any MCP client to read the 8 core principles. Highlights:

• Every uncertain input is a distribution. Treating a noisy input as deterministic understates total uncertainty by exactly the amount it could swing.
• Distribution fits use uncertainty=TRUE. Carry parameter uncertainty through the simulation; don't pretend the best-fit parameters are exact.
• Risk events use VoseRiskEvent, not probability * impact. The bimodal nature matters.
• Correlated inputs use copulas. Independent inputs that are actually correlated produce artificially tight outputs.

Architecture

Three internal layers plus two external integration paths:

┌──────────────────────────────────┐
│  MCP client                      │
│  (Claude Desktop, Code, etc.)    │
└────────────────┬─────────────────┘
                 │ JSON-RPC / stdio (or HTTP)
                 ▼
┌──────────────────────────────────┐
│  FastMCP layer (tools, resources,│
│   prompts)                       │
├──────────────────────────────────┤
│  ModelRiskBridge (domain)        │
│  + SimulationController          │
│  + ResultsReader                 │
├──────────────────────────────────┤
│  ExcelBridge      MrServiceBridge│
│  (xlwings)        (ctypes)       │
└──────┬───────────────────┬───────┘
       │ Application.Run   │ MRLIB_*
       │ + cell I/O        │ (read .vmrs)
       ▼                   ▼
┌──────────────┐   ┌──────────────────┐
│ Excel +      │   │  MRService.dll   │
│ ModelRisk XLL│   │  (SDK)           │
└──────────────┘   └──────────────────┘

Two integration paths, each chosen for what it does best:

• Builds + simulation trigger → Excel COM via xlwings, plus Application.Run("VoseStartSimulCustom12", …) for the simulation kickoff. Mirrors what the ModelRisk ATL does internally; bypasses the fragile ATL CoClass Dispatch surface that doesn't expose IDispatch.
• Results read → MRService.dll directly via ctypes. Vose's official SDK opens .vmrs files, returns sample arrays, computes statistics. No COM round-trips per output; per-iteration sample arrays available for arbitrary downstream analysis.

More: docs/architecture.md, docs/com-surface.md.

Known caveats

• Starting ModelRisk is automatic (since 0.3.2). If no Excel is running when a tool is called, the server starts an attachable Excel and registers the ModelRisk XLL into it, so Vose functions resolve and simulations run. If Excel is already up but the add-in isn't live (e.g. ModelRisk's "Start with Excel" is off), the server auto-activates it before simulating, or returns a clear instruction if it can't. Disable auto-launch with MODELRISK_AUTO_LAUNCH=0 if you'd rather manage Excel yourself. (Earlier versions required you to open Excel + ModelRisk by hand first and could fail with an opaque "macro may not be available".)
• OneDrive-hosted workbooks: xlwings can fail to resolve the workbook's full path without ONEDRIVE_COMMERCIAL_WIN set. The bridge degrades gracefully — name-based operations still work, and run_simulation defaults the .vmrs save location to the user's Desktop when the workbook folder can't be resolved.
• Active simulation results: get_simulation_results reads from the .vmrs file produced by the most recent run_simulation call, or the most recent sibling .vmrs next to the workbook. Use set_active_vmrs(path) or read_vmrs(path) to point at a specific file.

License

MIT. See LICENSE.

Documentation

Links

• Vose Software: <https://www.vosesoftware.com>
• ModelRisk product page: <https://www.vosesoftware.com/products/modelrisk/>
• Source: <https://github.com/vosesoftware/modelrisk-mcp>
• Releases: <https://github.com/vosesoftware/modelrisk-mcp/releases>
• Issues: <https://github.com/vosesoftware/modelrisk-mcp/issues>
• Changelog: CHANGELOG.md
• MCP spec: <https://modelcontextprotocol.io>