Stella MCP Server

A Model Context Protocol (MCP) server for creating and manipulating Stella system dynamics models. This enables AI assistants like Claude to programmatically build, read, validate, and save .stmx files in the XMILE format.

What is this for?

Stella is a system dynamics modeling tool used for simulating complex systems in fields like ecology, biogeochemistry, economics, and engineering. This MCP server allows AI assistants to:

• Create models from scratch - Build stock-and-flow diagrams programmatically

• Read existing models - Parse and understand .stmx files

• Validate models - Check for errors like undefined variables or missing connections

• Modify models - Add stocks, flows, auxiliaries, and connectors

• Save models - Export valid XMILE files that open in Stella Professional

This is particularly useful for:

• Teaching system dynamics modeling

• Rapid prototyping of models through natural language

• Batch creation or modification of models

• Documenting and explaining existing models

Related MCP server: OpenKer Modeler MCP Server

Installation

From PyPI

pip install stella-mcp

From source

git clone https://github.com/bradleylab/stella-mcp.git
cd stella-mcp
pip install -e .

Requirements

• Python 3.10+

• mcp>=1.0.0

Configuration

Via uvx (no install required)

If you have uv installed, the lowest-friction configuration runs the published package directly:

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

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "stella": {
      "command": "stella-mcp"
    }
  }
}

Claude Code

Add to your .claude/settings.json:

{
  "mcpServers": {
    "stella": {
      "command": "stella-mcp"
    }
  }
}

Development mode

If running from source:

{
  "mcpServers": {
    "stella": {
      "command": "python",
      "args": ["-m", "stella_mcp.server"],
      "cwd": "/path/to/stella-mcp"
    }
  }
}

Recommended Agent Workflow

For a new model:

1. build_model with a stable model_id and the full set of stocks, auxiliaries, and flows in one call (connector sync and validation run by default, so the response doubles as an inspection).

2. Fix validation errors with update_*, rename_variable, or delete_variable.

3. Extend incrementally with add_variables (batch) or the single-add tools.

4. simulate to sanity-check behavior (requires the sim extra).

5. Save with save_model.

For imported models:

1. read_model with compat_mode="permissive" to inspect warnings.

2. Run inspect_model to understand model structure.

3. Use compat_mode="strict" before final save when round-trip fidelity matters.

Available Tools

Model Creation & I/O

Templates

Model Building

Notes:

• Tools accept optional model_id so one MCP session can manage multiple models safely.

• create_model and read_model set the session's current model_id and return it.

• add_flow and add_aux support optional graphical_function payloads (ypts plus exactly one of xscale or xpts).

• add_stock/add_flow/add_aux reject duplicate variable names across variable types; add_connector requires both variables to exist.

• set_connector_routing can target a connector by connector_uid or by from_var + to_var.

• save_model and get_model_xml accept auto_layout (default true) and resolve_layout_violations (default false).

• read_model, save_model, and get_model_xml accept compat_mode:

  • permissive (default): continue with warnings

  • strict: fail on compatibility issues

• set_module_style updates module view styling and persists those attributes in XMILE view <group .../> elements.

• save_as_template writes user templates to ~/.stella-mcp/templates by default (override via STELLA_MCP_TEMPLATE_DIR) and stores metadata in a .meta.json sidecar.

• Tool failures return structured MCP errors with error.code, error.category, and error.message.

Model Inspection

Batch Building

build_model creates and populates a model in one call. Items apply in the order stocks → auxs → flows → connectors → modules; the whole batch is all-or-nothing, and on failure the error names the failing item (error.stage + error.index). The same item arrays work on an existing model via add_variables.

{
  "name": "build_model",
  "arguments": {
    "name": "SIR",
    "model_id": "sir",
    "sim_specs": {"start": 0, "stop": 100, "dt": 0.125, "time_units": "Days"},
    "stocks": [
      {"name": "Susceptible", "initial_value": "9999", "units": "people"},
      {"name": "Infected", "initial_value": "1", "units": "people"},
      {"name": "Recovered", "initial_value": "0", "units": "people"}
    ],
    "auxs": [
      {"name": "contact_rate", "equation": "6"},
      {"name": "infectivity", "equation": "0.25"},
      {"name": "recovery_time", "equation": "2", "units": "days"},
      {"name": "total_population", "equation": "Susceptible + Infected + Recovered"}
    ],
    "flows": [
      {"name": "infection", "equation": "Susceptible * contact_rate * infectivity * Infected / total_population", "from_stock": "Susceptible", "to_stock": "Infected"},
      {"name": "recovery", "equation": "Infected / recovery_time", "from_stock": "Infected", "to_stock": "Recovered"}
    ],
    "modules": [
      {"name": "Disease Dynamics", "members": ["Susceptible", "Infected", "Recovered"]}
    ]
  }
}

Connector sync and validation run by default (disable with "sync_connectors": false / "validate": false); the response includes the full structured model summary, so no follow-up inspect_model call is needed.

Tool Payload Examples

Create and switch between session models:

{"name":"create_model","arguments":{"name":"Population","model_id":"pop_v1"}}

{"name":"create_model","arguments":{"name":"Carbon","model_id":"carbon_v1"}}

{"name":"list_models","arguments":{}}

{"name":"delete_model","arguments":{"model_id":"pop_v1"}}

{"name":"inspect_model","arguments":{"model_id":"sir_baseline","include_validation":true}}

List and load templates:

{"name":"list_templates","arguments":{}}

{"name":"list_templates","arguments":{"source":"builtin","query":"epidem","tags":["epidemiology"]}}

{"name":"get_template_info","arguments":{"template_name":"sir"}}

{"name":"load_template","arguments":{"template_name":"sir","model_id":"sir_baseline"}}

Save current model as a user template:

{"name":"save_as_template","arguments":{"model_id":"pop_v1","template_name":"my_population_template","description":"Baseline single-stock growth starter","tags":["intro","population"]}}

Create and manage modules:

{"name":"create_module","arguments":{"model_id":"sir_baseline","name":"Disease Dynamics","members":["Susceptible","Infected","Recovered"]}}

{"name":"add_to_module","arguments":{"model_id":"sir_baseline","module_name":"Disease Dynamics","members":["infection","recovery"]}}

{"name":"list_modules","arguments":{"model_id":"sir_baseline"}}

{"name":"remove_from_module","arguments":{"model_id":"sir_baseline","module_name":"Disease Dynamics","members":["recovery"]}}

{"name":"rename_module","arguments":{"model_id":"sir_baseline","module_name":"Disease Dynamics","new_name":"Disease Core"}}

{"name":"delete_module","arguments":{"model_id":"sir_baseline","module_name":"Disease Core"}}

Rename and delete variables safely:

{"name":"rename_variable","arguments":{"model_id":"sir_baseline","old_name":"population_total","new_name":"total_population"}}

{"name":"delete_variable","arguments":{"model_id":"sir_baseline","name":"recovery"}}

{"name":"delete_variable","arguments":{"model_id":"sir_baseline","name":"Susceptible","force":true}}

Update an existing variable:

{"name":"update_flow","arguments":{"model_id":"pop_v1","name":"growth","equation":"Population * growth_rate * stress_modifier"}}

Infer missing connectors from equations:

{"name":"sync_connectors_from_equations","arguments":{"model_id":"pop_v1"}}

Set module view geometry directly:

{"name":"set_module_view","arguments":{"model_id":"sir_baseline","module_name":"Disease Dynamics","x":420,"y":280,"width":420,"height":240}}

Set module view style:

{"name":"set_module_style","arguments":{"model_id":"sir_baseline","module_name":"Disease Dynamics","border_color":"#666666","background":"#FFF7E6","font_color":"#333333","font_size":"10pt","label_side":"top"}}

Auto-place module boxes from current member positions:

{"name":"auto_place_module_boxes","arguments":{"model_id":"sir_baseline","padding":40,"only_missing":true}}

Target a specific model in later calls:

{"name":"add_stock","arguments":{"model_id":"pop_v1","name":"Population","initial_value":"100"}}

Read with strict compatibility checks:

{"name":"read_model","arguments":{"filepath":"./external_model.stmx","model_id":"imported","compat_mode":"strict"}}

Preview XML in permissive mode (default) and return compatibility warnings when present:

{"name":"get_model_xml","arguments":{"model_id":"imported","compat_mode":"permissive"}}

Valid graphical function payload:

{
  "name": "add_aux",
  "arguments": {
    "model_id": "pop_v1",
    "name": "lookup_rate",
    "equation": "GRAPH(Time)",
    "graphical_function": {
      "xscale": {"min": 0, "max": 100},
      "ypts": [0.1, 0.2, 0.4, 0.6],
      "type": "continuous"
    }
  }
}

Invalid graphical function payload (rejected):

{
  "name": "add_aux",
  "arguments": {
    "name": "bad_lookup",
    "equation": "GRAPH(Time)",
    "graphical_function": {
      "xscale": {"min": 0, "max": 100},
      "xpts": [0, 10, 20, 30],
      "ypts": [0.1, 0.2, 0.4, 0.6]
    }
  }
}

Example Usage

Creating a simple population model

User: Create a simple exponential growth model with a population starting at 100
      and a growth rate of 0.1 per year

Claude: [Uses create_model, add_stock, add_aux, add_flow, add_connector, save_model]
        Creates population_growth.stmx with:
        - Stock: Population (initial=100)
        - Aux: growth_rate (0.1)
        - Flow: growth (Population * growth_rate) into Population

Reading and analyzing an existing model

User: Read the carbon cycle model and explain what it does

Claude: [Uses read_model, list_variables]
        This model has 3 stocks (Atmosphere, Land Biota, Soil) and 6 flows
        representing carbon exchange through photosynthesis, respiration...

Building a biogeochemical model

User: Create a two-box ocean model with surface and deep nutrients

Claude: [Uses create_model, add_stock (x4), add_aux (x8), add_flow (x6), save_model]
        Creates a model with nutrient cycling between surface and deep ocean
        including upwelling, downwelling, biological uptake, and remineralization

Diagram Preview

The render_diagram tool renders the model as an SVG stock-and-flow diagram — stocks as rectangles, auxiliaries as circles, flows as valved pipes (clouds mark sources/sinks), and dependency connectors as arcs. The SVG is returned inline so an agent can inspect the layout, and optionally written to a file you can open in any browser. It runs auto-layout first by default, so a freshly built model renders without manual positioning.

{"name":"render_diagram","arguments":{"model_id":"sir_baseline","filepath":"./sir.svg"}}

The diagram below is the built-in sir template rendered by render_diagram (no manual positioning):

Simulation

The simulate tool runs the current model and returns downsampled time series plus per-variable summaries (initial/final/min/max), closing the build→verify loop without opening Stella. It requires the optional PySD dependency:

pip install 'stella-mcp[sim]'

{"name":"simulate","arguments":{"model_id":"pop_v1","overrides":{"growth_rate":0.05},"include":["Population"],"max_points":50}}

Notes and caveats:

• PySD integrates with Euler only — models whose method is RK4 simulate with Euler and the response carries a warning. Results can differ from Stella for stiff systems.

• PySD supports a subset of XMILE; unsupported constructs fail with a structured error rather than wrong numbers.

• overrides accepts variable names in display ("growth rate") or underscore (growth_rate) form and replaces the variable with a constant.

• save_results_csv writes the full-resolution results table with a time column.

• The session model is never modified by simulation (the run uses a throwaway copy).

Scenario Comparison

The compare_scenarios tool answers "what happens under these alternative assumptions?" — it runs several named override sets against a baseline (the unmodified model by default) and reports how each diverges. Also requires the sim extra.

{"name":"compare_scenarios","arguments":{"model_id":"pop_v1","include":["Population"],"scenarios":[{"name":"low growth","overrides":{"growth_rate":0.02}},{"name":"high growth","overrides":{"growth_rate":0.08}}]}}

Each scenario reports its own downsampled series plus delta_vs_baseline per variable: final_abs, final_pct (percent change of the final value), and max_abs. Notes:

• Every override name across all scenarios is validated before any run, so a typo fails fast and atomically — no scenario runs half-applied.

• A scenario whose run produces NaN/inf reports the warning in that scenario's warnings without aborting the others; final_pct is null when the baseline final is zero (no divide-by-zero).

• baseline is optional — pass an override set to measure deltas against, or omit it to compare against the unmodified model.

• save_comparison_csv writes a wide table with one column per variable__scenario (and variable__baseline).

• The compiled model is reused across every scenario in one call, so a comparison is roughly as cheap as a single simulation plus one run per scenario.

Sensitivity Analysis

The sensitivity_analysis tool answers "which parameters actually move the outcome?" — it sweeps each parameter one at a time across a range (holding the others at their baseline) and reports how a single chosen output metric responds. Also requires the sim extra.

{"name":"sensitivity_analysis","arguments":{"model_id":"pop_v1","parameters":[{"name":"growth_rate","start":0.02,"stop":0.08,"steps":7}],"output":{"variable":"Population","metric":"final"}}}

For each parameter it returns the metric at every swept value, a range_sensitivity (the metric's average slope across the swept range), and a baseline-normalized elasticity (≈ Δoutput% / Δparam%) so parameters can be ranked by influence. Notes:

• One-at-a-time only. mode accepts "oat"; full-factorial (grid) and Monte-Carlo sampling are reserved for a future release.

• metric is one of final, max, min, mean, or time_to_threshold (which needs an output.threshold and reports the first time the series crosses it). max/min/mean cover finite values only; a non-finite or never-crossing run reports null for that point with a warning.

• A parameter spec is either start/stop/steps (evenly spaced, steps ≥ 2) or an explicit values list (≥ 2 entries).

• max_runs (default 200) caps the total swept runs; an oversized sweep errors rather than silently truncating. OAT runs are a sum across parameters, not a product, so the cap only trips on genuinely large sweeps.

• elasticity is null when it cannot be defined (a non-constant parameter, or a zero baseline metric/parameter); range_sensitivity is still reported.

• save_sweep_csv writes a long parameter, value, metric table.

• Like scenario comparison, the model is compiled once and reused across the whole sweep.

MCP Resources & Prompts

Beyond tools, the server exposes MCP-native affordances:

• Tool annotations. Every tool carries hints (readOnlyHint, destructiveHint, idempotentHint) so clients can manage permissions and parallelize read-only calls. Inspection tools (inspect_model, validate_model, list_*, get_model_xml) are read-only; delete_* are marked destructive.

• Resources. Templates and session models are readable as resources:

  • stella://templates/{name} — a built-in or user template's .stmx

  • stella://models/{model_id} — a session model's current XMILE export

• Prompt. A build-stella-model prompt (argument: description) encodes the recommended build → validate → simulate → render → save workflow, so it is discoverable inside MCP clients.

Validation

The validate_model tool checks for:

• Undefined variables - References to variables that don't exist

• Mass balance issues - Stocks without flows, flows referencing non-existent stocks

• Missing connections - Equations using variables without connectors (warning)

• Connector endpoint integrity - Connectors pointing at missing variables (error)

• Orphan flows - Flows not connected to any stock

• Circular dependencies - Infinite loops in auxiliary calculations

• Module integrity - Empty modules (warning) and modules referencing missing members (error)

• Units present - A stock or flow missing units while others define them (warning)

• Units consistency - A flow whose units don't read as stock-units/time-unit when every attached stock shares the same units (warning; conservative — stays silent on conversion flows and anything it can't confidently parse)

• Unused auxiliaries - An auxiliary referenced by no equation or connector (warning); stocks and flows are never flagged

XMILE Compatibility

• Output files use the XMILE standard

• Compatible with Stella Professional 1.9+ and Stella Architect

• Auto-layout positions elements reasonably; adjust manually in Stella if needed

• Variable names with spaces are converted to underscores internally

• Parser normalizes imported stock inflow/outflow and connector endpoint references

• Time-step export avoids lossy reciprocal rounding (non-exact reciprocals are exported as plain dt)

• Import/export preserves unknown attrs/elements on supported sections (header, sim_specs, variables, views/model extras) to reduce round-trip data loss

• Compatibility corpus regression tests live in tests/fixtures/compat_corpus/ and run in CI

• Maintainer helper: python scripts/sync_compat_corpus_manifest.py --check validates corpus manifest sync

Project Structure

stella-mcp/
├── README.md
├── LICENSE
├── pyproject.toml
└── stella_mcp/
    ├── __init__.py
    ├── server.py      # MCP server wiring + schemas
    ├── tool_handlers.py # Tool handler implementations/registration
    ├── tool_schemas.py  # MCP tool schema definitions
    ├── xmile.py       # Core model types + layout logic
    ├── xmile_io.py    # XMILE parsing/export helpers
    └── validator.py   # Model validation logic

Contributing

Contributions are welcome! Please feel free to submit issues or pull requests.

Maintainer Release

PyPI publishing is handled by .github/workflows/publish.yml using PyPI Trusted Publishing. To release a new version:

1. Update the version in pyproject.toml and stella_mcp/__init__.py, and move the [Unreleased] items in CHANGELOG.md under the new version heading.

2. Merge the release changes to main.

3. Create and publish a GitHub release with a matching tag, for example v0.5.0.

The GitHub release event builds the source distribution and wheel, then publishes them to PyPI through the configured trusted publisher.

License

MIT License - see LICENSE for details.

Acknowledgments

• Model Context Protocol by Anthropic

• ISEE Systems for Stella and the XMILE format