FlexSim MCP Server
---
title: FlexSim MCP Server
emoji: ⚙️
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: "5.0.0"
app_file: app.py
pinned: false
---
# FlexSim MCP Server (incl. FlexSimPy)

Control and automate FlexSim simulations via the Model Context Protocol (MCP). This repository exposes a FastMCP server that launches FlexSim (via FlexSimPy) and provides tools to open models, run/stop, step, query/set node values, evaluate FlexScript, export results, and more.
Why this matters
- Bridges digital twins and AI assistants (Claude/Cursor) for manufacturing/warehouse analysis and “what-if” studies.
- Demonstrates robust systems engineering: GUI process orchestration, async I/O, protocol handling, Windows quirks, and reproducible automation.
---
## Design
```
┌─────────────────────┐ JSON-RPC ┌─────────────────────┐
│ MCP Client │◄────────────────────► │ MCP Server │
│ (Claude/Cursor) │ stdin/stdout │ (flexsim_mcp.py) │
└─────────────────────┘ └──────────┬──────────┘
│
│ Python API
▼
┌─────────────────────┐
│ FlexSimPy │
│ (Python SDK) │
└──────────┬──────────┘
│
│ COM/IPC
▼
┌─────────────────────┐
│ FlexSim │
│ (Simulation Engine)│
└─────────────────────┘
```
---
## Features
- MCP server entry point: `mcp_server/flexsim_mcp.py`
- Tools exposed (69 total: 22 base + 47 opencode; AST manda — full catalog in `GUIA_DESDE_CERO.md` §7):
- Simulation control: `flexsim_open_model`, `flexsim_reset`, `flexsim_run`, `flexsim_run_to_time`, `flexsim_stop`, `flexsim_step`, `flexsim_get_time`
- Model/script: `flexsim_execute_script`, `flexsim_evaluate_node`, `flexsim_save_model`, `flexsim_save_model_as`, `flexsim_close_model`, `flexsim_new_model`
- Node access: `flexsim_get_node_value`, `flexsim_set_node_value`, `flexsim_list_tree`
- Results/stats: `flexsim_get_statistics`, `flexsim_export_results`, `flexsim_run_experiments`, `flexsim_chart_snapshot`
- Logging: `mcp_server/flexsim_mcp.log`
- Reproducible environment: `uv sync`
- FlexSimPy build automation: `utility/build_automation.py`
- Integration tests: `tests/integration/test_mcp_client.py` (stdio-based MCP client)
- GUI sanity test: `utility/integration_test.py`
- Configurable behavior via `config.toml` and environment overrides
---
## Requirements
- OS: Windows
- Python: 3.12 (default; configurable via `config.toml` or `FLEXSIM_PYTHON_VERSION`)
- Package manager: `uv` (https://astral.sh)
- FlexSim:
- Real runs require a local FlexSim installation (or repo-local mirror in `FlexSimDev/program`)
- FlexSimPy SDK submodule is provided at `depends/FlexSimPy` (see `.gitmodules`)
---
## Quickstart
Guided setup (recommended)
```/dev/null/shell.sh#L1-11
# From the repo root
python setup_mcp_server.py
# The setup helper will:
# - Summarize your configuration (install locations, Python version, submodule presence)
# - Ensure `uv` is installed and run `uv sync`
# - Build/install FlexSimPy via `utility/build_automation.py`
# - Optionally run a GUI integration sanity test
# - Print next steps to run the MCP server and tests
```
Manual setup
```/dev/null/shell.sh#L1-15
# 1) Install dependencies
uv sync
# 2) Check (and optionally build) FlexSimPy
uv run python utility/build_automation.py --status
# To force a build with a specific Python version:
# uv run python utility/build_automation.py --python-version 3.12
# 3) Launch MCP server in test mode (keeps the process alive; launches FlexSim on first tool call)
uv run mcp_server/flexsim_mcp.py --test-mode
# 4) Run integration tests (spawns the server as a subprocess; stdio MCP client)
uv run python tests/integration/test_mcp_client.py
```
---
## Configuration
Primary configuration lives in `config.toml` (repo root). You can also override at runtime using environment variables.
Example (excerpt from `config.toml`):
```flexsim_mcp/config.toml#L1-36
[flexsim]
install_path = "FlexSimDev/program"
src_path = "C:\\Program Files\\FlexSim 2025 Update 2"
[python]
version = "3.12"
[server]
name = "FlexSimPy MCP Server"
version = "0.1.0"
http_endpoint = "http://127.0.0.1:8088/mcp"
[session]
reuse_policy = "singleton"
[logging]
level = "INFO"
log_file = "flexsim_mcp.log"
[build]
flexsimpy_dir = "depends/FlexSimPy"
auto_build = true
auto_install = true
```
Environment overrides (see `utility/config.py`):
- `FLEXSIM_CONFIG_PATH` – path to an alternate `config.toml`
- `FLEXSIM_INSTALL_PATH` – override FlexSim program path
- `FLEXSIM_PYTHON_VERSION` – override Python version for FlexSimPy
- `FLEXSIM_LOG_LEVEL` – override logging level
---
## Running the MCP server
Two modes:
- Test mode (interactive; no stdio protocol)
```/dev/null/shell.sh#L1-3
uv run mcp_server/flexsim_mcp.py --test-mode
# Keeps running; FlexSim launches on first tool call; intended for manual testing
```
- MCP stdio mode (default when no flags are passed)
```/dev/null/shell.sh#L1-3
uv run mcp_server/flexsim_mcp.py
# Expects MCP JSON-RPC messages via stdin/stdout; used by clients/tests
```
Client configuration (for Claude/Cursor/etc.):
```flexsim_mcp/mcp_server_config.json#L1-14
{
"mcpServers": {
"flexsim": {
"command": "uv",
"args": [
"--directory",
"<repo-flexsim_mcp>/",
"run",
"mcp_server/flexsim_mcp.py"
]
}
}
}
```
---
## Example tool calls (JSON-RPC via MCP)
Open a model
```/dev/null/jsonrpc.txt#L1-15
{"jsonrpc":"2.0","id":1,"method":"tools/call",
"params":{
"name":"flexsim_open_model",
"arguments":{"params":{
"model_path":"C:/path/to/Model.fsm"
}}
}}
```
Run to time and read time
```/dev/null/jsonrpc.txt#L1-30
{"jsonrpc":"2.0","id":2,"method":"tools/call",
"params":{"name":"flexsim_run_to_time","arguments":{"params":{"target_time":300}}}}
{"jsonrpc":"2.0","id":3,"method":"tools/call",
"params":{"name":"flexsim_get_time"}}
```
Evaluate FlexScript
```/dev/null/jsonrpc.txt#L1-16
{"jsonrpc":"2.0","id":4,"method":"tools/call",
"params":{"name":"flexsim_execute_script","arguments":{"params":{
"script":"return getmodeltime();"
}}}}
```
Get/Set node value
```/dev/null/jsonrpc.txt#L1-33
{"jsonrpc":"2.0","id":5,"method":"tools/call",
"params":{"name":"flexsim_get_node_value","arguments":{"params":{
"node_path":"Model/Queue1/stats/input"
}}}}
{"jsonrpc":"2.0","id":6,"method":"tools/call",
"params":{"name":"flexsim_set_node_value","arguments":{"params":{
"node_path":"Model/Processor1/variables/processtime",
"value":5.0
}}}}
```
More FlexScript examples and behavior: see `eval-flexscript.md`.
---
## Tests
Integration tests (stdio client; verifies tools and basic run sequence):
```/dev/null/shell.sh#L1-2
uv run python tests/integration/test_mcp_client.py
# Test 1: open model; Test 2: run for 20s, stop, verify time > 0
```
GUI sanity test (manual, interactive menu):
```/dev/null/shell.sh#L1-2
uv run python utility/integration_test.py
# Launches FlexSim GUI and provides simple interactive actions
```
---
## Repository structure
- `mcp_server/flexsim_mcp.py` — FastMCP server and FlexSimPy controller logic (entry point)
- `utility/`
- `config.py` — configuration loader with env overrides
- `build_automation.py` — FlexSimPy build/install orchestration (MSBuild resolution, .pyd install)
- `copy_flexsim.py` — copies a FlexSim installation into `FlexSimDev/` (uses `flexsim.src_path`)
- `integration_test.py` — GUI-backed sanity test (suppresses CEF stderr, safe cleanup)
- `utility.py` — helpers: suppress stderr, kill FlexSim processes, force exit
- `tests/integration/`
- `test_mcp_client.py` — async stdio client that starts the server and calls tools
- `README.md` — test descriptions and sample outputs
- `depends/FlexSimPy` — FlexSimPy SDK submodule
- `FlexSimDev/program` — optional repo-local mirror of FlexSim install
- `config.toml`, `mcp_server_config.json`, `AGENTS.md`, `eval-flexscript.md`
---
## Batch Experiment Orchestrator (MVP implementado)
Sweep de parámetros con réplicas, métricas y exportación unificada `summary.csv` (y `manifest.json`), para estudios “what-if” a escala.
- Guía: `Batch_Experiment_Orchestrator.md`
- Estado: implementado en `mcp_server/flexsim_mcp.py` (`flexsim_run_experiments`, con `flexsim_experiment_plan` como pre-vuelo de solo lectura; reutiliza `flexsim_get_statistics` / `flexsim_export_results`)
- MVP: `flexsim_run_experiments` con scenarios, seeds, run_to_time, run_speed, statistics, artifact_dir, stop_on_error (requiere ventana FlexSim)
---
## Troubleshooting
- FlexSim not found:
- Update `flexsim.install_path` in `config.toml` or set `FLEXSIM_INSTALL_PATH`
- Optionally use `utility/copy_flexsim.py` to mirror an existing install into `FlexSimDev/`
- FlexSimPy unavailable:
- Check status: `uv run python utility/build_automation.py --status`
- Rebuild for the active Python version and ensure the `.pyd` is installed
- Stdio client issues:
- Ensure you run the server without `--test-mode` when using stdio clients
- Check `mcp_server/flexsim_mcp.log` for details
- GUI noise (CEF) or shutdown issues:
- The utilities suppress CEF stderr and force a safe interpreter exit when needed
---
## Guides (Spanish)
- `GUIA_DESDE_CERO.md` — install and configure everything from scratch, for any user and any MCP client.
- `AGENTE_FLEXSIM.md` — verified instructions for an AI agent (Claude Code, OpenCode...) to build, connect, configure, run and measure models.
- `uv run python utility/doctor.py --live` — checks your environment and tells you how to fix what is missing.
- Two backends: FlexSim launched from Python (evaluation license, runs and measures) and your normal FlexSim window (creates objects); see the guides.
TDQS
Scored across 73 tools
Many tools overlap in reading statistics (flexsim_get_statistics, flexsim_fluid_level, flexsim_get_performance_measure, flexsim_chart_snapshot), creating objects (flexsim_create_objects, flexsim_create_any, flexsim_person_create, flexsim_dashboard_create, flexsim_collector_create, flexsim_fluid_tank), and connecting objects (flexsim_connect_objects, flexsim_attach_resource, flexsim_agv_connect, flexsim_pf_connect). Despite detailed descriptions and cross-references, the boundaries between tools remain fuzzy, forcing agents to parse experimental notes and backend hints to choose correctly.
All tool names follow a consistent flexsim_ prefix and snake_case convention with descriptive terms. Minor deviations exist (e.g., verb-first vs. noun-first ordering, abbreviations like pf and agv), but there is no mixing of casing or chaotic naming styles.
With 73 tools, the set is far beyond the typical 3-15 range and even exceeds the 50+ threshold for an extreme mismatch. Many tools are experimental, overlapping, or superseded, and do not all earn their place, indicating an overgrown surface for the server's purpose.
The surface covers many operations (create, connect, run, read statistics, export results), but notable gaps remain: no general delete tool, incomplete ProcessFlow building and chart binding, no verified experiment launching, and several read-only diagnostic stubs. These gaps will cause agents to work around missing functionality.