MCP for OpenFilters

An unofficial connector built on OpenFilters: not affiliated with or endorsed by the OpenFilters authors or Polytechnique Montréal.

Design, simulate and optimize optical thin-film coatings by talking to Claude.

This project is an MCP server around the calculation core of OpenFilters (Stéphane Larouche and Ludvik Martinu, Polytechnique Montréal), ported to Python 3 and run without its GUI. Claude plans the work and reads compact summaries; the server does the computing — in parallel on all CPU cores — and writes full results to disk, including .ofp project files that open in the OpenFilters GUI.

Claude orchestrates; the server computes. One tool call can launch a whole campaign.

Built and tested with Claude. The server follows the open MCP standard, but it has so far been developed and tested only with Claude (Claude Desktop and Claude Code). Other MCP clients may work, but they are untested: tool descriptions, plots and long-running jobs may behave differently there.

What it can do

Example requests to Claude:

• "Design a dichroic on BK7 with R > 99 % at 532 and 1064 nm and T > 90 % at 633 nm for 0–20°, at most 40 layers of TiO2/SiO2. Run 100 needle optimizations from random starts and rank the best 10 by yield at ±1 % thickness error."
• "Add my ALD Al2O3 from this ellipsometry file and sweep the AR coating thickness ±10 %."
• "Find thin-film HfO2 data on refractiveindex.info, import it and redo the mirror with it."
• "Where is the electric-field peak of this mirror at 1064 nm, and what monitoring signal should I expect at 633 nm at the end of each layer?"

Ready-made workflows (prompts) appear as / commands in Claude Code and in the attachment menu of Claude Desktop: design an antireflection coating, a mirror, an edge filter or dichroic, a narrow band-pass filter, and review a design for manufacturing. The design-format guide, the material list and the OpenFilters example projects are available as resources (openfilters://guide, openfilters://materials, openfilters://examples/<name>). Long optimizations and yield analyses report progress to apps that show it.

Installation

Requires Python ≥ 3.10. OpenFilters itself does not need to be installed: its calculation engine is built in (the OpenFilters GUI is only useful to open the .ofp files).

With uv (one command)

With uv installed, the server runs straight from PyPI, with the right package for your computer chosen automatically:

claude mcp add --scope user mcp-for-openfilters -- uvx mcp-for-openfilters

For Claude Desktop, add to claude_desktop_config.json (Settings → Developer → Edit config), then fully quit and reopen Claude Desktop:

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

uvx keeps its own cached environment; uvx mcp-for-openfilters@latest picks up a new version.

With the installer

From the release page, download install_mcp_for_openfilters.py and the package for your computer (or all .whl files) into one folder, then:

python install_mcp_for_openfilters.py

It creates a private environment in ~/.mcp-for-openfilters, installs the package, registers the server in Claude Desktop (backing up its settings) and Claude Code, and runs a self-test. Then fully quit and reopen Claude Desktop. --uninstall removes it; --from-github downloads the packages with the GitHub CLI.

Packages are built for Windows (64-bit), Linux (x86-64) and macOS (Apple Silicon), Python 3.10–3.14, with the compiled fast core included: no compiler needed.

Manually

pip install mcp_for_openfilters-<version>-<python>-<platform>.whl
claude mcp add --scope user mcp-for-openfilters -- mcp-for-openfilters

For Claude Desktop, add to claude_desktop_config.json (Settings → Developer → Edit config):

{"mcpServers": {"mcp-for-openfilters": {"command": "<full path to mcp-for-openfilters>"}}}

Upgrading from openfilters-mcp (before 0.2.0)

The project was called openfilters-mcp until version 0.2.0. Fully quit Claude Desktop and Claude Code, then run the new installer: it removes the old openfilters registration and the old private environment (~/.openfilters-mcp) and registers mcp-for-openfilters. Your data folder ~/OpenFilters-MCP is kept and still used, and the old OPENFILTERS_MCP_* environment variables are still read. After a manual install, pip uninstall openfilters-mcp and replace the openfilters entry in your Claude settings.

From source (development)

git clone https://github.com/BBrunoEsteves/mcp-for-openfilters
cd mcp-for-openfilters
python -m venv .venv
.venv/Scripts/pip install -e ".[dev]"      # .venv/bin/pip on macOS/Linux
python dev/fetch_compiled.py               # fast C++ core compiled by GitHub Actions (or build it locally)

Configuration

The names used before 0.2.0 (OPENFILTERS_MCP_HOME, OPENFILTERS_MCP_WORKERS) still work, and an existing ~/OpenFilters-MCP folder is used when ~/MCP-for-OpenFilters does not exist.

Conventions

• wavelengths and thicknesses in nm, angles in degrees, R/T/A as fractions (0–1);
• complex index N = n − ik (k ≥ 0 absorbs), as in OpenFilters and Macleod;
• layers are listed from the substrate outward;
• the merit function is OpenFilters' χ² = Σ ((target − value)/tolerance)², with satisfied inequality targets contributing zero.

A design:

{"substrate": "FusedSilica", "incident_medium": "void", "center_wavelength": 550,
 "layers": [{"material": "TiO2", "thickness": 58.6}, {"material": "SiO2", "qwot": 1}]}

or with a stack formula: {"formula": {"formula": "(HL)^8 H", "materials": {"H": "TiO2", "L": "SiO2"}}}.

Materials given as n,k tables have no data outside their wavelength range. There the design field "table_extrapolation" decides: "hold" (default: n,k stay at the value at the nearest end of the data), "openfilters" (OpenFilters' own extrapolation, identical to its GUI) or "refuse". Every result computed outside the data says so, with the n,k used and what the other choice would give; optimizations, campaigns and yield analyses do not start until the choice is made (Claude asks you).

A target:

{"kind": "R", "from": 1000, "to": 1100, "step": 5, "value": 0.99, "tolerance": 0.005,
 "inequality": ">=", "angle": 0, "polarization": "unpolarized"}

Validation

The test suite (pytest) checks the ported core against analytic results (quarter-wave stacks) and an independent transfer-matrix implementation (dispersive dielectrics, absorbing metals, 0–60°, s/p/unpolarized; agreement < 1e-9) and textbook closed forms (quarter-wave antireflection and absentee layers, V-coat zero, Brewster angle, quarter-wave stack peak reflectance and stop-band width), round-trips every example project through .ofp files, and runs the refinement and needle optimizations of the OpenFilters examples.

How the port was made

src/openfilters_core is generated from the original OpenFilters 1.1.1 sources by tools/port_from_py2.py, an auditable script (mechanical Python 2 → 3 conversions, a short list of targeted fixes, package-relative imports). The _abeles C++ extension uses the Python 3 C-API port by Emil Kleijn; _moremath is ported by the same script. See src/openfilters_core/PROVENANCE.md.

Privacy

Everything runs on your computer: no data is sent anywhere and there is no telemetry. The only network access is downloading public refractiveindex.info data when you ask for it. See PRIVACY.md.

Citing

If you use this in published work, please cite OpenFilters:

S. Larouche and L. Martinu, "OpenFilters: open-source software for the design, optimization, and synthesis of optical filters," Applied Optics 47, C219–C230 (2008).

Contact

• Bug reports and feature requests: GitHub Issues (please include the output of the server_info tool).
• Collaboration and other questions: Bruno Esteves, Center for MicroElectroMechanical Systems (CMEMS), University of Minho: mcp.for.openfilters@gmail.com.

License

GPL-2.0-or-later, like OpenFilters. OpenFilters is © Stéphane Larouche; this project is not affiliated with or endorsed by its authors or Polytechnique Montréal.