MCP for OpenFilters v0.2.2

OpenFilters thin-film design for Claude (Claude Desktop and Claude Code), with the OpenFilters
calculation engine built in. Nothing else to install besides Python.

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.

Changes in this version

0.2.2: the easiest way in: download mcp-for-openfilters-0.2.2.mcpb below and open it, and Claude Desktop
installs the extension (the server comes from PyPI). Also available with one command, uvx mcp-for-openfilters
(see "Install" below). No change to the calculations; the 0.2.0 changes below are what is new since 0.1.x.

New name: MCP for OpenFilters (was openfilters-mcp), so that it is not mistaken for part of OpenFilters
itself; it remains an unofficial connector built on the OpenFilters calculation engine. The package and command
are now mcp-for-openfilters and the server is registered as mcp-for-openfilters. Upgrading: fully quit
Claude Desktop and Claude Code, then run the new installer; it removes the old openfilters registration and
environment. Your data folder (~/OpenFilters-MCP) is kept and still used.

Fixes (please update):

• Optimizations could return broken designs (every thickness "NaN") that were reported as meeting all targets
  with 100 % yield and ranked first in campaigns. They are now detected and never returned.
• Targets like "R <= 0.2 %" counted as met up to 1.2 % (the default tolerance was used as a pass margin); they
  are now met only when they hold.
• Your own material with the same name as a built-in one (e.g. TiO2) was ignored in favour of the built-in
  data.
• Installs from source (computers without a ready-made package) silently used the slow calculation core.
• Optimizations ran a slow solver even with the compiled core installed (now about 3x faster).
• Group delay and GDD were reported in s and s^2 under fs labels, and GD/GDD targets never worked; everything
  is now in fs and fs^2, as in the OpenFilters GUI.
• A mixture index outside the mixture's range crashed the calculation process; it is now refused with the range.
• Saving a project with characters outside Latin-1 (e.g. "λ") failed on Windows.
• The electric field inside graded layers was only sampled at step boundaries (peaks could be missed).
• A polarization angle (e.g. 30°, the fourth choice of OpenFilters' dialogs) was refused by every tool.
• Optimizations with an absorptance (A) target always failed (a bug in OpenFilters itself).
• About 4 % of optimizations with few target points (e.g. a single laser wavelength and three or more layers)
  ended with every thickness "NaN": OpenFilters' compiled solver used uninitialized memory. Fixed in the
  calculation core.
• On macOS, reflectance beyond the critical angle (light from the substrate side at steep angles) was slightly
  wrong.
• The needle method could crawl for its whole time limit, or stop with new layers left at 0 nm and return a
  much worse design (a two-layer V-coat campaign: 10 of 12 starts met the target, now 12 of 12, in seconds).
• Cancelled campaigns now rank and save the starts that finished (duplicates marked, project file written).

New:

• simulate_random_errors: OpenFilters' Preproduction > Simulate random errors (mean, standard deviation and
  worst case of the curves for random thickness errors, no targets needed).
• Spectroscopic (broadband) optical monitoring, and monitoring from the substrate side.
• Thermal model: thermal_sensitivity and thermal_sweep give a peak's shift with temperature
  (dn/dT, expansion, Poisson ratio), kelvin per FWHM and the dn/dT that makes a design athermal.
• Graded (rugate, inhomogeneous) layers: generated profiles (rugate with apodization and several notches,
  quintic and linear ramps, custom profiles) with index matching to the substrate and the medium; supported by
  plots, the electric field, tolerance analysis and layer sensitivity.
• Layers kept fixed in every optimization ("fixed": true, e.g. an electrode).
• Tabulated materials used beyond their data: choose "hold", "openfilters" or "refuse" (table_extrapolation);
  long jobs ask first.
• Materials from the refractiveindex.info database: search it and import any entry, on demand.
• remove_material deletes one of your own materials.
• The needle method keeps improving where OpenFilters' own needle search stops early.
• Ready-made workflows (antireflection, mirror, edge filter, band-pass, manufacturing review) in the "+"
  menu of Claude Desktop and as / commands in Claude Code; OpenFilters example projects to attach.
• Progress updates during long optimizations and yield analyses.
• Peak width (FWHM) and threshold ranges in spectra; a margin option for inequality targets, and warnings
  when a design meets its targets on the grid but not between grid points.

Install

1. Install Python 3.10 or newer (skip if you have it), then open a new terminal:

   • Windows: winget install Python.Python.3.13 (or from python.org)
   • macOS: brew install python (or from python.org; Apple's built-in python3 3.9 is too old)
   • Linux (Debian/Ubuntu): sudo apt install python3 python3-venv

2. Download into one folder:

   • install_mcp_for_openfilters.py
   • the package for your computer. File names say which: cp312 = Python 3.12;
     win_amd64 = Windows, macosx_11_0_arm64 = Mac with Apple Silicon, manylinux = Linux.
     If unsure, download all .whl files: the right one is picked automatically.

3. In a terminal, in that folder, run:

   • Windows: py install_mcp_for_openfilters.py
   • macOS/Linux: python3 install_mcp_for_openfilters.py

   The installer creates a private Python environment, installs the
   package, registers it in Claude Desktop and Claude Code, and runs a self-test.

4. Fully quit Claude Desktop and reopen it. In a new chat: "Run the MCP for OpenFilters server_info tool."

To update an existing installation, fully quit Claude Desktop and Claude Code first (the installer
replaces the environment the running server uses), then run the installer again.

To remove it: py install_mcp_for_openfilters.py --uninstall (python3 on macOS/Linux) (your data folder is kept).

Your materials, projects and results are stored in ~/MCP-for-OpenFilters (or in ~/OpenFilters-MCP if you
used an earlier version). .ofp project files open in
the OpenFilters GUI.

MCP for OpenFilters v0.2.1

OpenFilters thin-film design for Claude (Claude Desktop and Claude Code), with the OpenFilters
calculation engine built in. Nothing else to install besides Python.

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.

Changes in this version

0.2.1: first release on PyPI, so it can also be run with one command, uvx mcp-for-openfilters (see
"Install" below), and contact details for bug reports and collaboration. No change to the calculations; the
0.2.0 changes below are what is new since 0.1.x.

New name: MCP for OpenFilters (was openfilters-mcp), so that it is not mistaken for part of OpenFilters
itself; it remains an unofficial connector built on the OpenFilters calculation engine. The package and command
are now mcp-for-openfilters and the server is registered as mcp-for-openfilters. Upgrading: fully quit
Claude Desktop and Claude Code, then run the new installer; it removes the old openfilters registration and
environment. Your data folder (~/OpenFilters-MCP) is kept and still used.

Fixes (please update):

• Optimizations could return broken designs (every thickness "NaN") that were reported as meeting all targets
  with 100 % yield and ranked first in campaigns. They are now detected and never returned.
• Targets like "R <= 0.2 %" counted as met up to 1.2 % (the default tolerance was used as a pass margin); they
  are now met only when they hold.
• Your own material with the same name as a built-in one (e.g. TiO2) was ignored in favour of the built-in
  data.
• Installs from source (computers without a ready-made package) silently used the slow calculation core.
• Optimizations ran a slow solver even with the compiled core installed (now about 3x faster).
• Group delay and GDD were reported in s and s^2 under fs labels, and GD/GDD targets never worked; everything
  is now in fs and fs^2, as in the OpenFilters GUI.
• A mixture index outside the mixture's range crashed the calculation process; it is now refused with the range.
• Saving a project with characters outside Latin-1 (e.g. "λ") failed on Windows.
• The electric field inside graded layers was only sampled at step boundaries (peaks could be missed).
• A polarization angle (e.g. 30°, the fourth choice of OpenFilters' dialogs) was refused by every tool.
• Optimizations with an absorptance (A) target always failed (a bug in OpenFilters itself).
• About 4 % of optimizations with few target points (e.g. a single laser wavelength and three or more layers)
  ended with every thickness "NaN": OpenFilters' compiled solver used uninitialized memory. Fixed in the
  calculation core.
• On macOS, reflectance beyond the critical angle (light from the substrate side at steep angles) was slightly
  wrong.
• The needle method could crawl for its whole time limit, or stop with new layers left at 0 nm and return a
  much worse design (a two-layer V-coat campaign: 10 of 12 starts met the target, now 12 of 12, in seconds).
• Cancelled campaigns now rank and save the starts that finished (duplicates marked, project file written).

New:

• simulate_random_errors: OpenFilters' Preproduction > Simulate random errors (mean, standard deviation and
  worst case of the curves for random thickness errors, no targets needed).
• Spectroscopic (broadband) optical monitoring, and monitoring from the substrate side.
• Thermal model: thermal_sensitivity and thermal_sweep give a peak's shift with temperature
  (dn/dT, expansion, Poisson ratio), kelvin per FWHM and the dn/dT that makes a design athermal.
• Graded (rugate, inhomogeneous) layers: generated profiles (rugate with apodization and several notches,
  quintic and linear ramps, custom profiles) with index matching to the substrate and the medium; supported by
  plots, the electric field, tolerance analysis and layer sensitivity.
• Layers kept fixed in every optimization ("fixed": true, e.g. an electrode).
• Tabulated materials used beyond their data: choose "hold", "openfilters" or "refuse" (table_extrapolation);
  long jobs ask first.
• Materials from the refractiveindex.info database: search it and import any entry, on demand.
• remove_material deletes one of your own materials.
• The needle method keeps improving where OpenFilters' own needle search stops early.
• Ready-made workflows (antireflection, mirror, edge filter, band-pass, manufacturing review) in the "+"
  menu of Claude Desktop and as / commands in Claude Code; OpenFilters example projects to attach.
• Progress updates during long optimizations and yield analyses.
• Peak width (FWHM) and threshold ranges in spectra; a margin option for inequality targets, and warnings
  when a design meets its targets on the grid but not between grid points.

Install

1. Install Python 3.10 or newer (skip if you have it), then open a new terminal:

   • Windows: winget install Python.Python.3.13 (or from python.org)
   • macOS: brew install python (or from python.org; Apple's built-in python3 3.9 is too old)
   • Linux (Debian/Ubuntu): sudo apt install python3 python3-venv

2. Download into one folder:

   • install_mcp_for_openfilters.py
   • the package for your computer. File names say which: cp312 = Python 3.12;
     win_amd64 = Windows, macosx_11_0_arm64 = Mac with Apple Silicon, manylinux = Linux.
     If unsure, download all .whl files: the right one is picked automatically.

3. In a terminal, in that folder, run:

   • Windows: py install_mcp_for_openfilters.py
   • macOS/Linux: python3 install_mcp_for_openfilters.py

   The installer creates a private Python environment, installs the
   package, registers it in Claude Desktop and Claude Code, and runs a self-test.

4. Fully quit Claude Desktop and reopen it. In a new chat: "Run the MCP for OpenFilters server_info tool."

To update an existing installation, fully quit Claude Desktop and Claude Code first (the installer
replaces the environment the running server uses), then run the installer again.

To remove it: py install_mcp_for_openfilters.py --uninstall (python3 on macOS/Linux) (your data folder is kept).

Your materials, projects and results are stored in ~/MCP-for-OpenFilters (or in ~/OpenFilters-MCP if you
used an earlier version). .ofp project files open in
the OpenFilters GUI.

MCP for OpenFilters v0.2.0

OpenFilters thin-film design for Claude (Claude Desktop and Claude Code), with the OpenFilters
calculation engine built in. Nothing else to install besides Python.

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.

Changes in this version

New name: MCP for OpenFilters (was openfilters-mcp), so that it is not mistaken for part of OpenFilters
itself; it remains an unofficial connector built on the OpenFilters calculation engine. The package and command
are now mcp-for-openfilters and the server is registered as mcp-for-openfilters. Upgrading: fully quit
Claude Desktop and Claude Code, then run the new installer; it removes the old openfilters registration and
environment. Your data folder (~/OpenFilters-MCP) is kept and still used.

Fixes (please update):

• Optimizations could return broken designs (every thickness "NaN") that were reported as meeting all targets
  with 100 % yield and ranked first in campaigns. They are now detected and never returned.
• Targets like "R <= 0.2 %" counted as met up to 1.2 % (the default tolerance was used as a pass margin); they
  are now met only when they hold.
• Your own material with the same name as a built-in one (e.g. TiO2) was ignored in favour of the built-in
  data.
• Installs from source (computers without a ready-made package) silently used the slow calculation core.
• Optimizations ran a slow solver even with the compiled core installed (now about 3x faster).
• Group delay and GDD were reported in s and s^2 under fs labels, and GD/GDD targets never worked; everything
  is now in fs and fs^2, as in the OpenFilters GUI.
• A mixture index outside the mixture's range crashed the calculation process; it is now refused with the range.
• Saving a project with characters outside Latin-1 (e.g. "λ") failed on Windows.
• The electric field inside graded layers was only sampled at step boundaries (peaks could be missed).
• A polarization angle (e.g. 30°, the fourth choice of OpenFilters' dialogs) was refused by every tool.
• Optimizations with an absorptance (A) target always failed (a bug in OpenFilters itself).
• About 4 % of optimizations with few target points (e.g. a single laser wavelength and three or more layers)
  ended with every thickness "NaN": OpenFilters' compiled solver used uninitialized memory. Fixed in the
  calculation core.
• On macOS, reflectance beyond the critical angle (light from the substrate side at steep angles) was slightly
  wrong.
• The needle method could crawl for its whole time limit, or stop with new layers left at 0 nm and return a
  much worse design (a two-layer V-coat campaign: 10 of 12 starts met the target, now 12 of 12, in seconds).
• Cancelled campaigns now rank and save the starts that finished (duplicates marked, project file written).

New:

• simulate_random_errors: OpenFilters' Preproduction > Simulate random errors (mean, standard deviation and
  worst case of the curves for random thickness errors, no targets needed).
• Spectroscopic (broadband) optical monitoring, and monitoring from the substrate side.
• Thermal model: thermal_sensitivity and thermal_sweep give a peak's shift with temperature
  (dn/dT, expansion, Poisson ratio), kelvin per FWHM and the dn/dT that makes a design athermal.
• Graded (rugate, inhomogeneous) layers: generated profiles (rugate with apodization and several notches,
  quintic and linear ramps, custom profiles) with index matching to the substrate and the medium; supported by
  plots, the electric field, tolerance analysis and layer sensitivity.
• Layers kept fixed in every optimization ("fixed": true, e.g. an electrode).
• Tabulated materials used beyond their data: choose "hold", "openfilters" or "refuse" (table_extrapolation);
  long jobs ask first.
• Materials from the refractiveindex.info database: search it and import any entry, on demand.
• remove_material deletes one of your own materials.
• The needle method keeps improving where OpenFilters' own needle search stops early.
• Ready-made workflows (antireflection, mirror, edge filter, band-pass, manufacturing review) in the "+"
  menu of Claude Desktop and as / commands in Claude Code; OpenFilters example projects to attach.
• Progress updates during long optimizations and yield analyses.
• Peak width (FWHM) and threshold ranges in spectra; a margin option for inequality targets, and warnings
  when a design meets its targets on the grid but not between grid points.

Install

1. Install Python 3.10 or newer (skip if you have it), then open a new terminal:

   • Windows: winget install Python.Python.3.13 (or from python.org)
   • macOS: brew install python (or from python.org; Apple's built-in python3 3.9 is too old)
   • Linux (Debian/Ubuntu): sudo apt install python3 python3-venv

2. Download into one folder:

   • install_mcp_for_openfilters.py
   • the package for your computer. File names say which: cp312 = Python 3.12;
     win_amd64 = Windows, macosx_11_0_arm64 = Mac with Apple Silicon, manylinux = Linux.
     If unsure, download all .whl files: the right one is picked automatically.

3. In a terminal, in that folder, run:

   • Windows: py install_mcp_for_openfilters.py
   • macOS/Linux: python3 install_mcp_for_openfilters.py

   The installer creates a private Python environment, installs the
   package, registers it in Claude Desktop and Claude Code, and runs a self-test.

4. Fully quit Claude Desktop and reopen it. In a new chat: "Run the MCP for OpenFilters server_info tool."

To update an existing installation, fully quit Claude Desktop and Claude Code first (the installer
replaces the environment the running server uses), then run the installer again.

To remove it: py install_mcp_for_openfilters.py --uninstall (python3 on macOS/Linux) (your data folder is kept).

Your materials, projects and results are stored in ~/MCP-for-OpenFilters (or in ~/OpenFilters-MCP if you
used an earlier version). .ofp project files open in
the OpenFilters GUI.