mofstructure-mcp
An MCP server for deconstructing and analyzing Metal-Organic Framework (MOF) structures, including SBU and ligand identification, guest removal, and COF stacking analysis. It enables AI agents to perform automated chemical informatics and structural characterization on crystal structures via CIF files.
README
mofstructure — MCP Edition
A lean, MCP-enabled repackaging of the outstanding mofstructure library by Dr. Dinga Wonanke.
Standing on the shoulders of giants
Let's be blunt: the original mofstructure is one of the most impressive pieces of
software in computational materials science written by a single researcher.
In a field where "framework analysis" usually means running a black-box binary and
hoping the output makes sense, Dr. Wonanke built something genuinely different.
The deconstruction engine inside mofdeconstructor.py — over 2 000 lines of
carefully reasoned graph theory — can take any arbitrary MOF crystal structure and
correctly identify every secondary building unit, every organic ligand, every metal
cluster, the SBU topology, the coordination number of every metal centre, and the
points of extension — all automatically, all in pure Python, without ever needing
a pre-curated database of known topologies to look things up in.
That is a hard problem. Most tools either require you to label things manually, rely
on heuristic pattern matching that breaks on anything non-standard, or are proprietary
closed-source binaries. mofstructure does it right: it constructs the full molecular
graph, identifies metal nodes, finds the organic bridges, locates the breaking points,
and assembles the building units — then hands you back proper ASE Atoms objects with
SMILES, InChI and InChIKey already computed via OpenBabel.
The guest-removal logic is similarly thoughtful. Rather than a naive "delete small molecules" heuristic it analyses graph connectivity to find genuinely unbound components, which means it works correctly even on unusual solvates and disordered structures that trip up simpler approaches.
And the COF stacking analysis, the PBC wrapping, the IUPAC name lookup database — all of it is the kind of infrastructure that takes years of domain knowledge to build correctly.
If you are doing any serious high-throughput MOF screening or computational reticular chemistry, reading the original codebase is essential. This fork would not exist without its extraordinary foundation.
Original repository: https://github.com/bafgreat/mofstructure
Original author: Dr. Dinga Wonanke — https://www.dingawonanke.com
Original paper: Wonanke et al., Journal of Open Source Software, 2024.
What is this fork?
This edition strips the original down to a focused MCP (Model Context Protocol) server so that AI agents — in particular those built with featherflow — can call MOF structure operations as tools over stdio or HTTP.
What was removed and why
| Removed | Reason |
|---|---|
Porosity calculation (porosity.py) |
Provided by the separate zeopp-backend MCP service |
Open metal site detection (get_oms) |
Provided by mofchecker-mcp and zeopp-backend |
obsolate/ directory |
Dead code, superseded by current implementation |
| Sphinx docs | Out of date; not needed for MCP deployment |
What was added
| Added | Description |
|---|---|
mofstructure/mcp_server.py |
FastMCP server exposing 5 tools over stdio |
mcp dependency |
MCP Python SDK |
mofstructure_mcp CLI entry point |
python -m mofstructure.mcp_server |
The three-service architecture
This fork is designed to work alongside two companion MCP services:
mofchecker-mcp → "Is this structure valid?" (quality checks)
mofstructure-mcp → "What is this structure made of?" (deconstruction + cleanup)
zeopp-backend → "How porous is this structure?" (geometric analysis)
Point your agent at all three and it can perform end-to-end MOF analysis from a raw CIF file.
Installation
Important — each MCP server should run in its own virtual environment.
git clone https://github.com/lichman0405/mofstructure.git
cd mofstructure
python -m venv .venv
# Linux / macOS
source .venv/bin/activate
# Windows PowerShell
.\.venv\Scripts\Activate.ps1
pip install -e .
Optional RDKit support:
pip install -e ".[rdkit]"
MCP tools
The server exposes five tools. Every tool accepts either a cif_path (absolute path
on the server filesystem) or cif_content (raw CIF text).
| Tool | Description |
|---|---|
remove_guest |
Remove unbound guest molecules; returns clean host CIF |
get_sbu |
Deconstruct into metal + organic SBUs with SMILES / InChI / InChIKey |
get_ligands |
Deconstruct into metal clusters + organic ligands with cheminformatics |
wrap_structure |
Wrap atoms into unit cell — fixes the "broken bond" PBC visualisation artefact |
analyze_cof_stacking |
Compute interlayer lateral offsets and heights; classify AA / AB / mixed stacking |
Running the server
stdio (for featherflow / Claude Desktop)
python -m mofstructure.mcp_server
# or, after pip install:
mofstructure_mcp
featherflow integration
Edit ~/.featherflow/config.json:
{
"tools": {
"mcpServers": {
"mofstructure": {
"command": "/path/to/mofstructure/.venv/bin/python",
"args": ["-m", "mofstructure.mcp_server"],
"toolTimeout": 120
},
"mofchecker": {
"command": "/path/to/mofchecker/.venv/bin/python",
"args": ["-m", "mofchecker.mcp_server"],
"toolTimeout": 120
},
"zeopp": {
"url": "http://localhost:9877/mcp",
"toolTimeout": 120
}
}
}
}
On Windows replace the path with C:/path/to/mofstructure/.venv/Scripts/python.exe.
Using as a Python library
The library API is unchanged from the original:
from mofstructure import structure
mof = structure.MOFstructure(filename="path/to/your.cif")
# Remove guest molecules
clean = mof.remove_guest()
# Deconstruct into SBUs (cheminfo=True computes SMILES / InChI / InChIKey)
metal_sbus, organic_sbus = mof.get_sbu(cheminfo=True)
for sbu in metal_sbus:
print(sbu.info['sbu_type'], sbu.info['smi'])
# Deconstruct into ligands and metal clusters
metal_clusters, organic_ligands = mof.get_ligands(cheminfo=True)
for ligand in organic_ligands:
print(ligand.info['inchikey'])
Command-line tools
| Command | Description |
|---|---|
mofstructure <cif_file> [output_dir] |
Deconstruct a single CIF into SBUs and ligands |
mofstructure_database <cif_folder> [output_dir] |
Batch-process a folder of CIFs into a database |
mofstructure_building_units <cif_folder> |
Batch deconstruction only |
mofstructure_curate <cif_folder> |
Curate / validate a folder of CIFs |
mofstructure_mcp |
Start the MCP server (stdio) |
License
MIT — same as the original project.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。