Foam-Agent
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Foam-AgentSimulate lid-driven cavity flow at Re=1000"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Foam-Agent
Foam-Agent automates the entire Foundation OpenFOAM v10 CFD workflow — meshing, case setup, execution, error correction, post-processing — from a single natural language prompt. This fork of csml-rpi/Foam-Agent restructures it around a "brain out, hands in" architecture:
Your agent harness Claude Code / Cursor / Codex / OpenCode / pi
(the BRAIN — its model guided by portable skills + subagents in agents/
does the CFD reasoning)
│ MCP (HTTP)
▼
foamagent-mcp in Docker 15 mechanical tools, ZERO API keys:
(the HANDS) RAG over v10 tutorials (local embeddings),
case file I/O, OpenFOAM execution, error
extraction, GMSH/PyVista scripts, SLURMThe intelligence comes from the AI subscription you already pay for. The container only needs OpenFOAM, the tutorial database, and local embeddings.
Features
Capability | |
🗣️ | Prompt → simulation — describe any CFD problem in plain language; the agent plans, writes all case files, runs, and reports |
📚 | Tutorial RAG — semantic retrieval over all Foundation v10 tutorials with local embeddings (key-free) |
🔁 | Auto error correction — failed runs are diagnosed and fixed in a dedicated debug loop until the case converges |
🕸️ | GMSH meshing — geometry described in words becomes a validated |
🖼️ | PyVista post-processing — headless field rendering to PNG |
🖥️ | HPC/SLURM — job submission and polling for cluster runs |
🔌 | 5 CLIs, one repo — MCP registration and skills committed for Claude Code, Cursor, Codex, OpenCode, and pi |
🔒 | Key-free server — the Docker container needs no LLM provider; your harness brings the model |
🧾 | Update contract — |
Related MCP server: COMSOL MCP Server
Quick start
1. Clone and start the server (Docker required; FAISS indices are baked into the image):
git clone https://github.com/KasperHonore/Foam-Agent.git
cd Foam-Agent
docker pull ghcr.io/kasperhonore/foamagent:latest
docker tag ghcr.io/kasperhonore/foamagent:latest foamagent:latest
docker run -d --name foamagent-mcp --restart unless-stopped -p 7860:7860 \
-v "$(pwd)/runs:/home/openfoam/Foam-Agent/runs" \
foamagent:latest python -m src.mcp.fastmcp_server --transport http --host 0.0.0.0 --port 78602. Open the repo in your AI CLI and let it finish the setup — no manual file editing:
claude # or cursor / codex / opencode / pi> onboard meThe foam-onboard skill health-checks the server, warms the embedding model (one-time ~1.2 GB download), offers a demo simulation, and gives you a tour. MCP registration is already committed for every supported CLI, so the tools are wired the moment you open the repo.
3. Simulate:
/foam Simulate lid-driven cavity flow at Re=1000Prefer to verify things yourself first? python scripts/doctor.py runs the same health checks without an agent and prints exact fix commands.
git lfs pull # FAISS indices ship via LFS
docker build -f docker/Dockerfile -t foamagent:latest . # first build: ~10 min, ~10 GBUsage
Where slash commands exist (Claude Code), use them; everywhere else, just say it in plain language — the skills trigger either way.
Command | Plain-language equivalent | What happens |
| "get me set up" / "onboard me" | Guided first-run: health check → warm-up → demo → tour |
| "simulate flow over a cylinder at Re=40" | Full pipeline: plan → generate case → run → debug loop → visualize |
| "the foam server isn't responding" | Doctor: diagnoses Docker/image/container/LFS and brings the server up |
— | "mesh a 2D channel with a cylinder using gmsh" | foam-mesher subagent: GMSH → gmshToFoam → checkMesh |
— | "plot the velocity field of the last run" | foam-visualizer subagent: headless PyVista → PNG |
How it works
"Simulate dam break with two fluids"
│
▼
PLAN find_similar_case → closest v10 tutorial as reference
│
▼
GENERATE write_case_file × N → 0/, system/, constant/, Allrun
│
▼
RUN run_case → success ─────────────► VISUALIZE (PyVista → PNG)
│ ▲
▼ errors │
DEBUG foam-debugger: diagnose → rewrite → rerun (until converged)Every simulation lands in its own directory under runs/, with full logs.
Updating
The clone is your workspace, and updates are designed to never touch your work:
git pull # skills, subagents, MCP configs, server code
docker pull ghcr.io/kasperhonore/foamagent:latest # the server image (then recreate the container)
docker tag ghcr.io/kasperhonore/foamagent:latest foamagent:latest
docker rm -f foamagent-mcp && docker run -d --name foamagent-mcp --restart unless-stopped -p 7860:7860 \
-v "$(pwd)/runs:/home/openfoam/Foam-Agent/runs" \
foamagent:latest python -m src.mcp.fastmcp_server --transport http --host 0.0.0.0 --port 7860The contract: git pull never touches your simulations (runs/, output/), your prompts and meshes at the repo root (user_requirement.txt, user_req_*.txt, *.msh), or your local agent settings (CLAUDE.md, .claude/settings.local.json, .claude/memory/) — they are all gitignored. runs/ is bind-mounted into the container, so simulation results live in your clone and survive container recreation too. Skills and their matching server version update together in lockstep.
Project structure
Foam-Agent/
├── agents/ # CANONICAL skills + subagents (edit here)
│ ├── skills/ # foam, foam-setup, foam-onboard
│ └── subagents/ # foam-debugger, foam-mesher, foam-visualizer
├── .claude/ .cursor/ .codex/ .opencode/ .pi/ # generated per-CLI copies + MCP configs
├── src/mcp/ # the FastMCP server (the "hands")
├── src/ # mechanics.py (mechanical layer) + ESI translation
├── database/faiss/ # pre-built tutorial indices (git-lfs)
├── docker/ # server image
├── examples/ # sample prompts + meshes (copy to root to use)
├── scripts/ # doctor.py, sync_agent_assets.py, ...
├── tests/ # key-free unit tests + e2e vs a running server
└── runs/ # YOUR simulations (gitignored)Skills and subagents
Canonical definitions live in agents/ and are fanned out to every tool's native location (.claude/, .cursor/, .codex/, .opencode/, .pi/) by python scripts/sync_agent_assets.py — edit the canonical files, never the generated copies.
Asset | Role |
| End-to-end orchestration: plan → generate case → run → debug loop → visualize, with reference docs on v10 conventions, file generation, multiphase/VOF, Allrun rules, error playbook, SLURM |
| Guided first-run: health check → warm-up → demo simulation → tour |
| Preflight/doctor for the server |
| Owns the diagnose → rewrite → rerun loop |
| GMSH mesh generation → gmshToFoam → checkMesh |
| Headless PyVista rendering |
Validated by autonomous end-to-end shakedowns: steady simpleFoam (backward-facing step, Re=800), transient multiphase interFoam (dam break), and a GMSH-meshed cylinder at Re=40 — all key-free, all physically verified.
MCP tools
Tool | What it does |
| Valid case domains/categories/solvers |
| Semantic search over v10 tutorials, Allrun scripts, command help |
| Best-matching tutorial + directory structure + Allrun references |
| Where a new case lives (under |
| Case file I/O on the server's filesystem |
| Execute Allrun, extract errors from logs |
| One-off utilities: |
| Server-side Python (PyVista, GMSH) with stdout capture |
| Visualization marker; patch names/types |
| Rules-based Foundation→ESI translation (best-effort) |
| HPC job submission and polling |
See src/mcp/README.md for details and local (non-Docker) installation.
Configuration
Environment variable | Purpose | Default |
| OpenFOAM v10 install (execution tools) | set by the Docker image |
|
|
|
| Embedding model for retrieval |
|
|
|
|
No LLM API keys are needed for the server or the skills.
Sample prompts and meshes
Sample prompts and meshes live in examples/. To write your own, copy one to the repo root and edit it there — root-level user_requirement.txt, user_req_*.txt and *.msh are gitignored, so updates never touch them.
Benchmarking against the original pipeline
The original self-contained LangGraph pipeline (foambench_main.py, made its own LLM calls, needed API keys) has been removed from main — this fork is key-free end to end. It is preserved at the legacy-pipeline git tag for a future harness-vs-harness-less comparison; check out the tag and follow its README to run it. Upstream's FoamBench evaluation of that pipeline reached 100% on 110 tasks with Claude Opus 4.6 at 25 correction loops.
Development
python -m pytest tests/test_mechanics_unit.py # key-free unit tests + asset drift check
python tests/test_lid_driven_cavity_mcp.py # deterministic e2e vs a running server
python scripts/sync_agent_assets.py # regenerate per-tool skill/agent copies
python scripts/doctor.py # validate the local setup (read-only)AGENTS.md documents the architecture for AI coding agents working on this repo.
Troubleshooting
First stop: python scripts/doctor.py — it checks LFS, Docker, image, container, and the MCP endpoint, and prints exact fix commands. For everything else (API keys, disk space, custom meshes, updating, ESI vs Foundation), see the FAQ.
Problem | Solution |
MCP connection refused | Container not running — run the |
First retrieval call takes minutes | One-time ~1.2 GB embedding model download inside the container — not a hang |
Retrieval errors / index not loaded |
|
| Recreate the container so the entrypoint sources OpenFOAM |
Share your simulation
Ran something cool? Open a "Share your simulation" issue — prompt, solver, and a picture is all it takes. Real-world cases directly shape which skills and error-playbook entries get improved next.
Citation
This fork builds on Foam-Agent by Yue et al. If you use it in research, please cite:
@article{yue2025foam,
title={Foam-Agent: Towards Automated Intelligent CFD Workflows},
author={Yue, Ling and Somasekharan, Nithin and Zhang, Tingwen and Cao, Yadi and Chen, Zhangze and Di, Shimin and Pan, Shaowu},
journal={arXiv preprint arXiv:2505.04997},
year={2025}
}
@article{somasekharan2026cfdllmbench,
title={CFDLLMBench: A Benchmark Suite for Evaluating Large Language Models in Computational Fluid Dynamics},
author={Somasekharan, Nithin and Yue, Ling and Cao, Yadi and Li, Weichao and Emami, Patrick and Bhargav, Pochinapeddi Sai and Acharya, Anurag and Xie, Xingyu and Pan, Shaowu},
journal={Journal of Data-centric Machine Learning Research},
year={2026},
url={https://openreview.net/forum?id=kTcH1MnkjY}
}This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Latest Blog Posts
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/KasperHonore/Foam-Agent'
If you have feedback or need assistance with the MCP directory API, please join our Discord server