Write a goal.md for scientific simulations and more!

FermiLink reads a simple markdown file describing your scientific goal,
then autonomously runs multi-job multi-day simulations on your laptop or HPC clusters.

---

Key Features

Package Knowledge Base Manager

Use FermiLink to conveniently load package knowledge base to your local workspace; and then apply any local agent for interactive research.

fermilink install <pkg-id>
cd /path/to/workspace/
fermilink init <pkg-id>
# then try any agent 
codex 

Autonomous Scientific Simulations

Beyond interactive use, FermiLink can also sustain multi-task multi-day simulations on either your laptop or HPC.

fermilink loop goal.md          # long-running simulations on workstations or HPC
fermilink reproduce paper.tex   # reproduction of a full research paper
fermilink research goal.md      # full research-paper-scale, multi-task campaigns

Autonomous Code Optimization (beta)

Apply FermiLink to your existing scientific code with a performance goal. It iteratively modifies hot paths and runs deterministic benchmarks.

At the end of the day, it provides optimized scientific code with improved speed while preserving computational accuracy.

fermilink optimize goal.md      # iterative code with correctness guarantees

---

Quickstart

# 1. Install
pip install fermilink

# 2. Set up an AI agent (pick one)
fermilink agent codex        # or: claude, gemini, opencode

# 3. Install a scientific package knowledge base
fermilink install meep       # 150+ packages available

# 4. (Optional) Configure the default SLURM HPC profile
fermilink hpc

# 5. Write your goal.md (see example above), then run
fermilink loop goal.md

First time? Just run fermilink with no arguments. Then, an interactive setup wizard will walk you through everything.

---

Design principles

The key design principle of FermiLink is the segregation of package knowledge bases and simulation workflows.

---

Supported Workflows

FermiLink supports the following seven major commands for various simulation workflows:

Command	Best for	Duration	Recommended Environment
exec	Quick one-off simulations	Minutes	Laptop / workstation
chat	Interactive conversation with agents	Hours	Laptop / workstation
loop	Iterative jobs with PID/SLURM monitoring	Hours -> days	Workstation / HPC
reproduce	Multi-task, paper-scale reproducation of papers	Days -> weeks	HPC clusters
research	Multi-task, paper-scale research	Days -> weeks	HPC clusters
optimize	Code performance tuning (beta)	Hours	Any
exploop	Iterative experimental jobs with PID monitoring	Hours -> days	Workstation

---

Built-in Scientific Packages

FermiLink ships with 150+ knowledge bases for packages spanning multiple scientific domains. Install any of them with a single command:

fermilink install <package>       # e.g., meep, lammps, pyscf, openfoam
fermilink list                    # see all installed packages

Don't see your package? Compile your own local knowledge base:

fermilink compile /path/to/your/code

Browse the full package list →

---

Interaction surfaces

FermiLink isn't just a CLI tool. Pick the interface that fits you:

• Command line — fermilink exec/loop/reproduce/research/optimize goal.md for headless, scriptable autonomy
• Web UI — fermilink start launches a ChatGPT-style browser interface for interactive sessions
• Telegram bot — fermilink gateway connects to Telegram so you can run and monitor HPC jobs from your phone

---

Documentation

Resource	Link
Full documentation	fermilink.org
Installation guide	fermilink.org/installation
Laptop tutorial	fermilink.org/tutorial_laptop
HPC tutorial	fermilink.org/tutorial_hpc

---

Citation

If FermiLink is useful in your research, please cite:

• Gang Meng†, Andres Felipe Bocanegra Vargas†, Xinwei Ji†, Federico Garcia-Gaitan, Felipe Reyes-Osorio, Jalil Varela-Manjarres, Yafei Ren, Mohammadhasan Dinpajooh, Branislav K. Nikolić, Tao E. Li. FermiLink: A Unified Agent Framework for Multidomain Autonomous Scientific Simulations. arXiv:2604.03460 (2026).

---

Contributing

We welcome contributions, from bug reports to new package knowledge bases. See CONTRIBUTING →

---

License

AGPL-3.0