CatGo — AI Workbench for Computational Materials Science
One desktop workbench that unifies structure building, input generation, HPC job submission and results analysis for computational chemistry — driven by natural language.
- Source repo
- Hello-QM/catgo-LRG
- Stars
- ★ 204
- Last updated
- 6d ago
- License
- AGPL-3.0
- Primary language
- TypeScript
- FA score
- 69/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformClaude CodeCodex (Partial support)
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A catalysis researcher who needs to fetch a Cu structure from Materials Project, cut a four-layer (111) slab, add 15 Å of vacuum, place O at a hollow site, and generate an ORCA input
- Not a fit if
- Teams expecting the latest 1.5.0+ source in this repo (newer releases ship as installers only)
- Users without their own cluster or licensed codes who want to run VASP/CP2K out of the box
- Users needing only lightweight structure viewing with no backend workflows or HPC
- Source review
- 69/100 · Some gaps
What does this agent do, and when should you use it?
CatGo is an AI workbench for computational materials science and chemistry, open-sourced at Hello-QM/catgo-LRG under AGPL-3.0. It brings an interactive 3D structure editor, the natural-language CatBot assistant, a visual DAG workflow engine, and remote HPC cluster access into a single workspace, built as a Tauri desktop app with a SvelteKit frontend and a FastAPI Python backend. It generates inputs and parses results for VASP, ORCA, CP2K, Quantum ESPRESSO, GPAW, DFTB+, SIESTA and LAMMPS, but bundles no commercial simulation codes — you supply the licensed software and compute. Entry points span Windows/macOS/Linux desktop installers, a static web app, a VS Code extension, iOS/Android mobile builds, the catgo CLI, and an MCP server. Note that open-source releases stop at 1.4.x; versions 1.5.0 and later ship only as installers from the official download center, with source not published in this repository.
CatGo reads dozens of structure and output formats — CIF, POSCAR, OUTCAR, vasprun.xml, CHGCAR, XYZ, LAMMPS dump, Gaussian cube and more — and lets you rotate, measure, and atom-edit crystals, molecules, surfaces and trajectories in its 3D viewer. It fetches structures from Materials Project and other database APIs, builds supercells and Miller-index slabs with vacuum, places adsorbates, and prepares defects, dopants, heterostructures and MOFs. It generates and exports inputs for ORCA, VASP, CP2K, Quantum ESPRESSO, ABACUS, Gaussian and LAMMPS, and connects optimization, single-point, frequency, DOS, NEB and MD steps via Quick Build recipes or a visual DAG. Through SSH it connects to your own cluster to browse remote files, submit scheduler jobs, and follow logs and convergence. CatBot performs structure retrieval, model construction, workflow editing and file analysis in natural language with visible tool calls. Analysis covers DOS/PDOS, bands, COHP/ICOHP, d-band centers, XRD, RDF, charge-density isosurfaces, free-energy diagrams and volcano plots, with export to figures, videos, CSV and 3D models.
- A catalysis researcher who needs to fetch a Cu structure from Materials Project, cut a four-layer (111) slab, add 15 Å of vacuum, place O at a hollow site, and generate an ORCA input
- A DFT user who wants a one-click VASP relaxation → static → DOS workflow for the current structure, with human review of inputs before submission
- A graduate student checking SLURM job status and browsing remote structure files over SSH from an iPhone or iPad
- A developer working in VS Code/Cursor who wants to view and export structures and trajectories directly in the editor
- A team automating file-first computational campaigns through MCP and the CLI with Claude Code, Codex or Gemini
- A researcher organizing OER/HER/ORR/CO₂RR/NRR adsorption energies, ZPE and thermal corrections, and plotting free-energy diagrams and volcano plots
How do you install or deploy this agent?
Fastest trial — open the web app (no install):
https://app.catgo-ucsd.orgFull experience — download installers from the official download center (1.5.0+ are installer-only; open-source releases stop at 1.4.x):
https://dl.catgo-ucsd.org/Windows ships CatGo_<ver>_x64-setup.exe/.msi, macOS (Apple Silicon) .dmg, Linux .deb/.rpm. Alternatively install the pip package, which bundles the full web UI:
bash
pip install catgo
# or: uv tool install catgoVS Code users can search CatGo in the Extensions marketplace; iOS users can join the public TestFlight beta at testflight.apple.com/join/FdHup5Hz. Mainland-China fallback via a PyPI mirror:
bash
pip install catgo -i https://pypi.tuna.tsinghua.edu.cn/simpleBuilding the open-source 1.4.x from source requires Node.js 20+, pnpm, Python 3.11, the Rust toolchain, and wasm-pack.
How do you use this agent?
After the pip install, launch and configure with:
bash
catgo # launch the app
catgo setup # register the MCP server and Claude Code skills
catgo serve
catgo status
catgo --helpDesktop flow: launch CatGo and drop in a structure/output file or fetch one from a database; prepare the model with the structure toolbar, Quick Build, or CatBot; connect your lab's cluster from the HPC panel via SSH; review generated inputs and scheduler settings before approving submission; then monitor the workflow and continue to analysis. Prompts you can send CatBot directly:
text
Fetch Cu from Materials Project, cut a four-layer (111) slab,
add 15 Å of vacuum, and place O at a hollow site.
text
Create a VASP relaxation → static → DOS workflow for the current structure.
Let me review the inputs first; do not submit it yet.
Before your first cluster submission, confirm cluster identity, scheduler, executable or module commands, Python environment, and POTCAR/pseudopotential locations.
What are this agent's strengths and limitations?
- Covers the full chain — structure building → input generation → HPC submission → post-processing — replacing the juggle between a structure builder, terminal, SSH client and plotting scripts
- Complete cross-platform coverage: desktop, web, VS Code extension, iOS (with a native Rust SSH transport), Android, CLI/MCP, with a purpose-built mobile interface rather than a compressed desktop layout
- CatBot tool calls remain visible and workflows include explicit human review gates, keeping researchers responsible for settings and conclusions
- Very broad read/write format support across VASP/CP2K/ORCA/QE/ABACUS/LAMMPS plus built-in catalysis analysis for free-energy diagrams and volcano plots
- Versions 1.5.0+ are closed-source and installer-only; GitHub Releases cover only 1.4.x, so the repo does not publish the latest application
- No commercial codes or pseudopotentials are bundled — VASP/CP2K etc. require your own licences, installations and cluster environment, and cluster configuration must be manually verified before first submission
- HPC paths (VASP, CP2K, xTB, etc.) go through a Python SSH adapter with simpler retry/resume behavior than the local DAG engine; QE, Gaussian and GROMACS are export-only with no executable workflow nodes
- The backend has no built-in authentication layer — a self-deployed web backend must stay on loopback or a controlled private network and must never face the public internet
How does this agent compare with similar options?
CatGo's 3D structure viewer, periodic table, and parts of its core UI originate from MatterViz by Janosh Riebesell and have been substantially reworked; on that foundation CatGo adds the catalysis pipeline, workflow engine, HPC integration, CatBot and a plugin system. Its CatRender is a Rust/WASM port of aligfellow/xyzrender (lineage includes xyz2svg).
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| CatGo — AI Workbench for Computational Materials Science This agent | 69 · Some gaps | Desktop appFree | ★ 204 | 6d ago | TypeScript | Claude Code |
| Metaflow | 77 · Good | Library / SDKFree | ★ 10k | 14d ago | Python | — |
| Luxas Research Colleague | 59 · Major gaps | CLIFree + model costs | ★ 1.2k | 22d ago | TypeScript | Claude API |
| EvoScientist: Self-Evolving AI Scientist | 51 · Major gaps | CLIFree + model costs | ★ 5k | 2d ago | Python | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Positives: README mandates human review of scheduler scripts and inputs before submission, states CatBot tool calls remain visible for review, and shows an explicit human review gate in the workflow figure; CI enforces release-rights verification, signer-certificate SHA-256 pinning, pinned whisper.cpp revisions with checksums, and legal-bundle verification — strong supply-chain hygiene. Source attribution to MatterViz is complete (third-party notices, CITATION, legal bundle), earning full marks for source_attribution. Deductions: least_privilege is 1 because the evidence set contains no code showing CatBot tool permission scoping or sandboxing — it is asserted, not shown; sensitive_data_handling is 1 because SSH password/key handling (mobile Rust transport, desktop sidecar) is not inspectable and the backend is self-described as having no built-in authentication layer; rollback is 1 — only editor undo/redo and retry policies are visible, and the README admits HPC retry/resume is weaker than the local DAG engine; dependency_security loses full marks because [email protected] is a known vulnerable release with no stated mitigation, despite otherwise excellent pinning, patching, and checksummed prebuilt binaries.
Positives: package. version 1.4.14 is consistent with the README's 'open source stops at 1.4.x'; pnpm frozen-lockfile in CI, a pinned nightly WASM toolchain kept identical across workflows, and conda/PyPI-mirror fallbacks. Deductions: failure_messages is 1 — CI script errors are detailed, but end-user-facing failure reporting (job, SSH, convergence failures) has no inspectable code in the evidence set; self_consistency falls short of full marks because the package description 'private fork' sits awkwardly with the public-repository framing, and the README evidence is truncated at the supported-software table.
Positives: full marks for audience_and_scenarios and capability_boundaries — a platform table maps each entry point to its best use, the static Web build explicitly excludes backend/HPC, Read vs Write semantics are separated, non-executable workflow engines and 'skills are guidance, not proof of execution' are called out — unusually honest boundary disclosure. Deductions: trigger_precision is 1 — nothing in the evidence shows how CatBot selects tools or guards against misfires; environment_fit is 2 — cross-platform and static-mode configuration is present, but license-code and cluster-environment fit is only inferable from referenced deploy READMEs.
Positives: full marks for install_notes (pip/uv/mamba, China PyPI mirror, TestFlight, per-platform deploy docs) and known_limitations (voluntarily discloses weaker HPC paths, unimplemented Gaussian/GROMACS/QE engines, no built-in Bader, unauthenticated web backend); full marks for license — complete AGPL-3.0 text, consistent badges and package.. Deductions: versioning_changelog is 1 — no CHANGELOG in evidence, and the 1.5.0 closed-source split means GitHub Releases no longer represent the latest version, breaking the version-tracing path; information_architecture is 2 because the README evidence is truncated and primary docs live off-repo; naming_stability is 2 due to the catgo-LRG / catgo / catgo-lrg mismatch; maintenance_responsibility is 2 — CI and release verification show active maintenance, but the closed fork weakens the community-verifiable maintenance commitment.
Positives: full marks for marginal_value — consolidating structure editing, input generation, DAG workflows, HPC submission, and analysis into one workbench targets a real fragmentation pain point. Deductions: output_usability is 2 — export formats are listed thoroughly, but usability of generated inputs (e.g., ORCA charge/spin validation) is not verifiable in the evidence; cost_benefit is 2 — heavy install footprint (Tauri + Python sidecar + Rust WASM) and a closed-source latest edition fetched from a private download center create a gap between what open-source users get and what is advertised.
Positives: full marks for fact_inference_separation — the README systematically distinguishes Read/Write/executable/guidance-only and disclaims that CatBot is an interface, not a replacement for scientific judgment; cross_source_corroboration gets 2 as package. (version, tests, legal scripts, patches) and CI workflows (release rights, cert pinning, checksums) corroborate README claims. Deductions: claim_traceability is 2 — core claims (CatBot review gates, SSH handling, DAG engine behavior) point to server/ and src/ code not present in this evidence set, so static review can confirm the claims exist but not that implementations match; confidence stays 'low' per protocol.
- Closed-source from 1.5.0: this repository contains only 1.4.x sources; installers come from a download center whose publisher is not registry-verified — verify hashes/signatures before installing.
- The backend is self-described as having no built-in authentication layer; never expose it directly to the public internet — loopback or a controlled private network only.
- [email protected] is a known vulnerable release with no stated mitigation; upgrade or sandbox before processing untrusted spreadsheets.
- CatBot's tool-permission boundaries are not visible in the source evidence; validate file and job-submission capabilities in a limited scope before wiring external agents (MCP/Claude Code).
- No CHANGELOG in the repository, and GitHub Releases no longer represent the latest version — version tracing depends on the external download center.
- HPC retry/resume is weaker than the local DAG engine; add your own checkpointing for long-running jobs.
FAQ
Can CatGo run VASP calculations for me out of the box?
Does using CatGo cost money?
Can I use it without installing the desktop app?
How do external AI agents integrate with CatGo?
catgo setup, which registers the MCP server and campaign skills for Claude Code; Codex, Gemini and other agents can use the same MCP endpoints and skills after manual configuration.