agentic-elastemp

An LLM agent for elastemp, the workflow that computes quasi-harmonic, temperature-dependent elastic constants from VASP and phonopy. It is built with LangGraph and uses Claude.

You ask about a project in plain language. The agent can tell you where the project stands, run the next analysis step, explain why a VASP run failed, and summarize the results. It never runs or submits VASP jobs; you still do that.

Install

pip install agentic-elastemp

This also installs elastemp, LangGraph and the Anthropic client. Python 3.10 or newer is required.

API key

The agent needs a Claude API key. It reads ANTHROPIC_API_KEY from the environment or from a .env file in the current or project folder:

cp .env.example .env      # then fill in ANTHROPIC_API_KEY=...

.env is in .gitignore. Never commit a real key; .env.example holds empty placeholders only. To choose the Claude model, set ELASTEMP_AGENT_MODEL in .env or pass --model.

Use

elastemp-agent --workdir runs/TiN                                   # conversation
elastemp-agent --workdir runs/TiN --ask "What should I run next?"   # one question

Type exit (or press Ctrl+D) to leave a conversation.

Example:

you> The VASP jobs finished. Where do we stand?
   [tool] project_status()

agent> The static runs are back: 43 of 44 converged, 1 failed. deformation-3/strain-2 hit NELM
(electronic SCF did not converge). Every deformation still has at least 7 converged points, so
get_elastic can drop that point and refit. You could also rerun it with a larger NELM first.

you> Rerun make_bulk_dynamic.
   [tool] run_step({"step": "make_bulk_dynamic"})

agent> I haven't run it. make_bulk_dynamic would delete 11 Phonon_calculation folders that already
contain finished DFPT runs. Do you want me to go ahead and delete them?

How it works

The agent has four tools. All of them are deterministic code on top of elastemp:

START -> agent --(tool calls)--> tools -> agent
               --(answer)--> verify --(ok)--> END
                                    --(invented numbers)--> agent (one rewrite)

Number check: after each answer, a verify node compares every number in it with the tool results. If the agent states a value that came from nowhere (an "experimental" value from memory, say), it is asked to rewrite the answer once. If it still does, the answer is shown with an [Unverified: ...] note.

Development

pip install -e ".[dev]"
pytest

The tests replace Claude with a scripted model, so they need no API key and make no network calls.

License

MIT. See LICENSE.