Connect Claude to your simulation data

Sign in with your existing Microsoft (Machine Phase Systems) account — no new credentials.

What you can do

Ask

The live queue and its order, job status and history, which methods this deployment has actually run, the starter templates, and the contents of job files.

Search

The job history server-side — by composition (“C12 O”, ranges, exclusions), method, status, free text or date — so “have we already computed this?” gets a real answer instead of a guess.

Read results

Energy, convergence, frontier orbitals and forces, with explicit units, straight from job detail. “Did my optimization converge?” is one question.

Act

Validate and submit jobs, cancel one that has not started yet, pause and resume, reorder the shared queue, edit metadata, delete or archive. Every write asks for your explicit confirmation first, and names the job’s owner when it is not you.

Administer

Admins also get an audited, read-only SQL tool over jobs data — never user records — for questions the other tools cannot express.

Guided workflows

Ask for these by name and Claude follows the whole checklist rather than remembering some of it:

• submit_job — the full pre-submission sequence: check whether it was already computed, start from a template, validate, and stop for your confirmation before anything is queued.
• check_already_computed — a composition-first existence check, including the name-versus-config verification the answer actually needs.
• triage_errors — pulls failed jobs and groups them by what really went wrong, separating config mistakes from infrastructure ones.

What it will not do

Stated plainly, because an unstated limit gets guessed at:

• Stop a job that has already started. There is no stop signal to the worker. Pausing an optimization halts it at the next geometry step, which on a long run can be an hour away; a single-point energy job is not interrupted at all. A job that has not been picked up yet can be cancelled outright.
• Read binary files such as .chk or thumbnails. Claude will point you at the job’s page instead of improvising a substitute.
• See archived jobs. Search covers live history, so “no results” is not proof that something was never run.
• Trust a job name. Names are typed by hand and routinely contradict the config — a job called “S=1” may be running multiplicity 3 — so Claude is told to identify jobs by composition and their .cfg, and to confirm deletions on the reference id rather than the label.

Server URL

Add to Claude

1. Desktop / Claude.ai: Settings → Connectors → Add custom connector, paste the URL above.
2. No Client ID or secret — the connector registers itself.
3. Claude Code: claude mcp add --transport http elementa https://mcp.simulations.mps.inc/mcp
4. Complete the Microsoft sign-in when prompted.

Every answer Claude gives you is stamped with which instance it came from and when it was read, so a result from a developer’s machine can never be mistaken for this one.

Develop & run locally

Developers run a local dev instance (elementa-mcp-dev) against a local API — there is no hosted dev API, and the dev server refuses to point at production. The runbook lives in the web-app repo: ops/mcp-local/README.md (the per-tool dev loop) and docs/deployment.md → “Local MCP dev deployment” (setup, prerequisites, smoke tests, graduating to prod).