How Matterloop works.

From question to recorded experiment

1. Create a project. Record its objective, hypothesis and evidence criteria. Add shared documents.
2. Create an experiment. Name it, add its specific context and documents, and prepare a plan and measurement checklist.
3. Upload a measurement. Select XRD, NMR, IR or Raman and confirm the acquisition metadata. Only uploaded methods appear in the experiment.
4. Inspect the result. Review plots, candidate matches, mismatches, reference sources and domain warnings. Revisit analysis history or record a separate scientific review.
5. Log what happened. Save observations, outcome, conclusion and next step. The log retains its evidence snapshot; logging does not automatically confirm a material.
6. Review the written draft. A separate worker generates a cited report from that snapshot. Export the draft, then use project or experiment chat to explore what to test next.

Supported files and scientific tasks

Measurements use UTF-8 text, CSV or XY, up to 10 MB. Documents accept PDF, DOCX, text, Markdown, CSV and finite JSON, up to 20 MB. PDF extraction requires a text layer; OCR is not included.

XRD

Two-theta degrees and intensity, with wavelength and radiation profile. The oxide library covers twelve reference phases. The separate measured-precursor task screens ten source-labelled compounds. Neither determines an arbitrary structure; the precursor task does not assign a crystallographic polymorph.

NMR

Reported distinct ¹H or ¹³C resonance positions in ppm. Candidate consistency uses the prepared reference catalog. Raw FIDs, integrations, multiplicities and atom assignments are outside this workflow.

IR

Wavenumber in cm⁻¹ and absorbance or explicitly declared transmittance. Sparse fitting uses twenty measured pure-component references. Fit coefficients are not concentrations or mass fractions.

Raman

Raman shift in cm⁻¹ and processed intensity. Supported mineral references use independently measured specimens. Laser wavelength is retained when recorded. Rankings do not establish mixture abundance.

Missing metadata is not supplied by a file conversion or proposed material name. Results retain input hashes, reference/model versions, parameters and warnings. Old analyses are not silently recomputed after a model change.

Chat, hypotheses and next experiments

Claude can discuss an entire project or a selected experiment. Independent conversations have automatic titles you can rename, archive and restore. Scoped read-only tools retrieve documents, measurements, recorded peaks, logs and earlier drafts. Citations and agent activity show the evidence used.

Hypotheses and next-experiment suggestions are proposals. Creating a suggested experiment creates an empty record; it does not execute laboratory work, copy measurements or confirm an interpretation. Provider failures are shown as errors.

Reviews, logs and reports

A scientific review is separate from the model result. Reviews and experiment logs remain append-only. Evidence reviews can record supporting, contradictory or inconclusive evidence against a project criterion, followed by a human proceed, revise or stop decision.

Logging an experiment queues an automatic report draft. Its status and failures remain visible, with an explicit retry action. Drafts retain the evidence captured at logging time; later reviews do not become part of an old snapshot. Reports export as PDF, Markdown or JSON. Deterministic analysis exports use PDF, CSV or JSON. All AI drafts require scientific review.

Accounts, storage and recovery

Production sign-in uses Resend email delivery and Neon identity storage, with no passwords. A single-use link expires after fifteen minutes. Sessions use secure HttpOnly cookies; logout revokes the session. Two short onboarding steps collect your name, role, organization or university, research focus and referral information, with independent and other options.

Research records and original measurement files live on the server. Project context, experiment records and saved conversations are mirrored with revision checks. Conflicts preserve local work rather than silently overwriting a newer version.

Uploaded project/experiment document originals remain in your browser’s IndexedDB. Extracted text and metadata are mirrored; original document bytes are not recoverable on another device in this release. Keep an independent source copy. Source-demo documents have a separate authenticated download fallback.

Archive/restore retains project, experiment and chat history. Permanent account erasure, organization permissions, general collaborative merging and high availability are not released.

Usage allowances

The pilot allows ten AI generations per user per UTC day, twenty globally per day, two chat requests per user per minute and two reports per user per day. One generation runs globally at a time. Failed admitted generations can consume the allowance. The account shows remaining usage.

Scientific processing and downloads do not use the AI allowance. Upload/file limits and server capacity still apply. Email sign-in also has cooldowns, per-address limits and IP throttling. These controls are not a guarantee of a fixed provider bill.

What runs behind the workspace

The browser uses Next.js, React, TypeScript and Plotly. FastAPI validates ownership, source links and metadata. CPU scientific workers use source-pinned reference assets with NumPy, SciPy, PyTorch and pymatgen. PydanticAI connects scoped tools to Claude. Research PostgreSQL and files remain on the single EC2 host; Neon stores identity, sessions and onboarding profiles.

The browser’s same-origin API is available under /api/v1 using the same session and CSRF checks. Project persistence is revisioned. There is no released machine-token or third-party OAuth API.

Read the evidence before relying on a result

Candidate rankings are not calibrated probabilities. Public development data do not establish transfer to every lab, and repeated execution does not create independent accuracy cases. The general XRD boundary benchmark currently has zero coverage. Independent scientist-adjudicated validation and measured time-saving studies remain future work.

Read the dated validation record or inspect the full benchmark dashboard after sign-in.