API documentation

JapanFold API

Fold proteins, co-fold complexes with ligands (and get binding affinity), design de-novo binders, and compute protein embeddings, over a free, public, keyless HTTP API. Boltz-2, ESMFold-2, Protenix-v2 and OpenDDE for structure prediction, BoltzGen and RFdiffusion3 for binder design, ESMC and SaProt for embeddings. No API key, no local GPU, nothing to install.

Base URL https://api.japanfold.com

Contract /v1/openapi.json OpenAPI 3.1, the source of truth

Fold your first protein API reference Try it in the browser

The model: submit → poll → download

Everything is an async job. Submit work, get back a job id, poll until the status is terminal, then download the results.

POST /v1/predictions   or   POST /v1/designs      →  { "id": "...", "status": "running", ... }
GET  /v1/jobs/{id}                                 →  poll until status is succeeded/failed/canceled
GET  /v1/jobs/{id}/results                          →  scores + a list of downloadable artifacts
GET  /v1/jobs/{id}/archive                          →  everything as one .zip

Statuses: queued → running → succeeded | failed | canceled.

Server artifacts are retained only temporarily. Download what you need and save it locally.

Fold your first protein in 3 calls

BASE=https://api.japanfold.com

# 1. submit: a bare `sequence` is the simplest input
JOB=$(curl -s -X POST $BASE/v1/predictions \
  -H 'Content-Type: application/json' \
  -d '{"model":"boltz2","name":"myprotein","sequence":"MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQ"}' \
  | python3 -c 'import sys,json; print(json.load(sys.stdin)["id"])')

# 2. poll until done. `Prefer: wait=60` blocks up to 60s so it often returns finished.
curl -s -H 'Prefer: wait=60' $BASE/v1/jobs/$JOB

# 3. once status=succeeded: read scores, then download the structures
curl -s $BASE/v1/jobs/$JOB/results
curl -s $BASE/v1/jobs/$JOB/archive -o myprotein.zip && unzip -oq myprotein.zip -d myprotein

That's the whole workflow. Complexes, ligands, affinity, binder design and parameter choice are all variations on these three calls.

Where to go next

• Authentication

  Keyless by default; an optional Bearer key scopes jobs to you instead of your IP.

• Predictions

  Input shapes, models, co-folding, affinity, params.

• Designs

  BoltzGen and RFdiffusion3 binder design.

• Embeddings

  ESMC and SaProt protein embeddings (per-residue + pooled).

• Jobs

  Polling, listing, cancel/delete, results, logs, artifacts, archive.

• Models & limits

  The model list, every parameter, and the caps.

• Accuracy

  Output parity against each model's reference GPU implementation.

• Errors

  The problem+json shape and the status codes.

• Examples

  End-to-end fold, co-fold+affinity and design in curl and Python.

• The JapanFold skill

  Fold and design straight from your AI agent.

Network egress

If your environment sandboxes outbound HTTP, allow the host api.japanfold.com. A 403 citing Cloudflare error 1010 is edge bot-filtering of your HTTP client, not an API error: retry with a browser-like User-Agent.

Powered by on