Japanfold MCP Server

JapanFold skill

Fold proteins, co-fold with ligands (+ binding affinity), design binders, and embed sequences from your AI agent, via the JapanFold API. Get a key at japanfold.aiand.com/account/token and set JAPANFOLD_API_KEY. No local GPU. Runs Boltz-2 / ESMFold-2 / OpenFold3 / OpenBind-0 / RoseTTAFold3 / OpenDDE for structures, AF2-IG for scoring a binder you already designed, BoltzGen / RFdiffusion3 / PXDesign / RFD4-Proteina for design, and ESMC / SaProt for embeddings, on Tenstorrent.

Agents can call the API directly, no install needed:

export JAPANFOLD_API_KEY=jf_live_...          # from japanfold.aiand.com/account/token
curl -s https://api.japanfold.aiand.com/v1/predictions \
  -H "Authorization: Bearer $JAPANFOLD_API_KEY" -H 'Content-Type: application/json' \
  -d '{"model":"esmfold2","sequence":"MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQ"}'

Base URL https://api.japanfold.aiand.com, key in Authorization: Bearer. GET /v1/models lists the model ids to send (boltz2, esmfold2, openfold3, ...) with their size limits. The agent-readable summary is llms.txt.

It's a single SKILL.md built on the open Agent Skills standard, so it installs into any compatible harness with one command.

Install

One line installs it everywhere the open standard is supported: Claude Code, Cursor, Codex, Gemini CLI, Cline, Windsurf, Copilot, Amp and the rest:

npx skills add japanfold/japanfold          # this project
npx skills add japanfold/japanfold -g       # global: every project / new chat

• Target specific agents: -a claude-code, -a cursor, -a codex, -a '*' (all).
• Prefer not to use the installer? It is one file. Download it unchanged into your agent's skills directory, for Claude Code: curl -fsSL --create-dirs -o ~/.claude/skills/japanfold/SKILL.md https://japanfold.aiand.com/skill.md. A web-fetch tool returns only a summary of the page, so an agent should use curl too.

Claude Science manages skills in-app (no installer): Customize → Skills, add from this repo (or paste SKILL.md) and publish it. Or skip install entirely: the API is public and self-describing, so just ask: "use the JapanFold API at api.japanfold.aiand.com to fold …".

Use

Once installed, ask your agent in plain language:

"Fold this sequence with Boltz-2 and report the confidence: MKTAYIAK…" "Design 10 nanobody binders against this target."

Or invoke it explicitly where supported: /japanfold.

MCP

The API is also a remote MCP server at https://api.japanfold.aiand.com/mcp, listed in the MCP Registry as io.github.japanfold/japanfold. claude.ai, Claude Desktop and ChatGPT add it as a connector and sign in, no key needed. Claude Code: claude mcp add --transport http japanfold https://api.japanfold.aiand.com/mcp --header "Authorization: Bearer $JAPANFOLD_API_KEY".

To publish a new registry version, bump version in server.json and run mcp-publisher login github && mcp-publisher publish.

The API

https://api.japanfold.aiand.com: async (submit → poll → download), a key on every call. Full contract at /v1/openapi.json. See SKILL.md for endpoints, examples, and limits.

Fast mode is off unless the agent sends "params": {"fast": true}. It gives higher throughput and may be slightly less accurate. The workbench turns it on for people clicking through; the API and this skill leave it off so a scripted run gets the accurate path. ESMFold-2 always runs fast on this hardware, and its job records fast: true.