Connect an agent to rnafold
New here? Start with the guide: a five-minute quickstart and worked recipes. This page is the detail on connecting.
One URL, one key. Nothing to install.
URL: https://mcp.rnafold.com/mcp
Auth: Authorization: Bearer rnaf_...
Create a key at rnafold.com/analyses (API keys), or connect through your client's own
sign-in (below). Connections made earlier to https://api.rnafold.com/mcp keep working; it is the same server under its older name.
A key carries your account's quota and nothing more; revoke it there and it stops working within a minute.
Claude Code
claude mcp add --transport http rnafold https://mcp.rnafold.com/mcp \
--header "Authorization: Bearer $RNAFOLD_API_KEY"
Then ask in plain language: "Does G10A change the structure of this tRNA? ..."
Cursor / any client that takes an mcpServers JSON
{"mcpServers": {"rnafold": {
"url": "https://mcp.rnafold.com/mcp",
"headers": {"Authorization": "Bearer rnaf_..."}}}}
Claude Desktop
Desktop's config takes local programs, so use the mcp-remote bridge:
{"mcpServers": {"rnafold": {"command": "npx", "args": [
"mcp-remote", "https://mcp.rnafold.com/mcp",
"--header", "Authorization: Bearer rnaf_..."]}}}
From code (Anthropic API)
Pass the server in the MCP connector with authorization_token set to your key.
claude.ai (web): connect without a key
In claude.ai, open Settings, then Connectors, then Add custom connector. Enter
https://mcp.rnafold.com/mcp
and press Connect. A page on rnafold.com opens (sign in if asked) that says which app is asking, where your browser will go next, and what the connection can and cannot do. Press Approve. Nothing is connected until you do.
- The connection is an API key that lasts an hour and renews itself while the app stays connected. It appears under API keys on the Analyses page, labelled "(connector)". Revoke it there to disconnect.
- It acts as your account and spends your allowance exactly like a key you made yourself. It cannot create, list or revoke keys.
- Other clients that sign in through a browser use the same flow.
What you can call
About 55 tools: sequence and structure utilities (free, no quota), RNA folding and ensemble sampling on CPU, groups and pipelines, and the GPU fold that compares several structure engines. Five recipes are offered as prompts (variant screen, mRNA panel, probing check, CRISPR target review, running a pipeline).
- Free tools are unmetered but rate limited.
- GPU folds and pipelines count against your monthly allowance exactly as on the
website. Estimates are free: ask
estimate_fold/estimate_pipelinefirst. - To get a fold's coordinates, ask for the structure:
get_structure(job_id, engine)returns the PDB as text (one engine's prediction; the platform names no consensus). A large molecule returns nothing rather than a truncated file; ask fortrace_only(one atom per residue) or a highermax_chars. A 28-nt RNA is about 49 KB of text, so expect it to use a lot of your assistant's context. - Results carry their caveats in the payload, under
caveats. (notesis a deprecated copy of the same list, kept for now; readcaveats.) Tools that only list or reformat things (reverse complement, listing your groups) have none.
Limits
Per key: 120 requests a minute, 10,000 a day, 8 running at once. Over a limit you
get HTTP 429 with a Retry-After header; nothing is queued. The server as a whole
caps concurrent requests and answers 503 with Retry-After when busy.
Your data
Sequences you send to GPU tools run on EU infrastructure and are kept as on the website (14 days by default). The request log records the time, status and a short hash of the key; never a sequence, a body or a key.
Note: an AI client you connect (Claude, Cursor, ...) sends your conversation, including any sequence you type, to its own provider. If a sequence must not leave EU infrastructure, do not paste it into a client hosted elsewhere.
Last changed 2026-11-18