API and MCP
Everything the website does, a script or an agent can do too. Same quota, same provenance, same refusals.
Getting a key
Sign in, open your analyses, and use the API keys panel. The key is shown once — it is stored only as a hash and cannot be recovered.
- · A key acts as your account and carries the same allowance. It grants nothing extra.
- · A key cannot create or revoke keys. That needs a signed-in browser, so a leaked key cannot replace itself.
- · Revoking takes effect immediately.
curl -H "Authorization: Bearer $RNAFOLD_API_KEY" \
https://api.rnafold.com/usageEndpoints
| POST | /analysis | Queue an analysis — a sequence, several components, a derived analysis on a parent, or an uploaded structure. |
| GET | /druggability | This account's analyses, most recent first. |
| GET | /druggability/{id} | One analysis: engine agreement, structural checks, novelty, provenance. |
| GET | /analysis/{id}/derived | What can be run next from this result. |
| GET | /analysis/{id}/coordinates | Coordinates, walking up the parent chain to whichever ancestor holds them. |
| GET | /analysis/{id}/structure/{engine} | One engine's structure. |
| GET | /usage | Allowance left, and the per-engine length caps. |
| POST | /groups | Make a set of complexes. A bundle that is refused comes back with its index and reason; scores cannot be supplied. |
| GET | /groups | This account's groups, and the caps for its tier. |
| GET | /groups/{id} | One group: each complex, your properties, and every score with its tool version. |
| GET | /groups/{id}/rank | Order a group by one metric from one tool version, flagging where replicate ranges overlap. |
| GET | /scorers | The tools that can be mapped over a group, and what each accepts. |
| POST | /groups/{id}/score | Run one scorer over a group as a CPU job. Complexes it does not accept come back with the reason; scores are written by the platform with the tool version and a control, never supplied. |
| GET | /pipelines/schema | The shape of a pipeline, with an example that needs you to choose a threshold. |
| POST | /pipelines/validate | Check a pipeline and the groups bound to it. Runs, saves and prices nothing. |
| POST | /pipelines/estimate | Price a run as a ceiling: every filter is assumed to remove nothing. A failed estimate has no numbers and is never a cost of zero. |
| POST | /pipelines/runs | Run a pipeline. Validated again on your groups as they are now, inputs snapshotted by complex id, one CPU job charged. Optional idempotency_key and budget_seconds; a repeated key returns the first run. |
| GET | /pipelines/runs | Your runs, newest first. |
| GET | /pipelines/runs/{id} | The full record: the snapshot, each node's counts, and for every filter the value, threshold and control (as numbers) of each complex it looked at. |
| GET | /pipelines/usage | What your runs count under the decided rule, recorded while metering is off. |
| POST | /pipelines/templates | Save a graph as the next version of a template, or get back the version it already is. Versions never change. Declare which filter fields a run may set; a threshold may be left unset, and then every run must supply it. |
| GET | /pipelines/templates | Your templates, newest change first. |
| GET | /pipelines/templates/{id} | A template and its versions (hash, note, editable fields). |
| GET | /pipelines/templates/{id}/versions/{v} | One version in full. Refused if it no longer matches its stored hash. |
| POST | /pipelines/templates/{id}/default | Choose which version a run uses when it names none (null clears it, so the latest is used). Your own default; it shares nothing. No version is altered. |
| DELETE | /pipelines/templates/{id} | Delete a template and its versions now. Runs that used it keep their own copy of the graph. |
| DELETE | /groups/{id} | Delete now, including any complex no other group still holds. |
| GET | /keys | Your API keys, redacted. Browser session only. |
| POST | /keys | Mint a key. The plaintext is in the response and nowhere else. |
| DELETE | /keys/{id} | Revoke a key. |
Submitting returns a job id; folding takes minutes, so poll rather than block. Every result carries its provenance — engine versions, parameters, what was withheld and why.
MCP server
For agents, with nothing to install: https://mcp.rnafold.com/mcp. Connect from claude.ai by adding it as a custom connector and signing in on rnafold.com, with no key to paste, or from any client that sends Authorization: Bearer rnaf_…. Create a key under API keys on the Analyses page; revoking a key, or a connected app, stops it at once.
claude mcp add --transport http rnafold https://mcp.rnafold.com/mcp --header "Authorization: Bearer $RNAFOLD_API_KEY"
{
"mcpServers": {
"rnafold": {
"url": "https://mcp.rnafold.com/mcp",
"headers": { "Authorization": "Bearer rnaf_..." }
}
}
}Most tools run in seconds on our CPU and cost nothing: sequence and structure utilities, RNA folding and accessibility, CRISPR guides, digest and primer design, vector liability scanning, and probing-data agreement. A fold runs several structure engines and takes minutes, so study_sequence and compare_variants start one and return at once, and wait_for_job collects it. A fold uses one run of your monthly allowance, and an estimate is free. A pipeline is validated, then estimated, then submitted, in that order; submitting needs an idempotency key and a budget so a retry cannot make or charge a second run. A score is a distribution with a tool version, written by the platform and never by you; your own measurements go in properties.
Every tool returns its caveats with the result. A result page can show which engines were withheld and why, whether a close relative sits in the training window, and what the control scored; a tool call has no page, so the payload carries them instead. The full walkthrough, with real examples, is in the guide.
What this does not do
- · No webhooks. Poll for completion.
- · No batch endpoint. Submit one at a time, within your concurrency limit.
- · No client library yet — the API is plain HTTP and JSON.
- · Engine length caps are a property of this deployment's hardware, not of the engines.