Loading…
Loading…
PaperTrail is a standalone evidence-verification app. Add it to Anthropic's Claude Science workbench as an MCP connector and a set of skills, and a scientist can verify an efficacy claim, pool a meta-analysis, screen a variant, or match a patient to trials — in plain language, without ever leaving the environment they already work in. Every answer comes back with a deterministic, grounded, citation-traced provenance trail.
48 tools · 8 skills · one connector
Pick one. The hosted Remote URL is fastest — nothing to install.
In Claude Science: Connectors → Add connector → Remote URL, then paste the hosted MCP endpoint. No install — all 48 public tools appear instantly.
https://papertrail-topaz-phi.vercel.app/api/mcpConnectors → Add connector → Local command. Build once (cd mcp && npm install && npm run build), then add this config. Runs the stdio MCP server on your machine (also exposes the two API-key tools).
{
"mcpServers": {
"papertrail": {
"command": "node",
"args": ["/absolute/path/to/papertrail/mcp/dist/server.js"],
"env": {
"PAPERTRAIL_BASE_URL": "https://papertrail-topaz-phi.vercel.app"
}
}
}
}Skills → Add skill → Import from GitHub (point at the repo), or Upload a skill with any SKILL.md. The 8 skills route Claude Science to the right PaperTrail tool automatically.
github.com/caelum0x/papertrail → skills/ folderNo connector needed for the plain-curl path: Network → Allowed domains → Add domainso Claude's own code can call the API directly during an analysis.
papertrail-topaz-phi.vercel.appEvery tool wraps a live, rate-limited endpoint returning a { success, data, error } envelope — the connector is optional.
curl -s https://papertrail-topaz-phi.vercel.app/api/verify \
-H 'content-type: application/json' \
-d '{"claim":"Drug X reduced major cardiovascular events by 30%"}'48 PaperTrail capabilities, callable by name from inside Claude Science.
Trace a claim to its primary source; recompute the number; ground every span.
Pool studies with no LLM in the numeric path — oracle-tested biostatistics.
Deterministic verdicts over open bio-data: FAERS, GWAS, Open Targets, ChEMBL, ClinVar, PharmGKB.
Grounded QA, deep research, mechanism assembly, knowledge graph — every claim cited.
The named-user tools — structure a bench notebook, match a patient to trials.
Drop these into Claude Science → Capabilities → Skills so the workbench reaches for PaperTrail on the right task, automatically.
Risk ratios, NNT, I², τ², GRADE — recomputed deterministically and oracle-tested against reference tools. The workbench gets a number it can defend.
Every flagged or cited span is a verbatim substring of the source. A span the engine can't locate is dropped — it structurally cannot make an unsourced claim.
When retrieval finds no confident match it returns no_support_foundrather than a confident guess. A wrong "confident" answer is worse than an honest "couldn't verify."