§10 · INSTALL
One command, then about three minutes.
Node 20 or newer, and an MCP client. No account, no API key, no build step.
All three tabs are the same stdio server. The JSON is that one command, written in the shape those two hosts read.
An NCBI key is optional, free, and instant. Without one you get 3 requests per second, and in testing 5 of 10 rapid unkeyed requests came back 429. With one, searches are faster and fail less.
§10.1 · QUICKSTART
Four steps, and none of them is a signup.
Node 20 or newer and an MCP client. There is no account to create, no key to paste, and nothing to compile.
- InstallOne line. npx pulls the package; no global install and nothing to compile.
- Scaffold the vaultCreates ~/health-vault with a README, a gitignore that already excludes your raw files, and empty markdown you can fill in by hand.
- Run the onboard promptThree minutes of questions: sex, birth year, ancestry, active conditions, what you're currently taking. That's enough to start ranking.
- Ask something"Should I take creatine?" comes back as graded results with tier badges, sample sizes, resolving PMIDs, and a line for each study saying how well it matches you.
You can also just write the files yourself. The vault is markdown and CSV on purpose, and hand-editing it is a supported path, not a fallback.
Hosts
- Claude CodeThe command above.
- Claude DesktopAdd the same server block to claude_desktop_config.json by hand, or install the .mcpb bundle and skip the JSON.
- CursorSame stdio server, added through Cursor's MCP settings.
- Any MCP clientIt's a plain stdio server. Nothing about it is Claude-specific.
§10.2 · TOOL SURFACE
Eight tools. Four resources. Three prompts.
Every extra tool costs schema tokens in every tools/list call, before anything is invoked, so the surface is deliberately small. These names and descriptions are the specification's, not a marketing summary.
- evidence_searchSearch primary literature for a health question, grade every result for evidence strength, and re-rank it for how well it applies to you specifically.read-only
- evidence_appraiseGrade one specific paper by PMID, DOI, URL, or citation string. Checks retraction, classifies species and design, scores strength and relevance. Returns NOT_FOUND rather than guessing.read-only
- health_context_readRead the local vault: profile, conditions, interventions, goals, genetics, or a summary of everything.read-only
- health_context_writeAdd or update vault entries: profile fields, conditions, medications and supplements with start and stop dates, goals, genetic variants. Also commits records staged by an import.writes
- health_import_fileParse a health data file into staged records for review: lab report PDF, lab CSV, Apple Health export.xml, or a consumer DNA raw file. Staging only.writes
- labs_timelineThe time series for one or more lab analytes: values, dates, reference ranges, flags, trend, and how each compares to your goals.read-only
- n_of_1_readoutWhether a lab marker changed after you started an intervention. Correlational only. Enumerates confounds and states explicitly what cannot be concluded.read-only
- pgx_reportCPIC-guideline pharmacogenomics for your genotype and current medications, with population allele-frequency context. Guideline-backed pairs only.v2
Disclaimer
Biocite is a literature-retrieval and evidence-grading tool. It is not a medical device, it does not diagnose, treat, or prescribe, and it is not a substitute for a clinician who can examine you. Evidence grades are computed from publication metadata by an automated rubric (cebm-x-1.0) — they are CEBM-inspired, not GRADE, and not a substitute for reading the paper. Personal-relevance scores describe how closely a study population resembles the profile you entered; they are not a prediction about you. Do not start, stop, or change a prescribed medication based on this output.
Returned as a field on every one of the eight tools above.
§11 · QUESTIONS
The ones that actually get asked.
Including the uncomfortable one about what leaves your machine, answered exactly rather than reassuringly.
Is there a hosted version?
No, and there won't be. A hosted vault would mean a breach surface, a compliance conversation, and asking you to trust a policy. Running locally removes the question instead of answering it.
So nothing at all leaves my machine?
Search terms do. Your question gets turned into queries and those go to NCBI, ClinicalTrials.gov, and Crossref, which see the terms and your IP exactly as they would if you searched PubMed yourself. Nothing from your vault is ever in a request. No lab value, no genotype, no condition, no date.
Is this medical advice?
No. It's a literature retrieval and grading tool. It doesn't diagnose, treat, or prescribe, and it isn't a substitute for a clinician who can examine you. Don't start, stop, or change a prescribed medication because of something it returned.
Why won't it recommend a dose?
Because that's prescribing, and it's the line between a research tool and a medical device. It'll show you the dose a trial used and quote the sentence it came from. What you do with that belongs in a conversation with someone who can examine you.
Why isn't the output labeled GRADE?
GRADE rates a body of evidence across five downgrade domains that can't be derived from publication metadata. Automating the parts we can and calling it GRADE would be borrowing credibility. The rubric is CEBM-inspired, versioned cebm-x-1.0, and published in full.
What happens when the evidence is thin?
It returns INSUFFICIENT and shows you what it found anyway, with tiers attached. That verdict is computed from tier counts in code, and it fires often. Most supplement questions genuinely deserve it.
What if a search fails?
It fails loudly. A rate limit or an outage returns an error containing the literal sentence "the literature search failed; this is not evidence of absence." A 429 quietly rendered as zero results is indistinguishable from proof that nothing exists, and that's a dangerous thing to be wrong about.
My ancestry isn't in most papers. Does that break it?
It's the normal case, and the tool says so. Only 1.9% of human RCTs from 2015 to 2026 carry any population indexing. Ancestry adds a bonus when a study reports a match and never filters anything out, because a hard ancestry filter would discard about 98% of the literature. When the data isn't there you get the sentence saying it isn't there.
Do I need an API key?
No. An NCBI key is free and instant and raises your rate limit, which makes searches faster and less likely to fail. Everything works without one.
Does it work with clients other than Claude?
Yes. It's a plain stdio MCP server with eight tools. Nothing in it is host-specific.
Can I show the output to my doctor?
That's what it's built for. Every result carries its PMID, its design, its sample size, and every term that produced its score. A clinician can check the basis for anything it said, which is the entire reason the scoring is exposed instead of hidden.
Is my vault a database I have to learn?
No. It's markdown files with YAML front matter and a couple of CSVs, in one folder. Open it in any editor. Put it in git if you want a history. The tool preserves keys it doesn't recognize and never rewrites your prose.
What does it cost?
The server is free and MIT-licensed, and it always will be. If a paid tier ever exists it'll be for the ingest and grading pipeline that makes messy personal data usable, never for the literature or the rubric.
Who is this for?
People who already read the papers and want the search conditioned on their own biology. If you've ever pasted a lab PDF into a chat window and gotten back a confident paragraph with no citations, this is the fix for that.