Caliper

Deterministic, version-pinned, unit-typed clinical calculations over MCP.

Caliper is a Rust Model Context Protocol server that exposes a curated set of clinical scores and equations as tools over stdio. It is built for one job and to do it precisely: take unit-typed inputs, apply a version-pinned, cited formula, and return a number with its interpretation band, the rules that fired, and a citation — or a typed error. It never guesses a unit, never substitutes a default, and never offers a diagnosis or treatment.

Calculation only. Not medical advice; not a medical device.

Invariants

Every tool enforces these, every time:

1. Unit-typed inputs. Every physical quantity is { "value": <n>, "unit": "<u>" }. A bare number where a unit is required is an error (UnitRequired); an unrecognized unit is UnknownUnit. Conversions are analyte-aware (creatinine ≠ bilirubin molar mass).
2. No silent defaults. A missing required input returns MissingRequiredInput. Caliper never fills in a value to make a calculation succeed.
3. Versioned + cited. Every result carries the exact formula version (e.g. OPTN-2016, CKD-EPI-2021) and a primary-source citation.
4. Stateless. No persistence, no PHI retention, no global mutable state, no logging of inputs. No network calls at runtime.
5. Calculation only. Every result carries a constant disclaimer. Interpretation bands are descriptive, never directive.

Where Caliper sits in the pipeline

Caliper covers ingest → compute: it validates unit-typed inputs and returns a cited, reproducible number with the rules that were applied. It deliberately stops there. Turning a MELD of 26 or an eGFR of 78 into a decision is the clinician's job, not the tool's.

Free-text and lab-dump text ingestion is supported as an explicitly advisory step (extract_inputs / prepare_score): the scanner proposes unit-typed candidates with provenance and never fabricates a unit. Turning a PDF into text is the host's job — point a document/OCR layer at Caliper's text input and the trust boundary stays pure (no document parsers, no new dependencies). Anything resembling a recommendation remains out of scope (see Roadmap).

Scores

25 scores across 11 domains:

id	name	version	domain
meld-na	MELD-Na	OPTN-2016	hepatology
meld-3	MELD 3.0	OPTN-2023	hepatology
ckd-epi-2021	eGFR (CKD-EPI creatinine, 2021 race-free)	CKD-EPI-2021	renal
cockcroft-gault	Cockcroft-Gault creatinine clearance	Cockcroft-Gault-1976	renal
schwartz-egfr	Pediatric eGFR (bedside Schwartz, 2009)	Schwartz-2009-bedside	renal
cha2ds2-vasc	CHA2DS2-VASc	Lip-2010	cardiology
has-bled	HAS-BLED	Pisters-2010	cardiology
curb-65	CURB-65	Lim-2003	pulmonary
crb-65	CRB-65	Lim-2003	pulmonary
wells-pe	Wells Criteria (PE)	Wells-2000	pulmonary
perc	PERC Rule	Kline-2004	pulmonary
westley-croup	Westley Croup Score	Westley-1978	pulmonary
news2	NEWS2	RCP-2017	acute
mews	MEWS	Subbe-2001	acute
qsofa	qSOFA	Sepsis-3-2016	icu
sirs	SIRS criteria	ACCP-SCCM-1992	icu
sofa	SOFA	1996	icu
gcs	Glasgow Coma Scale	1974	neuro
child-pugh	Child-Pugh	1973	hepatology
fib-4	FIB-4 index	Sterling-2006	hepatology
glasgow-blatchford	Glasgow-Blatchford Bleeding Score	Blatchford-2000	gastroenterology
padua-vte	Padua Prediction Score (VTE)	Barbar-2010	hematology
apgar	APGAR Score	Apgar-1953	neonatology
pews	Pediatric Early Warning Score (Brighton)	Brighton-2005	pediatrics
apls-weight	Pediatric weight estimate (APLS, age-based)	APLS-2011	pediatrics

See docs/SCHEMA.md for the full tool and input contract. Every score's coefficients and thresholds have been audited against their primary sources — see docs/COEFFICIENT_AUDIT.md (25/25 clean).

Tools

tool	purpose
list_scores	list scores, filter by domain / free-text query
score_inputs	the input contract for a score (fields, units, clamps, notes)
compute_score	compute a score → ScoreResult or typed CalcError
convert_units	analyte-aware unit conversion with the conversion basis
solve_for	monotone bisection: find the input value that hits a target score
score_series	compute across a time series with per-point deltas and trends
suggest_scores	suggest candidate scores for a context (does not compute)
extract_inputs	scan free text / lab-dump text into unit-typed candidates with provenance (advisory; never fabricates a unit, does not compute)
prepare_score	assemble inputs (text + explicit) and report readiness against a score's contract (does not compute)

Quickstart

Requires a stable Rust toolchain.

cargo build --release
cargo test            # one worked-example fixture per score, plus unit/ingest/server tests

The server speaks newline-delimited JSON-RPC 2.0 on stdin/stdout. A minimal session:

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{}}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"compute_score","arguments":{"id":"gcs","inputs":{"eye":3,"verbal":4,"motor":5}}}}' \
  | ./target/release/caliper-mcp

More request/response pairs are in examples/.

Transport decision (rmcp vs hand-rolled)

Caliper hand-rolls the JSON-RPC 2.0 stdio loop rather than depending on the official rmcp SDK. At build time rmcp 1.7.0 was the current stable release — stable and ergonomic — but it brings an async/tokio runtime and a large transitive tree. For a calculation-only server whose value is determinism, auditability, and a minimal trust surface, a synchronous request → compute → response loop keeps the whole crate at two dependencies (serde, serde_json), makes tool dispatch genuinely data-driven from the registry, and makes the full initialize → tools/call round-trip a pure, unit-testable function. Protocol version pinned to 2025-06-18.

Register with Claude Desktop (Windows)

Build the release binary, then add Caliper to %APPDATA%\Claude\claude_desktop_config.json (create the file if it does not exist):

{
  "mcpServers": {
    "caliper": {
      "command": "D:\\code\\caliper-mcp\\target\\release\\caliper-mcp.exe",
      "args": []
    }
  }
}

Use the absolute path to the built caliper-mcp.exe, escape backslashes, and restart Claude Desktop. Caliper will appear as a tool provider; try "list the renal scores" or "compute MELD-Na for creatinine 1.9 mg/dL, bilirubin 4 mg/dL, INR 1.5, sodium 130 mmol/L."

Roadmap

Shipped

• Free-text / lab-dump text ingestion (extract_inputs, prepare_score) — advisory, unit-safe, with provenance. Handles prose, multi-line lab dumps, and parenthetical/header units (0.2.0–0.3.0).
• Pediatric / weight-band scores: schwartz-egfr, apgar, westley-croup, pews, apls-weight (0.3.0).

Out of scope by design

• PDF/image parsing inside Caliper. Document → text (including OCR) is the host's job; Caliper ingests the resulting text so it can stay dependency-light (serde only) and keep its trust boundary pure. Point a PDF/OCR layer at extract_inputs.
• Anything beyond score + interpretation band + citation — no treatment, dosing, or guideline directives. (apls-weight estimates an unmeasured body weight; it is a calculation, not a dosing instruction.)
• Age-banded vital-sign auto-grading (e.g. full per-age PEWS/PEWS-variant grids). Caliper takes the assessed component levels directly and pins one cited variant; see docs/COEFFICIENT_AUDIT.md for the variant caveats.

License

Apache-2.0.