Our Mission
AI Toolkit
Overview DAG Studio SPIFD2 Workspace TARGET Checklist RWE MCP Registry AI Fluency
Reading List
This Week's Picks Archive How It Works
Blog
About
Contact
Open source · Apache-2.0 Official MCP Registry pip install target-mcp

Transparent Reporting for Target Trial Emulations

TARGET MCP

An MCP server that operationalizes the TARGET reporting guideline, so an AI agent can review whether an observational target-trial-emulation study reports what the checklist requires, and return a provenance-stamped result.

View on GitHub What it does
// illustrative example — a completed checklist Illustrative example of a completed TARGET reporting checklist rendered by TARGET MCP: the title, the assessed manuscript with its citation, the provenance line (spec target-0.1.0, scaffold mode, model claude-fable-5), and a completeness bar reading 26 reported, 11 partial, 2 not reported across 39 leaves. Illustrative example · a completed checklist rendered by TARGET MCP

In Brief

The reporting checklist, turned into a reviewable instrument.

TARGET, the reporting guideline for observational studies that emulate a target trial, becomes an MCP server: a connected AI agent reads a manuscript against the encoded checklist and hands back a completed, verifiable assessment.

// how a review runs

Orchestrator

LLM agent, in the loop

Drives the run, dispatches the scoring subagents, and owns every call to the server.

MCP
TARGET MCP server · 11 tools
Ingestparse_manuscript · parse_pmcid
Assessassess_manuscript · get_checklist
Finalizesubmit_scaffold_verdicts
Renderrender_checklist · _html · _docx
Corpus & validationaggregate_corpus · build_coding_sheet · validate_against_gold

Deterministic. Runs one scaffold pass on its own; the fan-out is a client-side accelerator on top.

Every call to the server is made by the orchestrator — the subagents are pure scorers with no MCP access.

01 Ingest the paper via MCP

parse_manuscript / parse_pmcid pulls the manuscript and its supplement into a source-tagged SectionMap and returns the text hash that keys the run.

02 Get the scaffold via MCP

assess_manuscript (scaffold mode) returns the rubric, the manuscript exactly as ingested, and the ordered 39 leaf ids. The server hands back the prompt to run — it does no scoring itself.

03 Fan out to parallel scorers client-side skill

The 39 leaves are split into ~11 batches — each target-trial specification leaf kept with its emulation pair — and dispatched to concurrent subagents on a pinned model. Each scores only its leaves, from the rubric and text; none touch the server. A ~20-minute pass collapses to about two.

1a·1b·1c 2–5 6a/7a · 6b/7b 6g/7g.i/7g.ii … ×11 batches · model pinned (e.g. fable)
04 Merge & verify orchestrator

The orchestrator normalizes ids, checks that all 39 leaves are present exactly once, re-dispatches any leaf a batch dropped, and orders the verdicts. The first verification layer, client-side.

05 Submit & validate via MCP server-side

submit_scaffold_verdicts re-checks coverage and verdict vocabulary and resolves every verbatim quote to a character offset in the source — an unfound quote is rejected. The second verification layer, on the server.

06 Provenance-stamped checklist artifact of record

The finished checklist falls out inline, carrying a stamp bound to the ingested text and the scoring prompt, so drift is detectable at the point of use. Rendered as HTML, Word, or Markdown.

Open source · Apache-2.0 pip install target-mcp com.blackswancausallabs/target-mcp in the MCP Registry Independent implementation, not endorsed by the guideline authors · Cashin et al. (2025): JAMA · BMJ
View on GitHub → Find it in the MCP Registry