Six open-source Model Context Protocol servers. Five for evidence-grounded work in pharma and healthcare: live literature and drug labels, a privacy boundary for clinical text, and deterministic checks that an answer is actually supported by its source. And one for the water and energy cost of AI itself.
6 servers, all open source
Listed in Anthropic’s MCP Directory
Published to the official MCP Registry
No LLM-as-judge in any check
See it work
One question, one real answer.
Each example below is genuine output from the server, captured in October 2026 and trimmed for length. Pick a server.
You ask
“Summarise this clinic note, but protect the patient’s details first.”
Claude calls
redacta.protect(text: "Patient: Margaret Ellison, DOB 14/03/1952, NHS number 943 476 5919. Seen by Dr Patel… Lives at OX7 5QA, tel 07700 900123…")
The model only ever sees the tokens. The map that reverses them stays inside the server. The clinician’s name is kept, because it isn’t a patient identifier. (Fictional patient.)
Returns
Patient: [PATIENT_NAME_1], DOB [DATE_OF_BIRTH_1], NHS number [NHS_NUMBER_1]. Seen by Dr Patel at the respiratory clinic. Lives at [POSTCODE_1], tel [PHONE_1]. Ongoing COPD; started ensifentrine 3 mg nebulised twice daily.
A built-in worked example, so it runs offline with no API key. Every value is tied to a verbatim sentence from the abstract.
Returns
Differs: assay / method
Howitz 2003 (Nature)
“lowers the Michaelis constant of SIRT1 for both the acetylated substrate and NAD(+)”
Beher 2009 (Chem Biol Drug Des)
“the Fluor de Lys-SIRT1 peptide is an artificial SIRT1 substrate…”
Ruled out: intervention (resveratrol in both). Not reported by either: dose, timing, sample size.
15claims verified against source
7fields “not reported”
0invented
Your agent drafts
“Start at 0.25 mg once weekly, then increase to 0.5 mg after 8 weeks.”
and checks it first
opengate.check_grounding(answer: "…after 8 weeks.", context: "…The recommended starting dose is 0.25 mg once weekly. After 4 weeks the dose should be increased to 0.5 mg once weekly.", expected_facts: ["0.25 mg"])
The context says 4 weeks. The draft says 8. No model is asked for an opinion: the number either traces to the source or it doesn’t.
Returns
✗ NOT GROUNDED
Issues:
• ungrounded number "8" — not in the provided context
You ask
“Check this extracted referral against the gold record.”
In the gold record, null means the document doesn’t state it. The date formats differ but normalise to the same day, so that field passes. The guessed date of birth doesn’t.
Returns
"passed": false
"issues": [
"FABRICATED field \"date_of_birth\": document does not state it (got \"1952-03-14\")"
]
"per_field": {
"patient_name": "correct",
"referral_date": "correct",
"date_of_birth": "fabricated",
"reason": "correct"
}
You ask
“How much water do 30 ChatGPT prompts a day use over a working year?”
Pure arithmetic over a published methodology, so the same question always gets the same answer. The boundary and assumptions come back with the number, so it can be checked rather than repeated.
Returns
"inputs": { "queries_per_day": 30, "usage": "simple", "boundary": "comprehensive", "days": 250 }
"per_query": { "energy_wh": 0.34, "water_ml": 1.39 }
"total": { "energy": "2.6 kWh", "water": "10.5 L" }
"assumptions": [
"Boundary: Onsite cooling plus water consumed generating the electricity (~4.1 mL/Wh…)",
"Chat-style prompts… Agentic sessions can be far higher…",
"… Excludes model training."
]
"methodology_version": "2026-09"
25.5hours of TV, in energy
21500 mL bottles
0models called
The servers
Install one, or all six.
Five run locally over stdio with npx. AI Footprint is hosted: connect by URL, with nothing to install. All six work with Claude Desktop, Claude Code, Cursor and any other MCP client.
Redacta
Privacy
A privacy boundary for clinical text.
Pseudonymises patient identifiers before the model sees them. The reversal map stays inside the server process, so the agent you’re protecting the data from never sees it. Output can be checked for identifiers that reappeared, so the boundary is verified rather than assumed.
Live access to PubMed, Europe PMC, US and UK drug labelling and ClinicalTrials.gov. Fourteen tools, no model in any of them, and every result carries its PMID, NCT ID or DOI. compare_labels pairs US and UK label sections so market differences are visible, not assumed.
search_pubmed
get_full_text
search_trials
get_uspi
get_smpc
compare_labels
+8 more
API key
NoneOptional free NCBI key for higher PubMed limits
Network calls
Public sources onlyPubMed, Europe PMC, DailyMed, emc, ClinicalTrials.gov
Extracts each study’s design, shows which methodological dimensions differ and which are identical (so ruled out), and grounds every value in a verbatim quote from the source. It never picks a winner. Fields it can’t ground come back as not reported, never guessed.
compare_studies
compare_example
list_examples
API key
AnthropicLive comparisons only; built-in examples need none
Let an agent check its own answer before it replies.
One tool, check_grounding, backed by OpenGATE’s deterministic grounding check. Required facts must appear, every number must trace to the context, and unanswerable questions must be declined, not fabricated. Reproducible verdicts with no LLM-as-judge.
Runs Groundwork’s deployment-readiness gate from inside a conversation: check a whole repo, a single answer for grounding, or a single extracted record against its schema, or scaffold the evaluation harness. A strong floor, not a safety certificate, and it says so.
check_readiness
check_answer_grounding
check_extraction
scaffold_harness
cost_summary
API key
NoneFor the checks themselves
Network calls
None of its ownThe readiness gate calls your system via its adapter
Four tools built on the calculator’s published methodology: energy first from 2025–26 vendor disclosures, water derived from it, onsite or comprehensive boundary. Estimate chat use, put any volume of water in context, or estimate an agent session from its token counts. Pure arithmetic, no model in any tool, and tests keep it identical to the page.
estimate_ai_footprint
get_methodology
compare_water_use
estimate_session_footprint
API key
None
Transport
Remote, streamable HTTPRead-only, rate-limited, inputs not logged
https://footprint.pharmatools.ai/mcp
Not an MCP client? The same tools are plain GET endpoints: try one.
AI Footprint is a remote server, so it goes in through Settings → Connectors → Add custom connector instead. Paste https://footprint.pharmatools.ai/mcp.
Run in your terminal. Add --scope user to make a server available in every project.
terminal
claude mcp add redacta -- npx -y redacta-mcp
claude mcp add pubcrawl -- npx -y @pharmatools/pubcrawl
claude mcp add studydiff -- npx -y studydiff-mcp
claude mcp add opengate -- npx -y @pharmatools/opengate-mcp
claude mcp add groundwork -- npx -y @pharmatools/groundwork mcp
claude mcp add --transport http ai-footprint https://footprint.pharmatools.ai/mcp
Paste into ~/.cursor/mcp.json for every project, or .cursor/mcp.json in one repo.
Already working with an agent that can edit files? Paste this and let it do the setup.
prompt
Set up the PharmaTools MCP servers described at https://www.pharmatools.ai/mcp. Add these six entries to my MCP client configuration. The first five run locally with npx:
- redacta: npx -y redacta-mcp
- pubcrawl: npx -y @pharmatools/pubcrawl
- studydiff: npx -y studydiff-mcp
- opengate: npx -y @pharmatools/opengate-mcp
- groundwork: npx -y @pharmatools/groundwork mcp
The sixth is a remote server over streamable HTTP, with no install:
- ai-footprint: https://footprint.pharmatools.ai/mcp
Show me the config change before you save it, then tell me how to restart the client and list the tools you can see.
In your environment
Or run it inside your own cloud.
For teams that need tools inside their own cloud project, behind their own sign-in, rather than installed on laptops. PubCrawl ships as a container that serves MCP over HTTP, and the steps below were run end to end on Google Cloud in November 2026.
Build it anywhere
A non-root image with a health check that shuts down cleanly when the platform stops it. CI builds and tests it on every push.
terminal
git clone https://github.com/nickjlamb/pubcrawl && cd pubcrawl
docker build -t pubcrawl . && docker run -p 8080:8080 pubcrawl
Put your sign-in in front
Deployed to Cloud Run with authentication required, requests without a Google identity are refused before they reach the server. PubCrawl holds no credentials of its own.
Agent Registry can’t read tools from a server that requires sign-in, so PubCrawl ships a tool spec generated from its own code, and a test fails if the two drift apart.