tag.bio MCP Request access
Model Context Protocol connector for Tag.bio

Analyse biomedical data using AI Harness. Compute stays at the data.

Validated analyses over governed data products such as TCGA and FAERS. Only results leave Tag.bio.

Connect in 2 minutes Browse data products
Works with Claude ChatGPT Claude Code Codex
Data products

Start with the data: TCGA, MSK, FAERS, METABRIC

A data product is one governed table, one row per entity, with its own validated analyses: survival, clustering, differential expression, cohort comparison.

cancer genomics · row = tumour sample

TCGA PanCancer Atlas

10,967
tumour samples
32
cohorts

Clinical, survival, mutation, copy-number and RSEM expression. Also available as the GDC harmonisation with TPM expression.

pharmacovigilance · row = report

FDA Adverse Event Reporting System (FAERS)

1,827,391
adverse event records
77,296
drug names
16,079
reaction terms

Drugs, indications, reactions and patient outcomes across 1,702,222 cases, ready for signal and disproportionality questions.

79,603 tumour samples

cBioPortal MSK

Memorial Sloan Kettering sequencing cohort, deduplicated across 112 studies.

99,791 samples

cBioPortal published studies

200 cancer genomics studies across 127 disease types, 2011 to 2025.

11,362 samples

cBioPortal pediatric

18 pediatric oncology studies from TARGET, St Jude and others.

1,980 patients

METABRIC breast cancer

Clinical and multi-omics data, with survival.

1,208 cell lines

Cell lines and models

CCLE, NCI-60 and Sanger lines, plus patient-derived xenografts.

row = gene or variant

Gene and variant annotation

Pathway and variant annotations that join onto results from the other products.

Also on the demo site: clinical trials, patient registries on FHIR and OMOP, single-cell RNA-seq. Enterprise sites add their own.

How it works

Compute at the data. Only results return.

Your AI assistant
Claude, ChatGPT, Claude Code, Codex
MCP connector
demo.tag.bio/mcp
17 tools
Gateway
Your permissions, checked on every call
Data products
Analyses run here
Sandbox
No internet
INSIDE TAG.BIO
question
results only
gateway access only
Row-level data never crosses the boundary.
Transparency

Deterministic protocols. Every step visible.

Deterministic Tag.bio protocols

The model chooses the analysis; the protocol computes it. Same cohort and arguments, same numbers, in chat and in Studio.

Every call in the chat

Each exchange shows the protocol and the exact arguments that ran.

Code you can read

Sandbox Python appears in the chat alongside its result.

Plot overall survival for TCGA breast cancer by molecular subtype.
tool call · shown in the chat
run_protocol(
  id       = "fc-cbioportal-tcga",
  protocol = "r_survival_km",
  cohorts  = background_cohort:
               Cancer Type = "Breast Cancer"
)
same inputs, same result
Kaplan-Meier curves by molecular subtype, with a log-rank test and risk table: 981 samples with survival time and status, 132 deaths.
Features

Everything a scientist needs, in the chat

Find your data

Browse the data products you are allowed to see and what each one contains.

list_data_products · describe_data_product · list_variables · list_levels

Query everything

Questions run over the full dataset, not a sample; only the answer comes back.

run_query

Run curated analyses

Each data product's own validated analyses: survival, UMAP clustering, cohort comparison, differential expression and more.

run_protocol · get_protocol_result · describe_protocol

Build cohorts like in Studio

Pick filters the way the Studio cohort builder does; the same cohort gives the same numbers in both.

cohort_options · build_cohort · argument_options

Charts in the chat

Projections and results come back as charts you can hover and explore.

render_chart

Real numbers from reports

Report-style analyses return their tables (counts, p-values, odds ratios), not just pictures.

report results

Big results, no flooding

Large tables become a short summary plus a full file to download or page through.

read_artifact · download links

Python next to the data

When no curated analysis fits, the assistant's Python runs in an isolated job inside Tag.bio.

run_analysis · get_job · cancel_job

Saved to Studio history

Runs from your assistant appear in Tag.bio Studio as cards you can reopen and share.

Studio history

Sign in, no keys

Connect with your normal Tag.bio login. Okta, Azure AD and SAML sites sign in the same way as Studio; API keys also work.

OAuth · API key

Guardrails built in

Refuses empty or impossible cohorts, checks arguments first, explains empty results, keeps p-values precise.

cohort guard · argument checks

Guided workflows

One-click starting points: explore a product, UMAP clustering, survival analysis, compare cohorts.

MCP prompts

Sandbox

The compute moves to the data, not the other way round

When a question needs a join, a custom model or anything no curated analysis covers, the assistant writes Python and Tag.bio runs it in an isolated sandbox pod inside the cluster. The code pulls what it needs under your entitlements, and only the small result comes back.

run_analysis · sandbox job running as you 
# host and key are preconfigured inside the job
from tagbiopy.fc import FC

fc = FC(fc_name='fc-...')
df = fc.df.select('[collection]', '[collection]').run()

summary = df.groupby('[group]').size()
result = summary.to_dict()   # only this leaves
no internet gateway only short-lived cancellable

Your login, your permissions

Every call and every job runs as you; you see exactly the data products your Tag.bio account can see.

No way out

Model-written code has no internet access. It can reach only the Tag.bio gateway, and it cannot install packages.

Short-lived by design

Each run is its own job. It starts for one question and can be cancelled from the chat at any time.

A scientific stack, preinstalled

numpy, pandas, scipy, scikit-learn, statsmodels, umap-learn, igraph, leidenalg, matplotlib and the Tag.bio Python SDK.

Results build on results

A job can load tables from earlier analyses directly, without pulling the data again.

Files that stay yours

Download links expire after 30 minutes; paged reads are yours alone, for 6 hours.

Example

One question, one answer

A real run on TCGA. The assistant picks the data product's own UMAP analysis, runs it inside Tag.bio and returns labelled clusters with a chart. No rows of patient data cross into the chat.

In TCGA, run a UMAP of the samples and colour by cancer type.
run_protocol · fc-cbioportal-tcga · umap · cluster_count = 10
UMAP of 10,071 TCGA tumour samples by gene expression, coloured by cancer type; most cancer types form separate clusters
10,071 tumour samples, 20,511 genes, 10 k-means clusters. Most cancer types form their own islands: breast cancer is one cluster (895 of 896 samples), and clear-cell and non-clear-cell renal tumours share another. Caveat: the cluster count was set to 10, and the chart plots 2,518 of the 10,071 points.
Get started

Connected in two minutes

Access is by request while the connector runs on the Tag.bio demo site. Email support@tag.bio, then add this address to your assistant:

https://demo.tag.bio/mcp
  1. Open Settings, then Connectors, then Add custom connector.
  2. Paste https://demo.tag.bio/mcp and choose Connect.
  3. Sign in with your Tag.bio login and choose Allow.
  4. Ask: "What data products can I use?"
Three things to try first
List the data products I can access and what each covers. Run a UMAP of the samples, coloured by cancer type. Compare survival between two groups and explain the caveats.

Enterprise customers on their own Tag.bio site use their site's address instead, for example https://your-site/mcp. Private-network sites work from Claude Code or Claude Desktop on that network.

Roadmap

Coming next

planned

Tag.bio AI skills

Open-source skills that teach your assistant to analyse and build Tag.bio data products.

Pricing

Pay with Tag.bio credits.

Start with 25 free credits.

Free start

$0

25 credits. Card check, nothing charged.

Monthly

$200 / month

400 credits a month.

Extra credits

$50 per 100 credits

Buy any time, on any plan.

  • Used after monthly credits.
  • Valid 12 months.
  • Optional auto-reload and spending limit.

1 credit = one query, analysis or Python job. Browsing and building cohorts are free.

FAQ

Questions

Which data can I use?

The data products your Tag.bio account can access. On the demo site these are public datasets such as cBioPortal, METABRIC and FAERS.

Does my data go to the AI model?

Analyses run inside Tag.bio. The results you ask for (counts, statistics, charts) are returned to your assistant.

Our Tag.bio site is on a private network. Can we use it?

Yes, from Claude Code or Claude Desktop on that network, using your site's own address.

New tools are not showing up.

Disconnect and reconnect the connector once; assistants cache the tool list.