Clinical Trial Registry Search (ct-registry)

🇨🇳 中文 / 🇺🇸 English

<div align="center"> <img src="assets/icon.svg" width="240" height="240" alt="ct-registry logo"/> </div>

Cross-source search of global clinical-trial registries (ClinicalTrials.gov, China CDE, WHO ICTRP, EU-CTR, ChiCTR, ISRCTN, DRKS, PubChem), normalized and aggregated into one actionable landscape — for trial planning, de-duplication, control-design benchmarking, and competitive intelligence. B-tier: ordinary input + public retrieval, quickly-adoptable.

Who This Is For

The ct-* clinical-trial skill family covers the entire clinical-trial lifecycle. It serves three groups:

• Clinical-trial practitioners at pharmaceutical companies — sponsors, CROs, and medical / statistical / regulatory roles;
• Clinicians and nurses who design, manage, or run clinical-trial projects;
• Medical students who want to learn clinical-trial methodology in a structured way.

How to Use (in conversation)

Just describe what you need in plain language. The assistant auto-decides which registries, which keywords (it localizes CN/EN for you), and which retrieval mode — you never pick flags yourself. Below are real examples from simple to complex.

Example 1 · Quick registry search

You say:

Search Pfizer's phase 3 oncology trials registered since 2022.

Assistant replies (sketch):

Queries ClinicalTrials.gov v2 directly, normalizes phase / status / sponsor, and returns a ranked table (count, scope, sample rows) plus a homepage link per trial.

How to trigger real computation:

Safe preview prints the plan only; say "run it" / "请直接计算" to execute the live retrieval.

Example 2 · Cross-source (CT.gov + China CDE)

You say:

Find trials on 奥希替尼 (osimertinib) in non-small-cell lung cancer, both globally and in China.

Assistant replies (sketch):

Proposes a scope of ClinicalTrials.gov v2 + China CDE, auto-translates the keyword to English for CT.gov and to Chinese (奥希替尼) for CDE, runs both, then merges them into one unified landscape (e.g. CT.gov 20 + CDE 2057 → 2077 records) with a phase / status / sponsor breakdown.

How to trigger real computation:

Safe preview prints the keyword map + scope; say "run it" to execute. CDE uses the unified Coze endpoint (token required); CT.gov is free and direct.

Example 3 · Complex — multi-parameter CDE routing menu

You say:

Search CDE for 化学药物 (chemical drug) 帕博利珠单抗 in 非小细胞肺癌, status 进行中, applicant 默沙东.

Assistant replies (sketch):

Recognizes this as a CDE advanced (structured) search. Instead of guessing, it pops a routing menu so you confirm each dimension before the query is built:

┌─ CDE search scope ──────────────────────────┐ │ 1) Keyword / drug name : 帕博利珠单抗 │ │ 2) Indication : 非小细胞肺癌 │ │ 3) Drug type : 化学药物 │ │ 4) Applicant (申办方) : 默沙东 │ │ 5) Trial status : 进行中 │ │ 6) Confirm & run │ │ 7) 还拿不准 → 详细解释差异 (explain differences) │ └─────────────────────────────────────────────┘

Selecting 7) 还拿不准 → 详细解释差异 opens a sub-explanation of what each field filters, why trial_status must use the exact enum value (进行中, not "正在进行"), and that CDE has no phase filter (phase comes from detail mode later). You adjust any item, then confirm.

How to trigger real computation:

The menu is a safe preview. After you confirm (item 6), say "run it" to send the payload to the CDE unified endpoint.

Example 4 · Vague — grill-me clarification

You say:

I'm not sure what I want yet — I just heard about a competitor's new drug.

Assistant replies (sketch):

Asks branch by branch to pin down your need, e.g.:

• "Do you know the drug name, or just the disease area / mechanism?"
• "Are you comparing globally, or specifically in China (CDE)?"
• "Is this for competitive landscaping, control-design benchmarking, or de-duplication before your own trial?"
• "Any phase / status / time window you care about?"

Once clarified (e.g. "a GLP-1 class drug, global + China, phase 3, last 3 years"), it proposes a scope and keywords, then previews the plan.

How to trigger real computation:

Safe preview throughout clarification; say "run it" once the scope is confirmed to execute.

Example 5 · Structured detail + document links

You say:

Pull the detailed records (sponsor, phase, endpoints) for the NSCLC CDE hits, and list any downloadable PDFs.

Assistant replies (sketch):

Runs CDE detail mode (65 fields per record) so sponsor / phase are populated (list mode shows them as Unknown). It renders a homepage link for every record and lists (but does not auto-download) any EU-CTR PDFs. PDFs download only after you explicitly confirm.

How to trigger real computation:

Detail fetch auto-runs if ≤100 items; if >100 it confirms the list first. Say "download the PDFs" to trigger the explicit download gate.

Example 6 · Broad global coverage via WHO ICTRP

You say:

Give me the global landscape of DPP-4 inhibitor trials for type 2 diabetes since 2023.

Assistant replies (sketch):

Uses WHO ICTRP (one call mirrors 14+ primary registries: CT.gov, EU-CTR, ISRCTN, DRKS, ChiCTR, jRCT, ANZCTR, CTRI …) plus China CDE (always searched independently, since WHO's English-title matching misses Chinese-registered trials), aggregates into one landscape with a competitor map.

How to trigger real computation:

Safe preview shows scope (WHO + CDE = two endpoints); say "run it" to execute. Counted as one demand against the shared quota.

What It Can Do — Scenarios

| Capability | Source(s) | Try saying | |---|---|---| | Search by disease / intervention / sponsor | CT.gov (direct); CDE / ChiCTR (user-paste or workflow) | "Find phase 3 oncology trials by Pfizer since 2022" | | Unified cross-registry normalization | All sources | "Merge CT.gov and CDE NSCLC trials into one table" | | Aggregation: phase / status / sponsor / timeline / competitor landscape | All sources | "Show the competitor landscape for osimertinib in NSCLC" | | China CDE advanced & multi-keyword search | China CDE (unified endpoint) | "CDE: 沙坦 + 进行中, chemical drug, since 2023" | | Global mirror via WHO ICTRP (14+ registries) | WHO ICTRP (unified endpoint) | "Global DPP-4 inhibitor T2D trials since 2023" | | Drug → target / property mapping | PubChem (direct) | "Map osimertinib to its target via PubChem" | | Structured detail (sponsor / phase / endpoints) | CDE detail (65 fields), EU-CTR docs | "Pull full details for these CDE registration numbers" | | Structured output: JSON / Markdown / PNG / Excel | — | "Export the landscape as an Excel workbook" | | Chained downstream analysis | → ct-pipeline / ct-protocol | "Feed the normalized results to ct-pipeline for intel" |

FAQ

Can I search with only some parameters? Yes. You never choose a mode or flag — the assistant auto-decides the CDE calling mode (search / combined / multi_keyword / detail) from your words. Leave any field blank; unused fields are simply omitted (empty fields would otherwise poison the query).

Are the counts per group or total? Totals are aggregated into one landscape. A cross-source run reports the grand total and the per-source breakdown (e.g. CT.gov 20 + CDE 2057 → 2077 records). The status distribution shows both RECRUITING and 已完成/进行中 in the same table.

How do I actually retrieve the data? By default you get a safe preview (the plan / payload only). Say "run it" / "请直接计算" to execute the live retrieval.

Is the output in Chinese when I'm on a Chinese system? Yes. The scripts auto-detect locale: all user-facing prompts switch to Chinese on a zh-* locale and to English otherwise. Raw data values (CDE Chinese status / indication, etc.) are always kept verbatim — never translated.

Why do some rows show phase / sponsor as Unknown? List-mode search returns only summary fields. sponsor and phase are populated only by CDE detail mode (65 fields) or by WHO detail. Request detail when you need sponsor / phase breakdowns.

Is it free, and is there a quota? Currently free. The shared third-party endpoint is metered by demand (one user request = 1 demand; WHO + CDE + keyword tweaks within it collapse to 1). Daily cap is 100 demands; direct Tier-1 sources (CT.gov v2, EU-CTR, PubChem) and preview are not counted.

Concurrency / parallel sessions (P1-7)? The quota counter config/usage.json and the output dir are shared files inside the skill directory. usage_guard.py serialises their read-modify-write with an advisory file lock, so parallel calls won't corrupt the counter. Best practice is still one skill per session; if multiple AI sessions drive the same skill directory at once, rely on that lock and avoid sharing an out_dir.

CDE standalone endpoint status? The original standalone CDE endpoint ct-searchcde.coze.site/run (CDE/search_cde_workflow.py) was RETIRED on 2026-08-12; --cde-legacy now only prints a deprecation warning and auto-routes to the unified endpoint search_ictrp.py --source chinadrugtrials. CDE/ is local-archive only and not shipped.

Is the result a regulatory submission? No. Output is for reference / planning only. CSR / filing documents must be produced separately per GCP.

Is the retrieval compliant, and does it hit rate limits? All queries go through official public APIs / public search interfaces (CT.gov v2, EU-CTR, PubChem, WHO ICTRP, China CDE unified endpoint) — no scraping of non-public pages, no bypassing of site protections. The backend calls third-party endpoints serially with exponential backoff, honouring their rate limits; on 429 / timeout it retries with backoff rather than hammering them. The 5-minute gateway cap on global mirrors (14+ registries) is a compliant wait, not a throttling penalty.

Before You Run — Time & Data Limits

When you mirror 14+ registries through WHO ICTRP in one call, the backend literally crawls those registries — it is not instant. Here's what to expect:

• How long: a live retrieval typically takes 1–5 minutes to come back. The skill first hands you a "submitted, running" receipt, then polls automatically and delivers the result when ready — you don't have to babysit it.

• Why it can hug the 5-minute ceiling: the unified third-party endpoint sits behind a hard 5-minute gateway cap. The workflow uses P4 async fire-and-forget (/run returns immediately with a run_id, then polls /run/status/{run_id} with exponential backoff) to bypass this wall. Broad queries like cancer, or "global + no country filter", can run right up against that wall and, rarely, time out. If it does, the skill tells you explicitly instead of silently dropping data.

• What happens after 5 minutes: if the retrieval volume exceeds the 5-minute cap and still isn't complete, the system returns whatever partial results have been collected so far (rather than erroring out), and marks the output with is_timeout: true. You can trade off between wait time and data completeness: wait longer for full results, or accept partial results for a quick preview.

• Roughly how many records in 5 minutes: it depends on how broad the query is:

  | Keyword type | Typical hit count | Can 5 min finish? | Empirical throughput | |-------------|------------------|-------------------|---------------------| | Narrow (e.g. osimertinib) | 20-200 | ✅ Easily | WHO ~100-150 pages/min, CDE ~200-300 pages/min | | Medium (e.g. diabetes) | 5,000-50,000 | ❌ Partial only | 5 min ≈ 1000-1500 records | | Broad (e.g. cancer) | 100,000+ | ❌ Only a small fraction | First page ~30s, ~2-3s/page thereafter (WHO) |

  What to do when you can't get everything: ① Accept partial results (enough for trend analysis); ② Narrow the scope (add intervention / sponsor / country filters); ③ Split into multiple queries (by dimension).

• How to keep the data volume sane: WHO global results can be huge. If a query returns too few rows or times out, it's usually too broad — narrow the keyword (e.g. osimertinib instead of cancer), or use advanced search (just say "advanced search" in conversation; the code auto-builds precise drug AND condition filters, returning a smaller, faster set). There's also a daily cap of 100 demands (see FAQ above) — normal use won't get near it.

Safety (safe preview)

Safe preview by default. Scripts only generate and display a plan / payload. Network requests run only when you explicitly say so ("run it" / "请直接计算"). Nothing is sent or fetched until then.

Public data only — zero confidential input. The skill reads public registry data; you never supply subject / protocol / CRF data, and none is ever transmitted.

Outbound data disclosure (egress). When you run a live retrieval, only public query terms (drug name, indication, registration number) leave your environment, sent to these public endpoints:

• ClinicalTrials.gov v2 — official REST API (direct, no token).
• PubChem — PUG-REST (drug → CID / properties / targets; direct, no token).
• China CDE — via the unified Coze /run endpoint ct-search.coze.site/run (third-party, Bearer token required — verified HTTP 401 without it).
• EU-CTR — pure-HTTP HTML parse of legacy EudraCT results (direct, no token, no browser).
• WHO ICTRP — unified Coze /run endpoint (third-party, Bearer token required); one call mirrors 14+ registries.
• ChiCTR / ISRCTN / DRKS — served by the same unified Coze /run endpoint via source=chictr|isrctn|drks (third-party, shared token).
• Bug reports (optional, opt-in only): if a likely skill defect is detected, the assistant may ask to send a sanitized bug report to the unified report endpoint https://ct-bugreport.coze.site/run (skill name/version/error type plus a problem description you review and approve — never your raw data). Nothing is sent without your confirmation; you can always decline.

First-outbound confirmation. The first time a search sends your terms to the unified endpoint, you'll see a one-time confirmation prompt (target server + what is sent = your public search terms, no personal info). After you approve, the endpoint is whitelisted and you won't be asked again this session. This follows the ct-base §5 library-wide outbound-authorization norm. Nothing is sent until you approve or say "run it".

WHO ICTRP and China CDE share one long-lived token on the unified endpoint. It is a PUBLIC shared credential, embedded as an XOR+base64 blob in config/keys.py (shipped with the package) so it works out-of-the-box. Resolution order: CLI(--token) > env(CT_REGISTRY_COZE_TOKEN, legacy ICTRP_WORKFLOW_TOKEN) > embedded blob (ct-base §5.236). No confidential data ever reaches any of these endpoints.

Coze key (unified-endpoint credential)

The unified Coze endpoint https://ct-search.coze.site/run (used by CDE, WHO ICTRP, ChiCTR, ISRCTN, DRKS) needs a Bearer token. It is a shared public credential — issued by the author, bound to the endpoint, not your personal secret.

• It just works: the token is embedded in config/keys.py (an obfuscated XOR+base64 blob) and ships with the skill, so retrieval runs out-of-the-box with no setup.
• To override it (e.g. the author re-issues the token): pass --token <JWT> on the command line, or set the CT_REGISTRY_COZE_TOKEN environment variable. Do not paste tokens into chat.
• Obfuscated, not encrypted: the encoding hides the string from casual viewing, not from a determined reader. Treat it as a credential, not a secret to protect at all costs.
• Security scanners: some automated scanners flag extsvc_client.py (HTTP/Bearer usage). The blob is a public shared credential, not a private key — there is no private secret in the repo. Override via CLI/env only.

Why You Can Trust the Results (anti-hallucination)

Every row in the output is retrieved, not invented:

• Source-authoritative. Records come straight from the registries' own APIs / endpoints (ClinicalTrials.gov v2, PubChem PUG-REST, EU-CTR HTML, and the unified Coze endpoint for CDE / WHO / ChiCTR / ISRCTN / DRKS). No LLM generation sits between you and the source data.
• Rule-based normalization. Merging / field mapping (phase, status, sponsor…) is deterministic code, not free-text rewriting. Raw values (CDE Chinese status, indication, etc.) are kept verbatim — never translated or altered.
• Counts are echoed, not estimated. Totals and per-source breakdowns are taken directly from each source's response; the skill does not round, guess, or extrapolate counts.
• Known limits are labelled, not hidden. A broad query that hits the 5-minute gateway returns partial results marked is_timeout: true (see Before You Run — Time & Data Limits); list-mode rows show Unknown for sponsor / phase (populated only by detail mode). These caveats are printed, not silently dropped.

Caveat: registries update on their own cadence and may have gaps or lags; always cross-check against the source before any regulatory / submission use.

Advanced Reference

The commands below are for developers / power users. In normal conversation you do not type these — the assistant builds and runs them for you behind the safe-preview gate.

⚠️ Commands that include --run perform live network retrieval (public query terms egress to the endpoints above). Omit --run — or stay in the conversation safe-preview — for the default no-network plan that prints only the intended call.

Requirements

• Python 3.10+ (Anaconda recommended).
• Required: requests, pandas, beautifulsoup4, lxml.
• Optional: matplotlib (PNG charts); playwright + playwright install chromium (CDE local scrape, kept only as last resort).

One-shot orchestration (CT.gov + PubChem + aggregate + report)

python scripts/ct_registry.py --cond "NSCLC" --status RECRUITING --with-pubchem --out-dir ./out

Cross-source orchestration (CT.gov + CDE merged into the SAME landscape)

Merge CT.gov + CDE in one run: English / Chinese primary terms with auto keyword localization, explicit CDE keyword override, advanced filters (--cde-mode combined), and multi-keyword AND (--cde-multi-keywords). Full CLI examples: references/cli_reference.md.

Full landscape (Tier-1 + Tier-2 external services)

One invocation covering CT.gov + EU-CTR + CDE + ChiCTR + ISRCTN + DRKS + WHO ICTRP, merged and aggregated into one report. Full CLI example: references/cli_reference.md.

Per-source direct scripts

# CT.gov (required, official API)
python scripts/search_ctgov.py --cond "non-small cell lung cancer" --status RECRUITING --max 50 --out ctgov.json
# CT.gov advanced search (full v2 query.* / filter.* coverage, v0.3.84+)
python scripts/search_ctgov.py \
  --cond "non-small cell lung cancer" \
  --phase PHASE1,PHASE3 --study-type INTERVENTIONAL \
  --status RECRUITING,COMPLETED \
  --start-date-since 2023-01-01 --last-update-until 2026-08-01 \
  --sort LastUpdatePostDate:desc --has-results --fields NCTId,BriefTitle --max 10
# CT.gov raw Expert Search expression passthrough (paste from website advanced search)
python scripts/search_ctgov.py --adv 'AREA[StudyType]OBSERVATIONAL AND AREA[LocationCountry]China' --max 5
# PubChem drug -> target
python scripts/enrich_pubchem.py --drug "osimertinib" --targets --out pubchem.json

CT.gov advanced flags (v0.3.85+, full list via --help): --query (query.term, AREA[] ok), --titles --outc --lead --id --locn --patient (other query.), --adv (raw filter.advanced), --phase --study-type --age-group --sex --has-results (assembled into filter.advanced AREA[] expressions, AND-combined with --adv), --sort (max 2; field[:asc|:desc] or @relevance), --ids (NCT batch), --geo (distance(lat,lon,dist) or bare lat,lon,dist; 1-500 mi / 1-805 km enforced), --fields, server-side date ranges --first-post-since/--until, --last-update-*, --start-date-*, --primary-completion-*, --completion-*, multi-value --status, and --post-status/--post-ids/--post-geo/--post-adv (postFilter. — same semantics but does NOT affect relevance ranking). NOTE: v2 has no flat phase/studyType/date params — they all go through filter.advanced; query.rmtln is a v1 legacy param removed in v2 (HTTP 400, verified against the live API 2026-08-13) — use --query/--adv for remote-trial search; --date-after keeps the legacy local post-filter (start-year lower bound).

Unified endpoint (WHO ICTRP + China CDE + ChiCTR / ISRCTN / DRKS)

WHO + CDE + ChiCTR / ISRCTN / DRKS share one Coze endpoint; the token is embedded in config/keys.py (shipped) — no manual step needed. To override on a rare 403, set CT_REGISTRY_COZE_TOKEN env or pass --token. Per-source CLI examples: references/cli_reference.md.

CDE four calling modes (RETIRED 2026-08-12; local archive reference only; production uses the unified endpoint)

Production CDE retrieval goes through the unified endpoint above; the retired standalone workflow is archived locally under CDE/ (NOT shipped).

Normalize → aggregate → report → Excel

python scripts/normalize.py --ctgov ctgov.json --cde cde.json --chictr chictr.json \
    --euctr euctr.json --isrctn isrctn.json --drks drks.json --out normalized.json
python scripts/aggregate.py --in normalized.json --out agg.json
python scripts/report.py --in agg.json --out report.md --png report.png
# Clinical-friendly 4-sheet Excel (auto-generated by the orchestrator; --no-excel to disable)
python scripts/export_xlsx.py --in normalized.json --out report.xlsx --title "Asciminib 2023-2026"

Structured detail + confirm-gated PDF download

Detail fetch (≤100 items auto, >100 confirm) and PDF download are gated by explicit confirmation; PDFs are never auto-downloaded. Direct doc download is resumable (skips existing files). CLI examples: references/cli_reference.md.

⚠️ Document-download duration & volume warning (measured 2026-08-12) download_docs.py downloads sequentially (one file at a time, no concurrency) and has no built-in quantity cap — it pulls every documents[].url in the input JSON. Only EU-CTR yields real downloadable links (CDE / WHO / CT.gov expose no protocol-PDF API), so a broad EU-CTR search (e.g. 40 records × ~3 docs each) can mean 100+ sequential downloads.

Measured behavior (sandbox, 2026-08-12):

• Per-file hard timeout --timeout default 60s; on failure it prints FAILED and moves on — there is no retry loop (the docstring says "per-file retries" but the code only tries once).
• Round-trip latency to euclinicaltrials.eu ≈ 1–3 s/request.
• This sandbox's egress is throttled to ~20–30 KB/s, so a 3 MB PDF would hit the 60s timeout and be skipped — i.e. in a throttled network, large docs fail. Production (Coze) bandwidth is normally far higher, but don't assume a doc will finish.

Planning rule of thumb: total wall-clock ≈ N_files × per_file_time, with a worst-case ceiling of N_files × 60s (a hung file burns the full timeout before skipping). For 100 files that is up to ~100 min in the worst case. EU-CTR protocol / IB / CSR PDFs are commonly 0.5–15 MB, so real per-file time varies with doc size and deployment bandwidth.

Mitigations:

• Cap how many docs are gathered: fetch_eu_ctr_docs.py --max N and/or download only a subset of records.
• Download in small confirmed batches; re-run is safe — files already in --out-dir are skipped (resumable).
• If you see many FAILED … timed out, raise --timeout (e.g. --timeout 180) or run from a faster network.

WHO ICTRP advanced fields (subset → CLI)

--who-title (+ operator), --who-condition, --who-intervention (alias --intr), --who-recruitment-status (alias --status), --who-sponsor, --who-country, --who-phase (comma-separated), --who-date-start / --who-date-end (DD/MM/YYYY), --who-with-results, --who-secondary-id. Any structured field auto-selects mode=combined.

CDE advanced fields (subset → CLI)

--reg-no, --indication, --case-no, --drugs-name, --drugs-type (enum: 中药/天然药物/化学药物/生物制品), --appliers, --communities, --researchers, --agencies, --trial-status (11-value enum: 进行中/尚未招募/招募中/招募完成/已完成/主动暂停/主动终止/IEC·IRB暂停/IEC·IRB终止/责令暂停/责令终止). Note: CDE advanced UI has no phase filter — phase is post-filtered from detail.

Interaction contract (two gates)

• Gate 1 — Pre-search brief (before any live-network call): the assistant prints keywords (EN+ZH per source), scope, any time/status filter, the demand_id grouping + quota impact, and the "free, 100 demands/day" note. It asks only when there is a real choice.
• Gate 2 — Post-list confirmation: it presents the list (count, scope, sample rows, Unknown phase/sponsor caveat). Detail fetch: ≤100 items auto-runs; >100 items confirm first. PDFs are never auto-downloaded.

Errors (quick reference)

| Error | Cause | Fix | |---|---|---| | CT.gov URLError | No network / proxy | Confirm clinicaltrials.gov reachable; set proxy | | CDE empty / "access blocked" | SafeDog WAF blocks browsers | Use unified endpoint workflow (preferred); or assisted paste → parse_cde.py (no egress) | | WHO / CDE HTTP 401 | Missing Authorization: Bearer <token> | Token is embedded in config/keys.py (shipped); if a 401 persists, set CT_REGISTRY_COZE_TOKEN or pass --token | | WHO / CDE HTTP 403 (rare) | Corrupted / revoked token blob | Re-issue via env CT_REGISTRY_COZE_TOKEN / --token (token is long-lived; 403 is not expiry) | | CDE HTTP 500 字段类型错误 | Field sent as {"value":x} or project_list as array | Use plain strings; project_list must be a JSON string | | ISRCTN 404 | Public API dead (2026-07-20) | Use unified endpoint source=isrctn | | CDE read timed out | Large result / transient gateway (capped ~300 s/run) | Retry with default 300 s; transient, not a payload error |

Pipeline

ct-registry → ct-pipeline (consumes normalized.json for competitor intel) / ct-protocol (design benchmarking).

---

Version: v0.3.81 | License: MIT | Authors: medstatstar, phoe-zip

For feature requests, bug reports, or other feedback, please contact the author directly at medstatstar@gmail.com (Wintone Zhang / 张文彤).

Confidentiality Notice

The CT series consists of 20+ specialized domain skills, organized into two tiers — A, B — by "confidential-data-exfiltration risk + whether external retrieval is needed", providing full coverage of the entire new-drug clinical trial (Clinical Trial) lifecycle.

• Tier A (non-confidential · public): takes only ordinary (non-confidential) input; runs fully locally (network=off) or performs public retrieval (network=public-retrieval, e.g. ct-registry / ct-advisor) — never involves confidential information. Tier A skills are published openly on GitHub.
• Tier B (confidential · internal): involve strictly confidential clinical-trial data and internal information from pharma sponsors (e.g., ct-analysis, ct-sdtm, ct-eligibility); Tier B is processed locally (egress=none, data never leaves the machine) or requires approved egress (egress=approval-req, e.g. ct-eligibility). These skills are designated for internal enterprise use only and are not publicly released at present.

If you do have a genuine need for these confidential skills, please contact the author to request custom installation.

📧 Contact: medstatstar@gmail.com (Wintone Zhang / 张文彤)