Varys
Varys is a market intelligence tool for health IT, enabling vendor and health system discovery and profiling.
Varys
A market intelligence tool for the health IT ecosystem with four capabilities:
- Vendor discovery — Build a competitor list from natural language. "Find AI scribe competitors to Nuance" → curated company list → CSV ready for profiling.
- Vendor profiling — Profile health IT companies for competitive analysis. Who are they, what do they sell, who have they sold to, how are they funded, and what is their regulatory status?
- Health system discovery — Discover hospitals and health systems by US state from the public CMS dataset. No input list needed — a free, open-source alternative to Definitive Healthcare.
- Health system profiling — Profile hospitals and health systems for BD prospecting. EHR vendor, bed count, CMS star rating, payer mix, CIO name, and more.
Two ways to use this tool:
One at a time (interactive) — Load a skill file into any AI assistant (Claude, ChatGPT, Gemini, etc.) and research entities conversationally. Ask follow-up questions, refine results, and build a competitor list interactively before exporting.
<img src="https://github.com/user-attachments/assets/42710642-6c8f-43dc-addc-522c7ae22aa3" width="600"/><br/><br/>
Batch processing (CLI) — Run varys.py to process a CSV list end-to-end. Outputs clean values and source-cited results at any scale.
Useful for
BD teams, competitive intelligence analysts, and market researchers in health IT who need structured, source-cited data without a Definitive Healthcare subscription.
Quickstart
Interactive (AI assistant)
Load a skill file from the skills/ folder into your AI assistant of choice. Then either ask:
Find me AI scribe competitors to Nuance
The assistant will propose a list and let you refine it ("remove Nuance itself", "add Suki", "only keep Series B+") before exporting to CSV.
Or use a slash command:
/profile-health-it-vendor Abridge
Batch (CLI)
# Install dependencies
pip install -r requirements.txt
# Set your API key
export ANTHROPIC_API_KEY=sk-ant-...
# --- Vendors ---
# Discover competitors via natural language (prompts for query)
python varys.py discover vendor --output discovered_vendors.csv
# Profile vendors from a list
python varys.py profile vendor --input discovered_vendors.csv --output results.csv
# Or do both in one shot
python varys.py pipeline vendor --output results.csv
# --- Health systems ---
# Discover all hospitals in California from CMS public data
python varys.py discover health-system --state CA --output ca_hospitals.csv
# Profile health systems from a list
python varys.py profile health-system --input ca_hospitals.csv --output ca_results.csv
# Or do both in one shot
python varys.py pipeline health-system --state CA --output ca_results.csv
Architecture
flowchart LR
A["Natural language query"] -->|discover| B["Entity list"]
B --> C["Profile loop<br/>with one context window per entity"]
D["CSV input"] --> C
C --> E["results.csv"]
C --> F["results_sources.csv"]
Output
Every run writes two CSVs:
| File | Contents | Use case |
|---|---|---|
results.csv | Clean values only | Downstream consumption, import, sharing |
results_sources.csv | Values + source URLs + confidence levels | QA, verification, auditing |
Vendor skill fields
results_sources.csv adds _source and _confidence for every field.
| Field | Values |
|---|---|
product_category | AI Scribe / EHR / RCM / Care Management / CDT / Patient Engagement / Clinical Decision Support / Interoperability / Other |
primary_customer | Provider / Payer / Employer / DTC |
business_model | SaaS / Per-Seat / PMPM / Implementation Fee / Usage-Based / Other |
fda_status | Not Required / Cleared / Breakthrough Device / PMA / Pending / Unknown |
funding_stage | Seed / Series A / Series B / Series C / Series D+ / Public / Profitable / Unknown |
clinical_evidence | true / false |
Free-text fields: entity_name, ehr_integrations, notable_health_system_customers, total_funding, key_investors, num_employees, headquarters, founded_year
Health system skill fields
results_sources.csv adds _source and _confidence for every field.
| Field | Values |
|---|---|
ownership_type | Non-profit / For-profit / Academic / Government / Unknown |
ehr_vendor | Epic / Oracle Health / Meditech / Allscripts / athenahealth / Other / Unknown |
cms_star_rating | 1 / 2 / 3 / 4 / 5 / null |
teaching_hospital | true / false |
vbc_participation | true / false |
innovation_program | true / false |
geographic_region | Northeast / Southeast / Midwest / Southwest / West |
Free-text fields: entity_name, health_system, bed_count, payer_mix, annual_revenue, recent_tech_announcements, cio_name
CLI subcommands
varys.py discover vendor # prompts for query → writes entity CSV
varys.py discover health-system --state CA # CMS data → writes entity CSV
varys.py profile vendor --input vendors.csv
varys.py profile health-system --input hospitals.csv
varys.py pipeline vendor # discover + profile in one shot (interactive query)
varys.py pipeline health-system --state CA # discover + profile in one shot
discover flags:
| Flag | Default | Description |
|---|---|---|
--state | (required for health-system) | Two-letter state code (e.g. CA, NY). |
--output | (varies) | vendor-results.csv or <STATE>-health-systems.csv |
--model | claude-sonnet-4-6 | Anthropic model (vendor only). Override via ANTHROPIC_MODEL. |
profile and pipeline flags:
| Flag | Default | Description |
|---|---|---|
--input | (required for profile) | Input CSV with entity_name column. |
--output | see below | Output CSV. A _sources.csv is auto-written alongside it. |
--batch | false | Async mode via Anthropic's Batches API. See Async mode vs agentic. |
--concurrency | 1 | Parallel API calls. Default is 1 (sequential). Increase based on your Anthropic rate limit tier — see Concurrency and rate limits. |
--model | claude-sonnet-4-6 | Anthropic model. Override via ANTHROPIC_MODEL. |
--yes | false | Skip the cost confirmation prompt. |
Default --output filenames: varys-vendor-research-results.csv, varys-health-system-research-results.csv, vendor-pipeline-results.csv, <state>-pipeline-results.csv.
Requirements
ANTHROPIC_API_KEY— set in environment before runninganthropic>=0.84.0pyyaml>=6.0
Async mode (--batch) vs agentic
Anthropic's Messages Batches API offers a 50% discount off regular API pricing by processing requests asynchronously. --batch uses this to keep costs low without sacrificing quality — it starts with a cost-efficient single-shot pass, then automatically runs the full agentic loop only on entities where the batch returned weak or missing fields. This is the tradeoff: lower cost at the expense of a longer total turnaround.
--batch (hybrid) | Default (agentic) | |
|---|---|---|
| How it works | Single-shot with adaptive thinking (50% discount), then agentic follow-up on weak fields only | Full multi-round agentic loop from the start |
| Well-known companies | Excellent — high confidence, minimal follow-up needed | Excellent |
| Obscure/early-stage companies | Good — adaptive thinking helps, agentic follow-up catches gaps | Best — multi-round web research from the start |
| Cost | ~$0.02/company in tests | ~$0.20/company |
| Turnaround | ~1 hour for 10 companies (async, Anthropic-side queue) | ~2–3 min per entity sequentially |
| Rate limit pressure | None during batch phase; agentic follow-up is subject to your TPM limit | Subject to your TPM limit throughout |
Rule of thumb: use --batch when you want to optimize for cost over latency, or your list is mostly well-known companies. Use agentic when you need results immediately or your list is mostly obscure/early-stage companies.
The CLI always shows a cost estimate and requires confirmation before any API call. Use --yes to skip this in CI or scripted workflows.
Cost estimates based on claude-sonnet-4-6 (March 2026). Model pricing changes — always check anthropic.com/pricing before large runs.
Concurrency and rate limits
The default --concurrency 1 runs entities sequentially. Agentic calls with web search accumulate large contexts — a single round can send 10,000–20,000 input tokens. Running multiple entities in parallel multiplies this and can trigger 429 rate limit errors.
Adjust based on your Anthropic usage tier:
| Rate limit | Recommended --concurrency |
|---|---|
| 30,000 input tokens/min | 1 (default) — even sequential can occasionally hit limits on heavy searches |
| 100,000 input tokens/min | 3–5 |
| 200,000+ input tokens/min | 5–10 |
Rate limit errors are retried automatically with exponential backoff (60s, then 120s), so a run will always complete — it just takes longer at lower tiers.
--batch is unaffected by this limit — batch requests are processed asynchronously on Anthropic's side and have a separate rate limit bucket.
Further reading
For design decisions — why Python over an LLM orchestrator, how context isolation works, source priority per field, and how to tune research depth — see Architecture & Design Decisions.
If you use OpenAI or Gemini instead of Anthropic, see Using with other AI assistants to adapt varys.py to another provider.
This project was built while working through DeepLearning.AI's course, Agent Skills with Anthropic.