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.

Batch processing (CLI) — Run varys.py to process a CSV list end-to-end. Outputs clean values and source-cited results at any scale.

BD teams, competitive intelligence analysts, and market researchers in health IT who need structured, source-cited data without a Definitive Healthcare subscription.

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

# 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

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"]

Every run writes two CSVs:

results_sources.csv adds _source and _confidence for every field.

Free-text fields: entity_name, ehr_integrations, notable_health_system_customers, total_funding, key_investors, num_employees, headquarters, founded_year

results_sources.csv adds _source and _confidence for every field.

Free-text fields: entity_name, health_system, bed_count, payer_mix, annual_revenue, recent_tech_announcements, cio_name

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:

profile and pipeline flags:

Default --output filenames: varys-vendor-research-results.csv, varys-health-system-research-results.csv, vendor-pipeline-results.csv, <state>-pipeline-results.csv.

• ANTHROPIC_API_KEY — set in environment before running
• anthropic>=0.84.0
• pyyaml>=6.0

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.

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.

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 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.

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.