Finance Research Agent — Indian Equities (NSE/BSE)

An AI equity-research analyst for the Indian market: 28 structured data tools + RAG over concalls/filings + forensic-scoring skills, on the Claude Agent SDK and MCP. It runs real analyst workflows — sector screens, valuations, SWOTs, forensic audits — and ranks and scores evidence rather than issuing buy/sell calls or made-up price targets.

Not investment advice. Decision-support only. And no third-party data is shipped in this repo — you point the pipeline at your own accounts and it builds a local copy for the few companies you want to study. See DISCLAIMER.

See it in action

Four real, multi-tool analyses. Each links to the full worked example. (Figures are point-in-time snapshots from a local data lake — illustrative, not advice.)

1 · Sector analysis — "Is Indian IT a falling knife or a buying opportunity?"

sector_analysis → sector-scoped screen_stocks → financial_health on the leaders.

91 companies, ₹25.4L cr. Every major is 20–37% off its 52-week high and below its 200-DMA — a sector-wide de-rating — yet TCS still earns 63% ROCE / 52% ROE at a 14.9× P/E near its historical floor. The agent frames the one question that decides it (AI disruption: cyclical or structural?) and leaves the call to you. → full analysis

Related MCP server: IndiaQuant MCP

2 · Valuation — "Is Asian Paints still worth 57× earnings?"

valuation_summary (multiples + relative + 3-scenario DCF) cross-checked vs history + Graham Number.

All three DCF scenarios land below the market price: the ₹2,655 quote implies ~24.6% growth for a decade vs the ~10% actually delivered. High-quality business, priced for perfection. Every assumption is surfaced; the output is a range, not a "target." → full analysis

3 · SWOT — Titan, with the moat quantified

business_profile + competitive_position (VRIO-tested) + financial_health + valuation_summary.

Titan earns 2–3× the ROE/ROCE of Kalyan and Senco — a Valuable, Rare, Organised moat. The SWOT still surfaces the catches: cumulative CFO only 0.62× PAT (working-capital drain), 92% single-segment concentration, and a P/E of ~72–82. → full analysis

4 · Forensic audit — Deepak Nitrite, scores computed from raw statements

financial_health + forensic_checks + named scores computed from 12y of statements.

SCORECARD   Altman Z'' 9.84 (safe) · Piotroski 3/8 (weak) · Sloan accrual +0.15% (clean)
            Beneish M-Score: NOT COMPUTED — 4 of 8 inputs (receivables/COGS/SG&A/current-
            asset split) aren't in this data source, so the agent refuses to approximate it.

Three straight years of PAT decline, FCF negative two years, net debt swing of ₹1,971cr — but zero promoter pledge and clean accruals. The audit separates deteriorating from dishonest. → full analysis

More tools, one chart each

What it is

30 MCP tools over live market data, plus 8 research skills (dossier, forensics, SWOT, management-credibility, screening, ethics, risk-profiling) governed by a shared investing-principles rulebook (Graham / Greenblatt / Damodaran / Coffee Can / Piotroski / Altman / Sloan).

Data is fetched on demand — there is no lake to build before you can ask a question:

A local data lake still takes precedence if you have one — data_access reads it first and only falls back to the network — so a deliberately refreshed local copy stays authoritative.

On rate limits. screener.in serves one company per request and returns HTTP 429 after roughly seven consecutive fetches, so it is used only for a company actually being analysed, behind a circuit breaker. Breadth-first work (screening the universe) goes through yfinance instead, which tolerates concurrency. Nothing here bulk-scrapes.

On documents. A concall transcript is ~12k tokens, so company_documents lists a company's filings and read_document extracts one — read it directly; there is no vector index to build or keep in sync. (search_documents / topic_timeline still use an embedding index, worth building only for cross-company thematic search.)

• 🏗️ How it works, the tool list, engineering notes & setup → docs/ARCHITECTURE.md

• 📓 All worked examples → examples/

Quickstart

conda create -n finance-ai python=3.11 && conda activate finance-ai
pip install -r requirements.txt
cp .env.template .env && cp .mcp.json.example .mcp.json   # set your python path + repo cwd
python -m agent.finance_agent "Is Asian Paints still worth 57x earnings?"

No scraping step: the first question fetches what it needs. Optionally build the screening table once (yfinance, threaded) so universe-wide screens have metrics to filter on:

python -c "from agent import live; live.enrich_universe()"

Full setup in docs/ARCHITECTURE.md.

Licensing

Methodology adapts, in part, Anthropic's Apache-2.0 financial-services skills (see NOTICE). This project does not provide investment advice — see DISCLAIMER.