FinanceHarness 0.1.0

FinanceHarness 0.1.0

FinanceHarness is an autonomous financial deep research framework. Ask a
question; FinanceHarness plans the work, gathers evidence from the web and from
market and company data, runs the analysis, and writes a cited report.

Architecture

• Agent loop. Plan, act, observe over a constant tool registry. Execution
  modes (auto, research, analytical) are prompt variants, so switching never
  strands a run.
• Progressive disclosure. A small always-on core (web research, calc,
  planning, load_tool/load_skill); everything else waits in a catalog and
  loads on demand.
• Reference-chaining. One tool's output feeds the next by reference
  (prev:<call_id>.<path>), with no retyping.
• Model-agnostic seam. gemini (default), gpt, and self-hosted qwen
  behind one provider interface.
• Grounded. Every figure traces to the tool or source it came from, with an
  optional self-grounding review.

Tools and skills

• Web research (search, visit, compose_citations), equity and market data
  via yfinance, valuation (DCF, DCF-sensitivity, WACC), and risk (VaR, beta,
  correlation).
• Five composable SKILL.md workflows: ticker-snapshot, dcf-valuation,
  relative-valuation, consensus-check, equity-deep-dive.

Usage

uv tool install .             # installs `financeharness` (alias `fh`)
export GEMINI_API_KEY=...      # or OPENAI_API_KEY for --profile gpt
fh -p "Estimate NVDA's intrinsic value with a DCF."

An optional HTTP+SSE service is included (fh serve).