“"I'm not paying $125k/year/seat to get locked into a black box ecosystem that might change." — OldfieldFund, HN thread https://news.ycombinator.com/item?id=44576312 (accessed 2026-05-07)”
You know that feeling when you spend four hours building a comparable company analysis in Excel, pulling multiples from Bloomberg, formatting slides, and then the deal dies anyway? Financial analysts at banks, PE firms, and hedge funds repeat that cycle daily — comps, DCF models, earnings recaps, GL reconciliations, KYC reviews — each a multi-hour template task that requires domain knowledge but follows a predictable structure. Existing tools automate narrow slices: a Bloomberg terminal for data, PitchBook for deal history. None of them draft the actual work product end-to-end. This repo treats each workflow as a composable Claude agent: give it the right data connectors and domain prompts, and it drafts the output for a human to approve.
Each agent is a markdown file telling Claude how to behave for a specific finance workflow — think of it as a detailed job description in plain text. When you install an agent via Claude Code with one command, it also wires up MCP connectors to financial data providers like FactSet or PitchBook, so Claude can pull live market data when you run /comps or /dcf. A Python script called sync-agent-skills.py keeps skills synchronized across agents automatically, and check.py validates all manifest links before you deploy. You run the same agent two ways: as a plugin in the Claude Cowork web UI for interactive use, or as a headless Managed Agent via /v1/agents for automated pipelines. Every output is staged for human review — the agent drafts, a qualified professional approves.
If you are an engineer at a bank, asset manager, insurance firm, or fintech building internal AI workflows, this is months of prompt engineering and MCP connector wiring already done for you — fork it, adapt the markdown files to your firm's terminology, and deploy. Finance professionals who want to use it themselves need Claude Code and Anthropic API access, plus existing data provider subscriptions for the MCP connectors. Not useful if you need agents that execute autonomously — by design, every output requires explicit human sign-off, and the managed-agent architecture caps subagent depth ...
Yes, if your firm already pays for Anthropic API access and holds subscriptions to FactSet, PitchBook, or similar data providers — the 10 agent templates and 11 pre-built MCP connector definitions alone save significant integration time. The single-source-of-truth architecture is also worth studying for any enterprise AI deployment where consistency between interactive and automated environments matters. Be cautious: Claude Opus 4.7 scores 64.37% on the Vals AI Finance benchmark (roughly a 36% failure rate on benchmark tasks), and HN practitioners reported hallucination in financial document citation workflows — the repo's own disclaimer mandates human sign-off for every output.
Deep-dive insight, Easy and Pro modes, plus action playbooks — the full breakdown is one tap away.