Anthropic Open Sourced 11 Finance Agents and the Same Skills Run Headless
The anthropics/financial-services repo is trending at 236 stars a day and sitting on 35,300 total, and it is worth looking at even if you will never build a pitch deck. It is the reference implementation for Claude for Financial Advisors, which Anthropic launched on September 14. Repo at https://github.com/anthropics/financial-services.
What is inside is eleven named agents rather than a pile of prompts. Pitch Agent and Meeting Prep on the coverage side. Market Researcher, Earnings Reviewer and Model Builder for research and modeling. Valuation Reviewer, GL Reconciler, Month-End Closer and Statement Auditor for fund admin. KYC Screener for ops. They are packaged as six vertical plugins covering financial analysis, investment banking, equity research, private equity, fund admin and operations, and you install one with a single claude plugin install command.
The architectural detail that matters more than the domain is this: the same system prompt and the same skills run in two completely different execution modes. As a Cowork plugin a human drives them interactively. Through the Claude Managed Agents API the identical definition runs headless in a backend. That is the first serious vendor artifact saying an agent definition is portable between the interactive product and the batch product, and if that holds it kills a duplication tax every enterprise team is currently paying.
The connector list is the moat, not the code. Schwab, BlackRock, Addepar, Envestnet, iCapital, Orion, SS&C Black Diamond, Wealthbox, Wealth.com, Vanguard and Zocks on the wealth side, plus FactSet, S&P Global, Morningstar and LSEG on data. Pricing is Cowork plus a free plugin, roughly seventy to a hundred twenty dollars per user per month.
Read the legal note at the top, because it is the actual product boundary. These agents draft work product for human review, they do not make recommendations, execute trades or bind risk. Every vertical agent company is going to end up writing a sentence like that, and where the sentence lands is where the value stops.
← Back to all articles
What is inside is eleven named agents rather than a pile of prompts. Pitch Agent and Meeting Prep on the coverage side. Market Researcher, Earnings Reviewer and Model Builder for research and modeling. Valuation Reviewer, GL Reconciler, Month-End Closer and Statement Auditor for fund admin. KYC Screener for ops. They are packaged as six vertical plugins covering financial analysis, investment banking, equity research, private equity, fund admin and operations, and you install one with a single claude plugin install command.
The architectural detail that matters more than the domain is this: the same system prompt and the same skills run in two completely different execution modes. As a Cowork plugin a human drives them interactively. Through the Claude Managed Agents API the identical definition runs headless in a backend. That is the first serious vendor artifact saying an agent definition is portable between the interactive product and the batch product, and if that holds it kills a duplication tax every enterprise team is currently paying.
The connector list is the moat, not the code. Schwab, BlackRock, Addepar, Envestnet, iCapital, Orion, SS&C Black Diamond, Wealthbox, Wealth.com, Vanguard and Zocks on the wealth side, plus FactSet, S&P Global, Morningstar and LSEG on data. Pricing is Cowork plus a free plugin, roughly seventy to a hundred twenty dollars per user per month.
Read the legal note at the top, because it is the actual product boundary. These agents draft work product for human review, they do not make recommendations, execute trades or bind risk. Every vertical agent company is going to end up writing a sentence like that, and where the sentence lands is where the value stops.
Comments