Anthropic open-sources Claude finance agents, skills, and MCP connectors

What Anthropic released

Anthropic has published an open-source project called Claude for Financial Services: a repository of reference agents, reusable skills, and data connectors covering the financial workflows the company says it sees most often — investment banking, equity research, private equity, and wealth management. According to a dev.to post describing the release, the code lives in the anthropics/financial-services repository on GitHub, is written primarily in Python, and is licensed under Apache 2.0.

The notable architectural choice is that every component is available two ways from one source. The same agent can be installed as a Claude Cowork plugin for interactive use, or deployed through the Claude Managed Agents API behind a firm's own workflow engine. The system prompt and skills are identical in both cases; the operator decides where it runs. The dev.to post also stresses that this is a file-based reference implementation — markdown and JSON with no build step — rather than a hosted product, so teams can read it end to end, fork it, and tune it to their own conventions.

Ten agents across four functions

The repository ships ten named agents, each owning a workflow end to end and bundling the skills it needs. Grouped by function, according to the post:

• Coverage and advisory: a Pitch Agent that runs comps, precedents, and LBO analysis through to a branded pitch deck, and a Meeting Prep Agent that assembles client briefing packs.
• Research and modeling: a Market Researcher that turns a sector or theme into an industry overview, competitive landscape, peer comps, and an ideas shortlist; an Earnings Reviewer that takes an earnings call plus filings through a model update to a note draft; and a Model Builder that produces DCF, LBO, three-statement, and comps models in Excel.
• Fund admin and finance ops: a Valuation Reviewer that ingests GP packages and stages LP reporting, a GL Reconciler that traces breaks to root cause and routes them for sign-off, a Month-End Closer for accruals, roll-forwards, and variance commentary, and a Statement Auditor for LP statements.
• Operations and onboarding: a KYC Screener that parses onboarding documents, runs a rules engine, and flags gaps.

Vertical plugins and slash-command skills

Below the agents sit vertical plugins grouped by line of business. The core financial-analysis plugin carries the shared modeling skills and the data connectors, and the README says to install it first. On top of it sit investment-banking (CIMs, teasers, merger models, deal tracking), equity-research (earnings notes, initiations, catalyst tracking), private-equity (sourcing, diligence checklists, IC memos), fund-admin (GL reconciliation, NAV tie-out), and operations (KYC parsing and rules-grid evaluation), plus a claude-for-financial-advisors plugin for advisor workflows such as meeting prep, compliance pre-check, and rebalance review.

The skill layer is where the repository gets concrete. The dev.to post lists slash commands including /comps for trading-multiple analysis, /dcf with WACC and sensitivity analysis, /lbo, /3-statement-model for populating three-statement templates, and /debug-model, an Excel audit skill that does formula tracing, hardcode detection, and balance checks. Equity research adds /earnings and /earnings-preview plus initiation reports and a catalyst calendar; private equity adds IRR/MOIC sensitivity tables and /ic-memo for investment committee drafts.

Twelve connectors, and one documentation wrinkle

All of the data connectors are MCP servers, centralized in the core plugin and shared across the others, wiring Claude to terminals, research platforms, and document stores. The README's table lists twelve providers: Daloopa, Morningstar, S&P Global, FactSet, Moody's, MT Newswires, Aiera, LSEG, PitchBook, Chronograph, Egnyte, and Box. The dev.to post flags an inconsistency — the README text refers to "all 11 data connectors" while its own table lists twelve — and a more consequential caveat: MCP access may still require a subscription or API key from each provider. Pre-configured connectors do not mean free data.

Guardrails and vendor-reported numbers

Every agent is positioned as drafting analyst work product — models, memos, research notes, reconciliations — for review by a qualified professional. The repository states that the agents do not make investment recommendations, execute transactions, bind risk, post to a ledger, or approve onboarding; every output is staged for human sign-off. Because everything is plain files, a compliance officer can read a system prompt the way they would read a procedure document.

The post also relays Anthropic's own claims from its financial services announcement: that Claude Opus 4 passed five of seven levels of the Financial Modeling World Cup with 83% accuracy on complex Excel tasks, and that Claude 4 models outperform other frontier models as research agents on Vals AI's Finance Agent benchmark. These are vendor-reported figures about Anthropic's own models.

Why it matters

The bottleneck for AI in finance has rarely been raw capability — comps, reconciliations, and KYC checks are structured, repetitive, expensive work — but whether a risk committee will sign off on the implementation. An inspectable, Apache-licensed reference stack with explicit no-execution guardrails and mandatory human review is aimed squarely at that approval problem. It also gives MCP connectivity to a dozen enterprise data providers a standardized, forkable home, which could shorten integration work for firms already licensed for those feeds. The caveats are worth keeping in view: the benchmark numbers are Anthropic's own, the connector documentation disagrees with itself on the count, and access to the underlying data still costs money.