2026-05-10

Anthropic launches Finance Agents — 10 vertical templates, 64.37% on Vals AI benchmark, customers include Citadel and BNY

Anthropic ships 10 finance agent templates + governed data connectors (Moody's MCP, D&B). Claude Opus 4.7 leads Vals AI Finance benchmark at 64.37%.

Anthropic shipped its Finance Agents launch this week — 10 ready-to-run templates plus a slate of governed data connectors. The combination is the first concrete vertical agent cookbook library under the Managed Agents umbrella, and it’s the reference shape for how Anthropic wants third-party vertical builders (legal, healthcare, ops, security) to package their work.

The 10 templates

Template	What it does
Pitch builder	Drafts buy/sell pitches from data + research
Meeting preparer	Pre-call briefs from CRM + market data
Earnings reviewer	Tags 10-K / 10-Q against checklists
Model builder	Generates DCF, comps, sensitivity tables
Market researcher	Cross-source market sizing and trend
Valuation reviewer	Audits a third-party model for inconsistencies
General ledger reconciler	Matches transactions across systems
Month-end closer	Runs the close checklist end-to-end
Statement auditor	Cross-checks financial statement consistency
KYC screener	Sanctions / PEP / adverse-media screening

Each template ships three ways: as plugins inside Claude Cowork, as plugins inside Claude Code, and as cookbooks for Claude Managed Agents. The cookbook format is the load-bearing artifact — it’s the template definition, the system prompts, the tool list, the data-connector configuration, and the eval suite.

The data connectors

New governed data connectors include:

Moody’s MCP app (600M+ entities, the largest single source)
D&B (Dun & Bradstreet)
Fiscal AI
Financial Modeling Prep
Guidepoint
IBISWorld
SS&C Intralinks
Third Bridge
Verisk

The “governed” part matters. Each connector ships with auditable access logs, data-residency controls, and PII handling that satisfy financial regulator (SEC, FCA, MAS) compliance reviews. This is the part that small startups can’t replicate — it’s the relationship and legal work, not the technology.

Benchmark numbers

Claude Opus 4.7 leads the Vals AI Finance Agent benchmark at 64.37%, with reported customers including Citadel, FIS, BNY, Carlyle, Mizuho, Travelers, Walleye, and Hg. The 64.37% headline number sounds modest — finance agent tasks are graded strictly because mistakes have real financial-instrument consequences. Second-place models trail by 4-7 points.

Why this matters beyond finance

The template-as-plugin pattern is the reference for every vertical agent library Anthropic wants to ship. The structural elements are:

Cookbook-as-source-of-truth — the cookbook defines the agent and gets versioned in git
Three deployment surfaces — Cowork plugin, Claude Code plugin, Managed Agents cookbook
Governed data connectors — vertical-specific data sources with compliance baked in
Vertical-specific eval suite — tied to a benchmark that customers actually trust

If you’re building a vertical Claude plugin (legal, healthcare, ops, security), study the cookbook structure before shipping. The expected packaging shape just got concrete.

Practitioner note

Two things to do this week: (1) Read the cookbook for one of the 10 templates closely — the General Ledger Reconciler is the most generally instructive because the pattern (match-records-across-systems with explicit reconciliation rules) generalizes beyond finance. (2) If you’re targeting any regulated vertical, the data-connector list is your clue about what governance work is required to compete. You can’t replicate Moody’s MCP overnight, but you can identify the analogous data sources in your own vertical and start the integration conversations now. For non-finance verticals, the template-as-plugin architecture is the right shape to copy: cookbook + three surfaces + governed data + eval suite.