Building AI systems for financial services usually collapses under the weight of two conflicting realities: deterministic data integrity and the complex, messy ergonomics of institutional workflows. Financial engineering cannot tolerate hallucinated ledger entries or speculative DCF outputs. Anthropic's release of the open-source Claude for Financial Services repository cuts through this problem with an unexpected architectural choice. Instead of packaging heavy runtimes or complex Python orchestration scaffolding, the entire platform is driven by file-based system prompts, skills, commands, and standardized Model Context Protocol (MCP) data connectors.
It provides a production-grade blueprint for running specialized financial agents across local terminals, enterprise workspaces, and headless orchestrators without changing a single line of business logic.
┌────────────────────────┐
│ Anthropic Guidelines │
│ (Human-in-the-Loop) │
└───────────┬────────────┘
│
┌────────────────────────────┼────────────────────────────┐
│ │ │
▼ ▼ ▼
┌──────────────────┐ ┌──────────────────────┐ ┌──────────────────────────┐
│ Pitch / Research │ │ Valuation / Modeling │ │ Fund Admin & KYC Ops │
│ (Comps, Memos) │ │ (DCF, LBO, Excel) │ │ (GL Recs, Month-End) │
└─────────┬────────┘ └──────────┬───────────┘ └────────────┬─────────────┘
│ │ │
└─────────────────────┼──────────────────────────┘
│
▼
┌────────────────────────────────────────────────┐
│ Core Skills + Centralized MCP Connectors │
│ (FactSet, S&P, LSEG, Moody's, Morningstar...) │
└────────────────────────────────────────────────┘
The System Architecture: One Contract Across Three Runtimes
The architecture decouples the execution engine from domain knowledge. Claude for Financial Services operates on a shared codebase where prompts, conventions, and tool definitions are authored once in Markdown and JSON. You can mount these assets inside a Cowork UI session, install them directly into Claude Code, or deploy them headlessly via the Claude Managed Agents API.
The prompt engineering and tool interfaces remain identical across all deployment targets.
REPO ROOT
├── plugins/
│ ├── vertical-plugins/
│ │ ├── financial-analysis/ <-- Core modeling skills & .mcp.json connectors
│ │ │ ├── skills/
│ │ │ └── commands/
│ │ ├── investment-banking/
│ │ ├── equity-research/
│ │ ├── private-equity/
│ │ ├── wealth-management/
│ │ ├── fund-admin/
│ │ └── operations/
│ └── agent-plugins/
│ └── <slug>/ <-- Self-contained agent entry points
│ ├── agents/<slug>.md
│ └── skills/ <-- Synced from vertical source
├── managed-agent-cookbooks/ <-- agent.yaml + L1 subagents for headless execution
│ └── <slug>/
└── claude-for-msft-365-install/ <-- IT admin setup for tenant-hosted M365 add-ins
The design patterns across this directory structure eliminate build-step bloat:
| Component | Repository Path | Architectural Role |
|---|---|---|
| Agents | plugins/agent-plugins/<slug>/ | Standalone operational packages driving end-to-end workflows. |
| Skills | plugins/vertical-plugins/<vertical>/skills/ | Source domain methodologies that Claude auto-invokes during analysis. |
| Commands | plugins/vertical-plugins/<vertical>/commands/ | Explicit slash commands triggered manually by operators (e.g., /comps). |
| Connectors | plugins/vertical-plugins/financial-analysis/.mcp.json | Shared MCP definitions standardizing access to institutional data pipes. |
| Managed Wrappers | managed-agent-cookbooks/<slug>/ | agent.yaml specifications and steering logic for automated pipelines. |
To maintain consistency without runtime dependencies, domain skills live inside their respective vertical directories and synchronize to agent plugins via a simple repository script:
bashpython3 scripts/sync-agent-skills.py python3 scripts/check.py
Purpose-Built Agents: Human-in-the-Loop by Design
The platform strictly adheres to Anthropic Financial Services AI Guidelines: these agents draft artifacts; they do not finalize them. Claude will not place trades, underwrite risk, execute fund transfers, or push entries directly to a general ledger. Every output is prepared as a staged draft requiring professional sign-off.
Raw Data / Filings ──> Agent Processing ──> Staged Draft ──> Professional Review ──> Production Ledger / Pitch
The system organizes agents into domain-specific operational roles:
| Domain | Agent Slug | Operational Scope |
|---|---|---|
| Coverage & Advisory | pitch-agent | Assembles pitch decks, builds comparable analysis, and pulls precedent transactions. |
meeting-prep-agent | Generates comprehensive briefing packs ahead of client interactions. | |
| Research & Modeling | market-researcher | Builds industry overviews, landscape teardowns, and peer comp sets. |
earnings-reviewer | Parses earnings reports and filings to extract data and draft commentary. | |
model-builder | Constructs functional DCF, LBO, and three-statement models directly in Excel format. | |
| Fund Admin & Ops | valuation-reviewer | Ingests GP packages, calculates valuation sheets, and formats LP reporting. |
gl-reconciler | Tracks ledger breaks, isolates variance root causes, and routes for sign-off. | |
month-end-closer | Drafts variance commentary, processes accruals, and automates roll-forwards. | |
statement-auditor | Runs automated audit passes over LP statements prior to investor delivery. | |
| Operations | kyc-screener | Evaluates onboarding dossiers against strict rules engines and flags exceptions. |
Enterprise Data Integration: The MCP Layer
Large language models are useless in finance without low-latency access to ground-truth data. Instead of writing custom API clients for each platform, the system centralizes data access through Model Context Protocol (MCP) servers located in the core financial-analysis plugin.
┌────────────────────────┐
│ Claude Workspace │
└───────────┬────────────┘
│
plugins/.../.mcp.json
│
┌────────────────────────┼────────────────────────┐
▼ ▼ ▼
┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐
│ Market Data │ │ Financial Filings│ │ Document Storage │
│ (FactSet, LSEG, │ │ (Daloopa, S&P, │ │ (Egnyte, │
│ Morningstar) │ │ Moody's, Aiera) │ │ Chronograph) │
└──────────────────┘ └──────────────────┘ └──────────────────┘
The MCP configuration connects Claude to major financial platforms through standardized endpoints:
| Provider | MCP Server Endpoint | Core Domain Focus |
|---|---|---|
| Daloopa | https://mcp.daloopa.com/server/mcp | Fundamental financial modeling and historic data. |
| Morningstar | https://mcp.morningstar.com/mcp | Fund metrics, equities, and managed portfolios. |
| S&P Global | https://kfinance.kensho.com/integrations/mcp | S&P Capital IQ data, market summaries, and tear sheets. |
| FactSet | https://mcp.factset.com/mcp | Market analytics, estimates, and institutional data. |
| Moody's | https://api.moodys.com/genai-ready-data/m1/mcp | Credit research, risk assessment, and ratings. |
| MT Newswires | https://vast-mcp.blueskyapi.com/mtnewswires | Real-time financial breaking news streams. |
| Aiera | https://mcp-pub.aiera.com | Earnings calls, transcripts, and corporate events. |
| LSEG | https://api.analytics.lseg.com/lfa/mcp | Bond valuation, swap curves, option vol, and macro rates. |
| PitchBook | https://premium.mcp.pitchbook.com/mcp | Private equity metrics, VC tracking, and M&A data. |
| Chronograph | https://ai.chronograph.pe/mcp | Private capital portfolio metrics and fund management. |
| Egnyte | https://mcp-server.egnyte.com/mcp | Enterprise document stores and secure data rooms. |
Installing and Deploying Across Infrastructures
The repository structure makes it easy to move between a local terminal, a shared workspace, and cloud infrastructure.
1. Claude Code CLI Installation
Add the repository marketplace directly to Claude Code and install the core plugins and selected agents:
bash# Add marketplace repository claude plugin marketplace add anthropics/claude-for-financial-services # Core skills and data connectors (mandatory dependency) claude plugin install financial-analysis@claude-for-financial-services # Install domain agents claude plugin install pitch-agent@claude-for-financial-services claude plugin install gl-reconciler@claude-for-financial-services claude plugin install market-researcher@claude-for-financial-services # Install domain vertical skill extensions claude plugin install investment-banking@claude-for-financial-services claude plugin install equity-research@claude-for-financial-services
2. Headless API Deployment
For continuous deployment pipelines or scheduled automation runs, use the provisioning script to deploy managed agents directly to the Anthropic API:
bashexport ANTHROPIC_API_KEY=sk-ant-... scripts/deploy-managed-agent.sh gl-reconciler
3. Microsoft 365 Enterprise Add-in Integration
Enterprise security policies often require running models through your own cloud tenancy rather than public consumer endpoints. The claude-for-msft-365-install/ directory provides an administrative configuration tool. It provisions the Claude add-in for Word, Excel, PowerPoint, and Outlook, routing requests through private enterprise infrastructure on Amazon Bedrock, Google Vertex AI, or internal LLM gateways.
bashclaude plugin install claude-for-msft-365-install@claude-for-financial-services /claude-for-msft-365-install:setup
This configuration tool handles Azure admin consent, generates the custom add-in manifest, and writes per-user routing rules using the Microsoft Graph API.
Customizing the Stack
The repository is built to be customized. You can tailor it to your firm's private operations with simple modifications:
- Update MCP Endpoints: Point
.mcp.jsonto your internal data warehouses, valuation microservices, or private document stores. - Inject Proprietary Styles: Add custom presentation rules using
/ppt-templateto enforce internal brand standards. - Refine Domain Skills: Edit the Markdown files under
plugins/vertical-plugins/<vertical>/skills/to codify internal accounting rules, DCF conventions, and investment committee memos. - Scope Agent Boundaries: Modify
agents/<slug>.mdto match your firm's specific review gates and risk parameters.
Anthropic distributes the entire framework under the Apache License 2.0. You can inspect the source, adapt the skills, and integrate it with your existing market data feeds directly via the GitHub repository: https://github.com/anthropics/financial-services.