Building AI Workflows for Finance Without Code: Inside Anthropic’s Plugin Architecture
Hook
What if you could wire Claude to Bloomberg data, generate a 50-page pitch book, and build a three-statement model—all from a chat interface, without writing a single line of Python?
Context
Financial services firms have a workflow problem that generic AI assistants can’t solve. An equity research analyst doesn’t just need to “ask questions about a company”—they need to pull comps from FactSet, build a DCF model following firm-specific color conventions, generate sensitivity tables, export to Excel with proper formatting, and create a PowerPoint deck that matches house style. This requires orchestrating multiple proprietary data sources, embedding domain expertise, and producing publication-ready deliverables that meet regulatory and client expectations.
Traditional approaches fail at different points. Custom-coded solutions give you control but require engineering teams to maintain integrations with dozens of financial data APIs. Generic LLM wrappers lack the domain knowledge and can’t produce properly formatted financial models. Bloomberg and FactSet offer AI features, but they’re siloed within their platforms and can’t orchestrate end-to-end workflows. Anthropic’s financial-services-plugins attempts to solve this through a novel architecture: a file-based plugin system that extends Claude with financial capabilities using only markdown and JSON configuration files, powered by the Model Context Protocol (MCP) for data connectivity.
Technical Insight
The architecture separates concerns across three layers: skills (passive domain knowledge), commands (explicit user-triggered actions), and MCP connectors (external data integrations). This design means financial professionals can customize workflows without engineering resources, while maintaining clean separation between generic financial logic and firm-specific processes.
A plugin is just a directory structure with markdown files. Here’s what a simplified custom command might look like:
# commands/build_comps.md
name: Build Comparable Companies Analysis
trigger: /comps
parameters:
- ticker (required)
- sector (optional)
## Workflow
1. Use the FactSet MCP connector to retrieve:
- Current trading multiples (EV/EBITDA, P/E, P/B)
- Last twelve months financials
- Three-year CAGR for revenue and EBITDA
2. Identify 8-10 comparable public companies in the same sector
3. Build comparison table following house style:
- Sort by market cap descending
- Use firm standard formatting ($ in millions, multiples to 1 decimal)
- Highlight subject company in yellow
4. Calculate median and mean for each multiple
5. Export to Excel with:
- Blue font for formulas
- Black font for hardcoded data
- Data validation on ticker inputsThis markdown file becomes a reusable workflow. When a user types /comps AAPL, Claude orchestrates the entire sequence—no Python glue code required. The “blue font for formulas” detail matters enormously in finance; models that don’t follow these conventions get rejected by senior bankers.
The Model Context Protocol is where this gets powerful. MCP is Anthropic’s standard for connecting LLMs to data sources. Instead of hardcoding API calls, you declare connections in JSON:
{
"mcp_servers": {
"factset": {
"type": "mcp",
"module": "factset-mcp-server",
"credentials": "${FACTSET_API_KEY}",
"capabilities": [
"company_fundamentals",
"market_data",
"ownership_data"
]
},
"daloopa": {
"type": "mcp",
"module": "daloopa-mcp-server",
"credentials": "${DALOOPA_API_KEY}",
"capabilities": [
"detailed_models",
"historical_financials"
]
}
}
}Claude can now reference these connectors in workflows. The base “financial analysis” plugin ships with 11 MCP integrations—FactSet, S&P Capital IQ, Daloopa, LSEG, and others. The add-on plugins (investment banking, private equity, wealth management) don’t duplicate these connections; they just define domain-specific workflows that use the shared data layer.
This composability is the key architectural insight. The private equity plugin doesn’t need to know how to fetch financial data—it just declares “retrieve the last five years of EBITDA” and the MCP layer handles provider selection, authentication, caching, and error handling. If your firm switches from FactSet to Bloomberg, you update one JSON file rather than rewriting plugins.
The skills layer adds passive domain knowledge that’s always available. A skills/financial_modeling_standards.md file might contain:
# Financial Modeling Standards
When building Excel models:
- Blue font (#0000FF): formulas and calculations
- Black font (#000000): hardcoded historical data
- Green font (#00FF00): management guidance or assumptions
- Never merge cells in calculation areas
- Always include a version number and "last updated" date
- Sensitivity tables: inputs in left column, outputs across top row
- Use named ranges for key assumptionsClaude ingests this as context and automatically applies these conventions when generating models. You’re not prompting it every time to “use blue font for formulas”—it’s baked into the plugin’s knowledge.
Partner-built plugins are particularly interesting. LSEG (London Stock Exchange Group) built their own plugin that exposes Workspace data through MCP. S&P Global did the same for Capital IQ. This creates a marketplace dynamic where data providers maintain their own integrations rather than waiting for Anthropic to build them. For enterprises, this means you can build internal plugins for proprietary data—your deal database, your firm’s valuation models, your CRM—using the same architecture.
Gotcha
The enterprise-only licensing is a hard gate. This requires Claude for Enterprise (Cowork or Code subscription), which means individual developers, small RIAs, or boutique advisory firms are completely locked out. There’s no degraded free tier or hobbyist version—you either have an enterprise agreement with Anthropic or you can’t use any of this.
The data dependency stack is equally restrictive. The plugins are designed around premium financial data providers that charge institutional rates. FactSet subscriptions start around $20,000 annually per user. Daloopa is similarly expensive. S&P Capital IQ runs $30,000-$40,000. A firm could easily spend $100,000+ annually on data subscriptions before getting value from these plugins. The repository doesn’t include fallback integrations to free data sources like Yahoo Finance or FRED, so without the paid data, most commands simply won’t work. There’s also limited transparency about customization boundaries. The documentation shows you can modify commands and add skills, but it’s unclear whether you can fundamentally alter the execution model, add custom post-processing logic, or integrate with systems outside the MCP framework. The file-based approach is elegant for simple workflows but may hit walls for complex requirements like “send this model to our compliance system for review before delivering to the client” or “apply our proprietary risk adjustment to all DCF valuations.”
Verdict
Use if: You’re at a financial services firm (investment bank, private equity shop, asset manager) with Claude Enterprise licensing and existing subscriptions to at least two of the supported data providers like FactSet, Daloopa, or S&P Capital IQ. The workflow orchestration genuinely collapses hours of context-switching into single chat sessions, and the file-based customization means your analysts can encode firm-specific processes without engineering bottlenecks. This is transformative for standardizing junior analyst output and scaling institutional knowledge. Skip if: You’re an individual practitioner, at a smaller firm without enterprise AI budgets, or lack access to premium financial data sources. Also skip if you need deep programmatic control—the no-code approach is a feature for financial professionals but a limitation for developers who want to add complex business logic, custom integrations beyond MCP, or programmatic workflow orchestration. In those cases, you’re better off building custom agents with LangChain or LlamaIndex where you control the entire execution model.
