Starting point for financial AI that actually matches how the work gets done
Connecting Claude to financial workflows sounds straightforward — until you try to wire together pitch prep, GL reconciliation, equity research, and KYC screening into something a real team can use. The system prompts, skill files, data connectors, and agent handoff logic add up fast. anthropics/financial-services is Anthropic's reference template for that architecture, built around the workflows that actually show up in investment banking, equity research, private equity, and fund operations.
The problem it solves
The design cost is the real barrier. Connecting a single DCF model to Claude isn't hard. Building a coherent system where a Pitch Agent runs Comps → Precedents → LBO → branded deck end to end — and a GL Reconciler traces breaks → root cause → sign-off routing — requires upfront architectural decisions that most teams shouldn't have to make from scratch.
This repo bundles 10 named agents, vertical skill packs for 5 FSI segments, and 11 MCP data connectors into a file-based reference you can fork, install, and tune to your firm's conventions.
How it works
Three layers run together.
Agents own a workflow end to end. Each agent plugin is self-contained — it bundles the skills it uses, so installing the agent is all you need. The pitch-agent runs the full pitch workflow; the gl-reconciler handles break detection through sign-off routing.
Skills encode domain expertise and step-by-step methods as markdown and JSON files — no build step. The vertical plugins hold the canonical versions; each agent bundles a synced copy of the skills it needs.
Connectors are 11 MCP servers centralized in the financial-analysis core plugin: FactSet, PitchBook, Morningstar, S&P Global, Moody's, LSEG, Daloopa, MT Newswires, Aiera, Chronograph, and Egnyte. Connect the providers you already subscribe to.
The same system prompt and skills run two ways: as a Cowork plugin for team-facing deployment, or via the Claude Managed Agents API behind your own workflow engine. One source, two execution paths.
Setup
For Claude Code:
# Register the marketplace
claude plugin marketplace add anthropics/financial-services
# Core skills and connectors — install this first
claude plugin install financial-analysis@claude-for-financial-services
# Add the agents you need
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
# Or install vertical skill bundles without full agents
claude plugin install investment-banking@claude-for-financial-services
claude plugin install equity-research@claude-for-financial-services
For Managed Agent deployment:
export ANTHROPIC_API_KEY=sk-ant-...
scripts/deploy-managed-agent.sh gl-reconciler
The deploy script resolves file references, uploads skills, creates leaf-worker subagents, and POSTs the orchestrator to /v1/agents.
For Cowork, go to Settings → Plugins → Add plugin, paste https://github.com/anthropics/financial-services, and pick the agents and verticals you want.
Real scenarios
Scenario 1 — IB team preparing pitch materials
Install pitch-agent and you get /comps, /dcf, /lbo, /one-pager, and /cim available immediately in your session. The agent runs Comps → Precedents → LBO → branded deck in sequence. Every output is staged as a draft for human review — the agent does not execute transactions or finalize materials.
Scenario 2 — Fund ops team running month-end GL reconciliation
claude plugin install gl-reconciler@claude-for-financial-services
The gl-reconciler finds breaks, traces root cause, and routes for sign-off. The fund-admin vertical bundle adds accruals, roll-forwards, variance commentary, and NAV tie-out on top.
Trade-offs and limits
The repo is explicit about scope: these agents draft analyst work product for review by a qualified professional. They do not make investment recommendations, execute transactions, bind risk, post to a ledger, or approve onboarding. Every output is staged for human sign-off.
That means this is not the right fit for teams that need AI to publish outputs without review, or for workflows where regulatory judgments need to flow through without a human checkpoint.
The 11 MCP connectors require subscriptions or API keys from each provider. The repo is open; the data access is not.
The subagent delegation feature (callable_agents) is currently in Research Preview. Check each agent's README for security and handoff guidance before production deployment.
Alternatives
Build from scratch — Full control over every prompt and skill file, but the upfront cost of encoding financial domain conventions is real. This repo is most useful as a starting point that absorbs that cost, not as a locked-in dependency.
Vendor-bundled financial AI (Bloomberg AI, Kensho, etc.) — Data and UI come pre-integrated, which lowers time to first use. The trade-off is limited ability to swap connectors, customize skill logic, or route through your own orchestration layer. financial-services is designed to be forked and tuned, not consumed as-is.
Where to go from here
The real value of financial-services is reducing design cost, not replacing design judgment. Skill files are markdown — edit them to match your firm's terminology, process steps, and formatting standards. Swap .mcp.json entries to point at your internal data systems. Adjust agents/<slug>.md to match how your team actually runs the workflow.
A natural next step: combine pitch-agent with the equity-research vertical to keep research-to-pitch consistency across a coverage workflow — the shared skill layer is what makes that coherent rather than duplicated.
🐦 Faster updates on X: @baegseungh7061
📚 More in this series: AI Insights
💌 Subscribe: Follow on X or grab the RSS
댓글
댓글 쓰기