Claude for Financial Services - Anthropic's Reference Agents vs TradingAgents

We just covered TradingAgents — UCLA’s open multi-agent research framework that models a trading firm as Python + LangGraph, debates trades among LLM agents, and lands a decision.

This week Anthropic shipped its own answer to “AI in finance”: anthropics/financial-services.

Same domain, opposite design philosophy.

Where TradingAgents is an academic research framework that makes decisions, Claude for FS is an enterprise reference content pack that drafts work product for human sign-off.

Both are Apache 2.0; one is what an open-source research lab ships, the other is what the model vendor ships.

The contrast is worth understanding.

What is Claude for Financial Services?

Claude for Financial Services is Anthropic’s official open-source reference for FSI workflows. It ships 10 named agents (Pitch Agent, Market Researcher, GL Reconciler, KYC Screener, …), 6 vertical skill bundles (investment-banking, equity-research, private-equity, wealth-management, fund-admin, operations), and 11 commercial MCP data connectors (Daloopa, Morningstar, FactSet, Moody’s, PitchBook, LSEG, S&P Global, and more). Everything is markdown + JSON; no build step.

“Nothing in this repository constitutes investment, legal, tax, or accounting advice. These agents draft analyst work product — models, memos, research notes, reconciliations — 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 disclaimer is the entire design philosophy in one paragraph. Hold onto it.

Anthropic Financial Services on GitHub Claude Cowork Claude Managed Agents API docs

What makes it different

• 🏢 10 workflow-named agents — Pitch Agent, Meeting Prep, Market Researcher, Earnings Reviewer, Model Builder, Valuation Reviewer, GL Reconciler, Month-End Closer, Statement Auditor, KYC Screener
• 🧰 6 vertical skill bundles — investment-banking, equity-research, private-equity, wealth-management, fund-admin, operations (plus 2 partner-built: LSEG, S&P Global)
• 🔌 11 commercial MCP connectors — Daloopa, Morningstar, S&P Global, FactSet, Moody’s, MT Newswires, Aiera, LSEG, PitchBook, Chronograph, Egnyte
• 🪞 Two deployment paths from one source — Cowork plugin (interactive UI) OR Managed Agents API (headless); same system prompts and skill files, different wrappers
• 📝 Drafts, never decides — every output is staged for human sign-off; agents don’t execute trades, post to ledgers, or approve onboarding
• 🔧 Markdown + JSON only — no build step, no Node toolchain; python3 scripts/check.py lints manifests and verifies cross-file references
• 📦 Microsoft 365 add-in install tooling — separate Claude Code plugin that provisions the add-in against your own cloud (Vertex AI, Bedrock, internal LLM gateway)
• ⚖️ Apache 2.0 — fully open source

The TradingAgents comparison (the headline)

Both projects are Apache-licensed multi-agent systems applied to finance. They’re at opposite ends of nearly every design axis. Honest landscape:

The honest read:

• TradingAgents is the artifact — open framework, multi-provider, runs anywhere, makes trading decisions for research evaluation
• Claude for FS is the product enablement — vendor-locked, paid-tier data, runs on Anthropic platforms, drafts work for human review

They occupy completely different positions:

• Pick TradingAgents if you’re studying multi-agent LLM debate dynamics, want a template for building your own vertical multi-agent app, or want to run the whole stack against local Ollama models
• Pick Claude for FS if you’re an FSI firm already on Cowork / Anthropic API, you have commercial subscriptions to the data providers (Daloopa, Morningstar, FactSet, PitchBook), and you want production-grade reference workflows for analyst tasks
• Or run both — TradingAgents for market hypothesis exploration on the research side, Claude for FS for the actual deck / model / memo authoring on the production side

The deeper observation: these two projects illustrate the two patterns we’re going to see across every vertical as AI applications mature:

1. Open research framework — academic team writes a paper, ships open code, multi-provider, decision-making, self-hostable. (TradingAgents in finance; could be MedAgents in medicine, LegalAgents in law, etc.)
2. Vendor reference content — model provider ships markdown + skills + connectors that surface their platform’s capability in a specific vertical. (Claude for FS in finance; Anthropic could ship Claude for Legal, Claude for Healthcare, etc.)

These don’t compete — they’re at different layers of the stack.

Where Claude for FS fits in the AI stack

Recent posts on this site have mapped a five-layer AI tooling stack. Claude for FS sits at a new layer I haven’t drawn before — vendor-provided application content on top of the vendor’s runtime products:

Vendor vertical content is genuinely a different shape from everything we’ve covered. It’s not infrastructure (you can’t run Cowork yourself), it’s not a personal assistant, it’s not a coding agent. It’s enterprise enablement content shipped open-source so that a customer of the vendor’s commercial product can deploy industry-specific workflows on day one rather than authoring them from scratch.

The same pattern is likely to emerge for:

• Claude for Legal (contract review, citation checking, deposition prep)
• Claude for Healthcare (clinical note drafting, prior-auth packets)
• Claude for Engineering (design reviews, spec compliance, traceability)

Watch for those.

How It Actually Works

Two deployment paths, one source

        ┌───────────────────────────────────────────────┐
        │  plugins/agent-plugins/<slug>/                │
        │  ───────────────────────────                  │
        │  agents/<slug>.md       ← canonical sys-prompt │
        │  skills/                ← bundled skill copies │
        └────────┬─────────────────────────┬───────────┘
                 │                         │
                 │  same files, two wrappers
                 ▼                         ▼
        ┌────────────────────┐   ┌────────────────────────────────┐
        │  Claude Cowork     │   │  Claude Managed Agents API      │
        │  (interactive UI)  │   │  (headless / programmatic)      │
        └────────────────────┘   └────────────────────────────────┘

The agent.yaml under managed-agent-cookbooks/<slug>/ references the same files as the Cowork plugin. Customers can run the same Pitch Agent inside Cowork interactively, or call it via /v1/agents programmatically from their workflow engine. Switching modes doesn’t require maintaining two copies.

Skill authoring discipline

Skills are authored once in vertical-plugins/ and synced into agent bundles via python3 scripts/sync-agent-skills.py. The agent bundles contain synced copies (not symlinks) because plugin install needs each agent to be self-contained.

check.py lints every manifest, verifies all cross-file references resolve, and fails if any agent bundle’s skill copy has drifted from its vertical-plugin source. This drift check is the integrity glue holding the two-source-one-truth model together — without it, you’d inevitably get bundles with stale skill copies and out-of-sync Cowork/Managed Agent behavior.

Self-Hosting Claude for Financial Services

This isn’t traditional self-hosting — you can’t run Cowork on your own server, and the Managed Agents API runs on Anthropic’s infrastructure. But the content is open and there’s a genuine self-hosting angle for the Microsoft 365 routing path.

Cowork install (the consumer path)

In Cowork: Settings → Plugins → Add plugin. Either paste the repo URL https://github.com/anthropics/financial-services and pick from the marketplace list, or upload a zip of any directory under plugins/.

Claude Code install

claude plugin marketplace add anthropics/financial-services

# Core skills + connectors (install first)
claude plugin install financial-analysis@claude-for-financial-services

# Named agents — pick what you want
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

# Vertical bundles
claude plugin install investment-banking@claude-for-financial-services
claude plugin install equity-research@claude-for-financial-services

After install, agents appear in Cowork dispatch, skills fire automatically when relevant, slash commands available in your session (/comps, /dcf, /earnings, /ic-memo, …).

Claude Managed Agents (headless 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 multi-agent orchestration with handoffs between agents, see scripts/orchestrate.py — a reference event loop that routes handoff_request events.

claude plugin install claude-for-msft-365-install@claude-for-financial-services
/claude-for-msft-365-install:setup

Conclusion

Claude for Financial Services is what vendor-provided enterprise reference content looks like in 2026: open-source on GitHub, dual-deployment (interactive Cowork + headless Managed Agents API), 11 commercial MCP integrations, and explicit scope limits that the agents draft work product rather than make decisions. Pair it with TradingAgents on the research side and you have the full picture of how AI is showing up in finance — the open research framework that explores what’s possible, and the vendor-backed production content that ships what’s safe.

For self-hosters on this site, the actual self-hostable surface is narrow — the MSFT 365 add-in routing path lets you point Claude inside Office at your own cloud (Vertex / Bedrock / internal gateway), which is the real story for FSI firms with data-residency constraints. Everything else runs on Anthropic’s platforms.

Related tools worth knowing:

• TradingAgents — the academic counterpart; opposite design philosophy in nearly every dimension
• FinRL — reinforcement-learning finance, different paradigm
• Model Context Protocol — the protocol all 11 connectors speak
• OpenClaw / NanoClaw / Hermes Agent — personal assistant runtimes; can also consume MCP servers like the 11 in this repo
• Claude Cowork — the SaaS product Claude for FS plugs into