Anthropic Financial Services - Reference Agents, Skills and Connectors

Anthropic’s financial-services repository is a set of reference implementations for finance work: “Reference agents, skills, and data connectors for the financial-services workflows we see most — investment banking, equity research, private equity, and wealth management.” It is Apache-2.0 licensed, and its construction is deliberately plain: “Everything is file-based — markdown and JSON, no build step.”

Read the disclaimer first, because it defines the design

Most repositories bury this. Here it is load-bearing, and worth quoting at length because it describes what the agents are allowed to do:

“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 is not a liability shield bolted on afterwards; it is the boundary the whole collection is built around. Every agent produces a draft that a human signs. Compare that with TradingAgents, which simulates a decision all the way to an approved order, or dexter, which researches autonomously — this repository deliberately stops one step earlier, at the artefact rather than the action.

The ten agents

Each is a self-contained plugin under plugins/agent-plugins/:

Coverage and advisory — Pitch Agent (comps, precedents and an LBO through to a branded deck), Meeting Prep Agent (briefing pack before a client meeting).
Research and modelling — Market Researcher (sector to industry overview, landscape, peer comps, shortlist), Earnings Reviewer (call plus filings to model update to note draft), Model Builder (DCF, LBO, three-statement and comps, live in Excel).
Fund admin and finance ops — Valuation Reviewer, GL Reconciler (finds breaks, traces root cause, routes for sign-off), Month-End Closer, Statement Auditor.
Operations — KYC Screener (parses onboarding documents, runs the rules engine, flags gaps).

Underneath sit vertical plugins bundling skills and slash commands — /comps, /dcf, /lbo, /cim, /teaser, /merger-model, /earnings, /ic-memo, /screen-deal, /portfolio among them — plus partner-built plugins from LSEG and S&P Global.

The data connectors are the catch

The MCP server list is the part a firm will actually evaluate: Daloopa, Morningstar, S&P Global, FactSet, Moody’s, MT Newswires, Aiera, LSEG, PitchBook, Chronograph, Egnyte and Box.

And the honest note alongside it: “MCP access may require a subscription or API key from the provider.” That is the real cost structure. The agents are free and Apache-licensed; the data they are interesting against is not. A firm that already pays for FactSet or S&P gets leverage here; one that does not is looking at the connector list as a shopping bill.

One small discrepancy worth reporting, found by counting rather than reading: the page states the core plugin has “All 11 data connectors,” while the table beneath it lists twelve. Trivial, but it is the kind of drift the repository’s own scripts/check.py exists to catch.

Two deployment paths

Everything ships “two ways from one source” — as a Claude Cowork plugin, or through the Claude Managed Agents API behind your own workflow engine.

For Claude Code:

claude plugin marketplace add anthropics/financial-services
claude plugin install financial-analysis@claude-for-financial-services
claude plugin install pitch-agent@claude-for-financial-services

For managed agents:

export ANTHROPIC_API_KEY=sk-ant-...
scripts/deploy-managed-agent.sh gl-reconciler

Note that subagent delegation (callable_agents) is flagged as “a preview capability” — relevant if you were planning to compose agents rather than run them individually.

Contributors run python3 scripts/check.py before pushing; it lints manifests, verifies cross-file references and detects skill drift.

Who should look at this

Firms in the named verticals, and specifically the person who has to decide whether agent work belongs anywhere near regulated output. The value is less in any single agent than in the pattern: file-based definitions, no build step, every output staged for sign-off, and a stated refusal to let the agent act rather than draft.

If you are building your own finance agents, read this before designing them. The boundary it draws — draft, never execute — is the part worth copying, independent of whether you ever install a single plugin.