Anthropic has published Claude for Financial Services, an open-source repository of reference agents, skills, and data connectors for the financial-services workflows it sees most often: investment banking, equity research, private equity, and wealth management. The project lives at github.com/anthropics/financial-services, is written primarily in Python, and ships under the Apache License 2.0. What makes it unusual is that everything in it is available two ways from one source: you can install the same agent as a Claude Cowork plugin, or deploy it through the Claude Managed Agents API behind your own workflow engine. Same system prompt, same skills, you choose where it runs.
If you work in finance and you have been waiting for a concrete, inspectable example of what financial services AI looks like in production shape, this is the most useful thing Anthropic has released for you. It is not a demo and it is not a product you have to buy. It is a file-based reference implementation, markdown and JSON with no build step, that you can read cover to cover, fork, and tune to how your firm actually works.
Why this repository matters right now
Financial analysis is full of work that is structured, repetitive, and expensive: building comparable company analyses, updating models after earnings calls, reconciling general ledgers, screening KYC documents. These tasks follow known methods, which makes them good candidates for AI finance workflows, but they also sit inside regulated firms where a black-box answer is not acceptable. As a result, the question for most teams has not been whether AI can do the work, but whether it can do the work in a way a risk committee will sign off on.
Anthropic’s answer with this repository is twofold. First, every agent drafts analyst work product, models, memos, research notes, reconciliations, for review by a qualified professional. The repository states plainly that the agents do not make investment recommendations, execute transactions, bind risk, post to a ledger, or approve onboarding; every output is staged for human sign-off. Second, the whole thing is plain files. There is no hidden service doing something you cannot inspect. A compliance officer can read the system prompt of the KYC Screener the same way they would read a procedure document.
The timing also reflects a broader push. Anthropic’s own financial services announcement reports that Claude Opus 4 passed 5 out of 7 levels of the Financial Modeling World Cup competition with 83% accuracy on complex Excel tasks, and that Claude 4 models outperform other frontier models as research agents across financial tasks in Vals AI’s Finance Agent benchmark. Those are Anthropic’s claims about its own models, so treat them as vendor-reported figures, but they explain why the company felt confident publishing reference workflows rather than just marketing copy.
How the repository is organized
The repository splits its content into a few clear layers, and understanding the split is the key to using it well. At the top sit ten named agents, each one owning a workflow end to end. Underneath them are vertical plugins that bundle skills, slash commands, and data connectors by line of business. Alongside both are cookbooks for headless deployment and admin tooling for Microsoft 365.
Ten named agents for financial analysis workflows
Each agent is named for the workflow it runs, and each ships as a self-contained plugin that bundles the skills it uses, so installing the agent is all you need. The ten agents group into four functions:
Coverage and advisory: the Pitch Agent runs comps, precedents, and LBO analysis through to a branded pitch deck, and the Meeting Prep Agent assembles a briefing pack before every client meeting.
Research and modeling: the Market Researcher turns a sector or theme into an industry overview, competitive landscape, peer comps, and an ideas shortlist; the Earnings Reviewer takes an earnings call plus filings through a model update to a note draft; the Model Builder produces DCF, LBO, three-statement, and comps models live in Excel.
Fund admin and finance ops: the Valuation Reviewer ingests GP packages, runs the valuation template, and stages LP reporting; the GL Reconciler finds breaks, traces root cause, and routes items for sign-off; the Month-End Closer handles accruals, roll-forwards, and variance commentary; the Statement Auditor audits LP statements before distribution.
Operations and onboarding: the KYC Screener parses onboarding documents, runs the rules engine, and flags gaps.
Because each agent is a starting point rather than a finished product, the expectation is that you install the ones matching your work and then tune the prompts, skills, and connectors to your firm’s conventions.
Vertical plugins: skills and commands by business line
If you do not want a full agent, you can install the underlying vertical plugins on their own and just get the slash commands and connectors. The core plugin is financial-analysis, which carries the shared modeling skills and the data connectors, and the README says to install it first. On top of that sit investment-banking (CIMs, teasers, process letters, buyer lists, merger models, deal tracking), equity-research (earnings notes, initiations, model updates, thesis and catalyst tracking), private-equity (sourcing, screening, diligence checklists, IC memos, portfolio monitoring), fund-admin (GL recon, break tracing, accruals, roll-forwards, variance commentary, NAV tie-out), and operations (KYC document parsing and rules-grid evaluation). There is also a claude-for-financial-advisors plugin covering advisor workflows such as meeting prep, compliance pre-check, prospect intake, and rebalance review.
The skill-level detail is where the repository gets concrete. The financial-analysis plugin alone includes skills for comparable company analysis with trading multiples (/comps), discounted cash flow valuation with WACC and sensitivity analysis (/dcf), leveraged buyout modeling (/lbo), populating three-statement financial model templates (/3-statement-model), and an Excel model audit that does formula tracing, hardcode detection, and balance checks (/debug-model). The equity-research plugin adds earnings call analysis through /earnings and /earnings-preview, plus initiation reports, morning notes, and a catalyst calendar. For private equity AI work, the private-equity plugin covers deal sourcing, screening, diligence checklists, unit economics, IRR/MOIC sensitivity tables, and investment committee memo drafting through /ic-memo.
Financial services MCP data connectors
All of the data connectors are centralized in the financial-analysis core plugin and shared across the rest. They are MCP servers that wire Claude to terminals, research platforms, and document stores. The README’s connector table lists twelve providers: Daloopa, Morningstar, S&P Global, FactSet, Moody’s, MT Newswires, Aiera, LSEG, PitchBook, Chronograph, Egnyte, and Box. One wrinkle worth knowing: the README describes the core plugin as carrying “all 11 data connectors” while its own table lists twelve, so the documentation is not internally consistent on the count. In practice, what matters is the list itself, and the caveat the README attaches to it: MCP access may require a subscription or API key from the provider. These are enterprise data services, and the connectors being pre-configured does not mean the data is free.
Two ways to run the same agent
The dual-track design is the architectural idea at the heart of the repository, and it is worth understanding before you install anything.
Installing as Claude Cowork plugins
The interactive track is Cowork. In Cowork, you open Settings, then Plugins, then Add plugin, and either paste the repository URL (https://github.com/anthropics/financial-services) and pick agents and verticals from the marketplace list, or zip any directory under plugins/ and upload it directly. Once installed, agents appear in Cowork dispatch, skills fire automatically when relevant, and slash commands such as /comps, /dcf, /earnings, and /ic-memo become available in your session.
Installing through Claude Code
For Claude Code, the README gives the exact commands:
# Add the marketplace
claude plugin marketplace add anthropics/financial-services
# Core skills + connectors (install first)
claude plugin install financial-analysis@claude-for-financial-services
# Named agents — pick the ones 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 skill bundles
claude plugin install investment-banking@claude-for-financial-services
claude plugin install equity-research@claude-for-financial-servicesNote that the marketplace name after the @ is claude-for-financial-services, which is not the same as the repository name. Getting that string wrong is the most common install mistake.
Deploying through the Claude Managed Agents API
The headless track uses the Claude Managed Agents API. Each template under managed-agent-cookbooks/ references the same system prompt and skills as its plugin counterpart, and deployment is a two-line affair:
export ANTHROPIC_API_KEY=sk-ant-...
scripts/deploy-managed-agent.sh gl-reconcilerThe deploy script resolves file references, uploads skills, creates leaf-worker subagents, and POSTs the orchestrator to /v1/agents. The repository also includes scripts/orchestrate.py, a reference event loop that routes handoff_request events between agents through your own orchestration layer. However, one capability here is explicitly flagged as early: subagent delegation (callable_agents) is a research preview, and the per-agent READMEs carry security and handoff guidance for it.
What the agents actually do in practice
It helps to walk through a couple of workflows end to end, because the value is in the chaining rather than in any single skill.
Claude agents for investment banking workflows
Take the Pitch Agent as the clearest example of investment banking automation. It runs comparable company analysis, pulls precedent transactions, builds an LBO, and assembles the results into a branded pitch deck, end to end. The investment-banking vertical underneath it supplies the individual pieces: /one-pager for strip profiles, /cim for drafting Confidential Information Memorandums, /teaser for anonymous one-pagers, /buyer-list for the strategic and financial buyer universe, /merger-model for accretion/dilution analysis, and /deal-tracker for live deal milestones. Similarly, the /ppt-template command teaches Claude your firm’s branded PowerPoint layouts, so the output deck looks like your deck rather than a generic one.
Equity research AI: from earnings call to published note
The Earnings Reviewer shows the research loop. It ingests the earnings call and filings, updates the model, and drafts the note. The equity-research plugin’s skills cover the surrounding cadence of a coverage desk: pre-earnings scenario analysis, post-earnings quarterly updates, initiations, morning notes, thesis tracking, and a catalyst calendar. Meanwhile, the Model Builder handles the AI-powered DCF and LBO financial modeling side directly in Excel, which matters because Excel is where these models actually live in most firms.
Month-end close and KYC screening
On the operations side, the GL Reconciler finds breaks, traces the root cause, and routes items for sign-off, while the Month-End Closer produces accruals, roll-forwards, and variance commentary. The KYC Screener parses onboarding documents and runs a rules-grid evaluation, flagging gaps for a human to resolve. In each case the pattern is the same: the agent does the assembly and the first pass, and a person makes the decision.
Partner plugins and the Microsoft 365 add-in
Two partner-built plugins extend the core set. The LSEG plugin covers bond relative value, swap curves, FX carry, options vol, and macro-rates monitoring on LSEG data, and the S&P Global plugin produces tear sheets, earnings previews, and funding digests on S&P Capital IQ. Both live under plugins/partner-built/ and follow the same file-based structure as the rest.
Separately, the repository includes claude-for-msft-365-install/, admin tooling for firms that run Claude inside Excel, PowerPoint, Word, and Outlook through the Microsoft 365 add-in. It is a Claude Code plugin, not a Cowork plugin, and it walks an IT admin through generating the customized add-in manifest, granting Azure admin consent, and writing per-user routing config via Microsoft Graph. Notably, it provisions the add-in against your own cloud, Vertex AI, Bedrock, or an internal LLM gateway, instead of Anthropic’s API. Install it with:
claude plugin install claude-for-msft-365-install@claude-for-financial-services
/claude-for-msft-365-install:setupMaking the templates yours
Anthropic is explicit that these are reference templates that get better when you tune them. The suggested customization paths are practical: swap connectors by pointing .mcp.json at your own data providers and internal systems, drop your firm’s terminology and formatting standards into skill files, teach Claude your branded PowerPoint layouts with /ppt-template, and edit agents/<slug>.md to match how your team actually runs each workflow.
Contributing back follows the same file-based logic. New skills go under the relevant vertical’s skills/ directory, and python3 scripts/sync-agent-skills.py propagates them to any agent that bundles them. Before pushing, python3 scripts/check.py lints every manifest, verifies that all cross-file references resolve, and fails if any bundled skill has drifted from its vertical source. That drift check is a small detail, but it is the kind of thing that keeps a multi-plugin repository coherent as it grows.
Limitations and what to watch
A few honest caveats. First, the legal framing is not decorative: nothing in the repository is investment, legal, tax, or accounting advice, and your firm is responsible for verifying outputs and for regulatory compliance. Second, the connectors are only as useful as your subscriptions; a team without FactSet or PitchBook contracts will need to swap in what it actually has. Third, subagent delegation is a research preview, so headless multi-agent deployments deserve extra scrutiny before they touch real client work.
It is also worth keeping the benchmark numbers in perspective. Anthropic reports strong results for its models on financial tasks, and its financial services page cites compliance postures such as SOC 2 and FedRAMP, but neither the repository nor those pages publish pricing for the plugins, and there is no independent measurement yet of productivity impact across adopting firms. The fairest reading is that the repository gives you the machinery; the business case still has to be made inside your own workflow.
People Also Ask
How do I install the financial-services plugins in Claude Cowork?
From the Cowork interface, go to Settings, then Plugins, then Add plugin. You can either paste the repository URL, https://github.com/anthropics/financial-services, and select the agents and verticals you want from the marketplace list, or zip any directory under plugins/ and upload it directly. Install the financial-analysis core plugin first, since it carries the shared skills and all the data connectors.
How do I install the financial-services marketplace in Claude Code?
Run claude plugin marketplace add anthropics/financial-services, then install plugins with commands like claude plugin install financial-analysis@claude-for-financial-services. The marketplace name after the @ is claude-for-financial-services, spelled exactly like that, and the slash commands appear in a new session once installed.
What is the difference between a Claude Cowork plugin and a Managed Agent?
They are two deployment tracks for the same agent. The Cowork plugin runs interactively in your session, while the Managed Agent template deploys headlessly through the /v1/agents API behind your own workflow engine, using agent.yaml, leaf-worker subagents, and steering events. Both reference the same system prompt and skills from the same directory.
Do the MCP data connectors include access to the data providers?
No. The connectors for providers such as FactSet, Moody’s, LSEG, and PitchBook are pre-configured in the financial-analysis plugin, but the repository notes that MCP access may require a subscription or API key from the provider. You can also point .mcp.json at your own data sources instead.
The bottom line
Claude for Financial Services is best read as a reference architecture rather than a product. Its ten agents, six vertical plugins, partner integrations, and Managed Agent cookbooks show, in inspectable files, how Anthropic thinks AI finance workflows should be built: human sign-off on every output, schema-disciplined handoffs, and one source that runs both interactively and headlessly. For a finance team evaluating financial services AI, the fastest way to form an opinion is to clone the repository, install the financial-analysis core, and run /comps on a company you know well. What you see in that output will tell you more than any benchmark table.
Additional Resources
This article includes content created with AI.