What This Actually Is
Anthropic has published its reference agents for financial services under Apache-2.0. Ten named agents covering investment banking, equity research, private equity, fund administration and onboarding, plus the skills, slash commands and data connectors underneath them.
It passed 36,000 stars, the largest repository covered in this series, and it is worth being precise about what it contains. This is not a trading system and not a model. It is a plugin marketplace: markdown skill files, JSON connector configs and system prompts that run inside Claude Cowork or Claude Code.
A note on the GitHub language label, which says Python. The substance here is markdown and JSON with no build step. The Python is a handful of helper scripts for deployment and validation. If you clone this expecting a codebase, you will find a very well-organised set of prompts.
Nothing on this page is investment, legal, tax or accounting advice.
What These Agents Will Not Do
This deserves to come before the feature list, because the gap between how tools like this get described and what they actually do is where people get hurt.
Anthropic puts the limits in a callout near the top of the README rather than in a footer:
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.
Read that list again, because each item is a specific refusal. The KYC Screener does not approve onboarding, it flags gaps for someone who can. The GL Reconciler does not post to the ledger, it traces breaks and routes them for sign-off. The Model Builder does not tell you what a company is worth, it builds the model you then argue with.
So the honest framing is drafting, not deciding. That is a genuinely large amount of work in finance, because the first draft of a comps table or a reconciliation is hours of tedium with very little judgement in it. It is also not the part anyone is paid for.
If you came here expecting agents that trade, this is not that, and the repository is explicit that it is not.
The Ten Agents
Each agent is a self-contained plugin bundling the skills it needs, so installing one is all you do.
| Function | Agent | What it drafts |
|---|---|---|
| Coverage and advisory | Pitch Agent | Comps, precedents and an LBO into a branded pitch deck |
| Meeting Prep Agent | A briefing pack before a client meeting | |
| Research and modeling | Market Researcher | Industry overview, competitive landscape, peer comps, idea shortlist |
| Earnings Reviewer | Earnings call and filings into a model update and note draft | |
| Model Builder | DCF, LBO, three-statement and comps, live in Excel | |
| Fund admin | Valuation Reviewer | Ingests GP packages, runs the valuation template, stages LP reporting |
| GL Reconciler | Finds breaks, traces root cause, routes for sign-off | |
| Month-End Closer | Accruals, roll-forwards, variance commentary | |
| Statement Auditor | Audits LP statements before distribution | |
| Operations | KYC Screener | Parses onboarding docs, runs the rules grid, flags gaps |
The fund administration cluster is the most interesting and the least discussed. Month-end close and GL reconciliation are unglamorous, enormously time-consuming, and far more rule-bound than pitch work, which makes them a better fit for an agent than the deal-making tasks that get the attention.
Anthropic frames these as starting points to tune rather than finished products, which is the right expectation. A comps convention that matches your firm's house style is not something a generic prompt can know.
Skills, Commands and Connectors
The architecture separates four things, and understanding the split is what makes the repo navigable.
- Agents own a workflow end to end: a system prompt plus the skills it uses.
- Skills are domain expertise and step-by-step methods Claude draws on automatically when relevant. Authored once per vertical, then synced into each agent that needs them.
- Commands are slash actions you fire deliberately:
/comps,/dcf,/earnings,/ic-memo. - Connectors are MCP servers wiring Claude to actual data.
The skill-sync detail has a practical consequence. Skills live once in the vertical plugins and each agent bundles a copy, so when you tune a skill for your firm you are editing a bundled copy that upstream will later change. A text diff between your edited version and the refreshed upstream one is the least painful way to keep your customisations when you pull updates.
Nine vertical plugins are available if you want the pieces without a full agent: financial-analysis as the core, then investment-banking, equity-research, private-equity, fund-admin, operations, a financial-advisors bundle, and two partner-built plugins from LSEG and S&P Global.
Start with financial-analysis. It carries the shared modelling skills and all the connectors.
Free Software, Expensive Everything Else
This is the section that matters most, and it is the same lesson as every other free finance tool: the software is the cheap part.
The repository is Apache-2.0 and costs nothing. Then look at what it plugs into. Twelve MCP connectors ship in the core plugin:
| Category | Providers |
|---|---|
| Market and fundamental data | FactSet, S&P Global via Kensho, Morningstar, Daloopa, LSEG |
| Credit and ratings | Moody's |
| News and transcripts | MT Newswires, Aiera |
| Private markets | PitchBook, Chronograph |
| Documents | Box, Egnyte |
The README notes, in one quiet line, that "MCP access may require a subscription or API key from the provider."
That line is carrying a great deal of weight. FactSet, PitchBook, LSEG and S&P Capital IQ are enterprise terminals costing thousands to tens of thousands per seat per year. These are not services with a free tier you sign up for on a Saturday.
So the stack is: free prompts, a paid Claude subscription for Cowork or Claude Code, API spend if you deploy through Managed Agents, and enterprise data subscriptions that dwarf both.
Which tells you exactly who this is for. It is aimed at firms that already pay for FactSet and PitchBook, and for whom the agents are a way to get more out of seats they already hold. If you are an individual who wanted free Wall Street tooling, the prompts are genuinely free and interesting to read, and the data they expect is not available to you at any sensible price.
That is not a criticism of the repository. It is the structure of the industry, and the same wall any finance tool hits.
Two Ways to Run It
The same source serves both deployment paths, which is the neatest engineering idea in here.
As a plugin, inside Claude Cowork or 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-servicesAgents then appear in dispatch, skills fire when relevant, and the slash commands are available in your session.
As a managed agent, headless behind your own workflow engine:
export ANTHROPIC_API_KEY=sk-ant-...
scripts/deploy-managed-agent.sh gl-reconcilerThe deploy script resolves file references, uploads the skills, creates leaf-worker subagents and posts the orchestrator to the Managed Agents API. Same system prompt, same skills, different runtime.
Each managed agent is configured through an agent.yaml alongside its subagent definitions, and if you are hand-editing those a YAML formatter catches the indentation mistakes that otherwise surface as a deployment failure with an unhelpful message.
One flag worth noting: subagent delegation through callable_agents is labelled a research preview. Anthropic points to per-agent security notes for handoff guidance. Treat multi-agent handoff as the least settled part of this and read those notes before wiring it into anything that matters.
It Is Anthropic's Own Repo, and That Cuts Both Ways
Worth saying plainly rather than leaving as subtext.
This is published by Anthropic, the company that sells Claude, and it runs exclusively on Anthropic's paid products. It is a genuine open-source resource and a commercial on-ramp at the same time. Both things are true and neither cancels the other.
What that buys you. These prompts come from a team with direct visibility into how large financial firms actually deploy Claude. The workflow selection is not guesswork, and the refusals baked into the system prompts reflect compliance conversations most independent projects never get to have. That is real, hard-won content and it is unusual to see it published.
What to be alert to. Every path leads back to a Claude subscription or Claude API spend. There is no provider-agnostic version, and the skills are written for Claude's behaviour rather than as portable prompt engineering. That is fair enough for a first-party repo, and it does mean this is not a neutral survey of how to do finance with AI.
The Microsoft 365 tooling is the one part that cuts the other way. It provisions the Claude add-in against your own cloud, Vertex AI or Bedrock or an internal gateway, rather than Anthropic's API. That is a real concession to enterprises that will not send data to a vendor endpoint, and it is a point in the repo's favour.
Read the skill files even if you never install any of it. They are among the better public examples of writing domain expertise into a form a model can use.
Making It Yours
Anthropic is direct that these are templates rather than products, and the customisation surface is where the value actually lands.
- Swap connectors by pointing
.mcp.jsonat your own providers and internal systems. - Add firm context by dropping your terminology, processes and formatting standards into the skill files.
- Bring your templates so
/ppt-templateteaches Claude your branded deck layouts. - Adjust agent scope by editing the agent markdown to match how your team runs the workflow.
- Add your own workflows by copying the structure.
The first two matter most. A comps skill that does not know your firm's convention on calendarisation or your treatment of stock-based compensation will produce a table your MD sends back. Encoding those conventions once, in a file, is the actual work, and it is the part no vendor can do for you.
Financial data arrives as CSV more often than anything else, and a CSV to JSON converter is a quick way to reshape an export into something a connector or skill expects. When you are sanity-checking the margin and multiple arithmetic these agents produce, a markup and margin calculator is faster than opening a spreadsheet to verify one number.
Who This Is Actually For
A strong fit if you work at a firm that already pays for Claude and at least one major data provider, you are in one of the covered workflows, and you have someone who can tune prompts and own the output. The fund administration agents in particular target work that is genuinely repetitive and genuinely expensive in headcount.
A poor fit if you expected free access to financial data, or if you need an agent that decides rather than drafts, or if you are in a jurisdiction or function where staging output for human review does not resolve your regulatory position. The disclaimer puts responsibility for verification and compliance squarely on you, and that is not boilerplate in this industry.
Read it anyway if you are interested in how to structure agent skills. Ten workflow agents built on shared skills, with an explicit split between what is bundled and what is authored once, is a more mature pattern than most agent repositories demonstrate, and it transfers to domains that have nothing to do with finance.
The honest summary: this is a high-quality reference implementation, published by the vendor whose products it requires, aimed at professionals who already hold the expensive subscriptions. The prompts are free and worth reading. The stack around them is not free, and the agents deliberately stop short of the decisions. Both facts are stated clearly in the repository, which is more than most of this category manages.
