Anthropic just open-sourced Wall Street (11.4k stars)

0:01 3 days ago Anthropic dropped a repo on GitHub that might end Wall Street's monopoly on its own analyst workflow. Anthropic, the company that makes Claude, quietly shipped a repo called Claude for financial services. 3 days ago they restructured it from a generic skills bundle into 10 named end-to-end agents with job titles you would see on a Goldman analyst's business card, names like pitch agent, market researcher, earnings reviewer, model builder, GL reconciler, and KYC screener. The repo is Apache 2.0, hit 11,400 stars, and trended number one in Python the day it shipped. 11 licensed MCP data connectors are wired in, including SP Global FactSet PitchBook Moody's and LSEG. Every named agent ships its own skills

0:47 bundle and deploys two ways. Let me show you what's inside. Stars on the repo climbing fast since the restructure named agents that map to specific seats at a bank, licensed connectors plugging into the terminal data, and vertical bundles, one for each part of the firm. Okay, so the read me and the framing matters. The page opens with a disclaimer that none of this constitutes investment advice, and that every output is staged for human sign-off. The agents draft, the humans approve. That positioning is what lets Anthropic ship something this aggressive without sounding like they are replacing the analyst. The agent table is organized by function, grouping coverage and advisory at the top, then research and modeling, then the fund admin and finance ops cluster, then operate.

1:33 Below that, the vertical plugin section lists seven bundles, the core financial analysis bundle that ships shared modeling skills, plus six verticals running from investment banking and equity research through private equity, wealth management, and the back office stack. Two partner-built plugins also ship, one from LSEG and one from S&P Global. For 20 years the analyst workflow has been a closed shop. To build a pitch deck or run a DCF, you needed a Bloomberg terminal at $25,000 a year, an Excel template owned by a bulge bracket bank, a CapIQ subscription, and 3 years of unpaid weekends learning the modeling conventions. None of those vendors ever shipped the methodology. The methodology lived inside the banks.

2:19 What Anthropic just published is the entire skill ladder, the workflow itself, the prompts and conventions, the audit checklist, and the plumbing that lets a properly licensed seat actually pull the data once you have it. Forks the community has spun up since launch, commits sitting on main from years of compounding work, two partner-built plugins shipped alongside the agents, and how recently the restructure landed on the main branch. Okay, the named agents. Pitch agent runs comps, presidents, and an LBO end-to-end, outputs a branded pitch deck, the full thing a banking analyst would build over a weekend. Meeting prep generates a briefing pack before client meetings, pulling from your CRM, the news desk, and recent filings. Market researcher takes a sector or theme and

3:05 produces an industry overview, competitive landscape, peer comps, and a short list of investment ideas. Earnings reviewer ingests the call audio, the press release, and the 10-Q, then updates the model and drafts the research note. Model builder runs DCF, three statement, and comps live in Excel, hitting XL author headlessly when in managed agent mode. The fund admin and operations side is where this gets interesting because nobody usually open-sources middle office workflows. GL reconciler finds breaks in the general ledger, traces the root cause, roots for sign-off. Month-end closer handles accruals, roll forwards, and variance commentary. Valuation reviewer ingests GP packages, runs the firm's valuation template,

3:52 stages LP reporting statement auditor, audits LP statements before distribution. KYC screener parses onboarding documents, runs the rules engine, flags compliance gap. Every one of these is something a junior accountant or a compliance analyst was getting paid 80,000 a year to do 6 months ago. Under the hood, the architecture is what other indie builders should study. Every named agent is a self-contained cloud code plugin, a folder with an agent.md file declaring the system prompt, a skills folder with a bundled markdown files, a commands folder with a slash commands, and a .mcp.json declaring which connectors that agent is allowed to call. Skills are authored once in the vertical plugins, then synced into each agent by

4:37 a script called sync_agent_skills.py. There is a check.py that lints manifests. A deploy_managed_agent.sh that pushes the same agent to the /v1/agents API, and an orchestrate.py showing the reference event loop for routing handoff requests between agents. The whole thing is markdown and JSON, no build step. Install on cloud code is three commands: add the marketplace, install the financial analysis core, which carries the data connectors, cherry-pick whichever agents and verticals match your work. Pitch agent for IB coverage, TL reconciler for fund admin, wealth for the advisor stack.

5:22 Each agent bundles its own skills, so there is no manual dependency tracking. The May 5th restructure commit is what flipped this repo. Every vertical moved from the repo root into a plugins/vertical plugins folder. The 10 named agent plugins landed under plugins/agent plugins, each one self-contained. The team added managed agent cookbooks with one folder per agent, including sub-agent isolation and steering examples for headless deployment. They added a fund admin and operations vertical, so the finance ops and onboarding agents shipped with real domain skills. They tightened tool grants. No agent can shell out to bash, hit web fetch, or call an undeclared connector. And they added a GitLeaks based secret scan workflow because everyone using this is

6:09 going to be holding licensed terminal credentials. The tradeoffs are real. The connectors all require our license seats, and those subscriptions still cost real money. So, you cannot clone this repo and walk away with a turnkey terminal stack out of the box. The agents are reference templates, not finished product, so you have to tune the prompt and swap in your firm's data sources before any of this runs in production. But if you build software for the financial industry in 2026, this is the new ceiling for what an open reference implementation looks like, and the conversation with your enterprise prospects just changed. They can spin up the analyst stack themselves and compare it to whatever you're charging for. Links in the description. See you tomorrow.