Anthropic’s official open-source reference stack for FSI workflows: 10 named agents, 7 vertical skill bundles, 11 MCP data connectors, and Managed-Agent cookbooks. One source, two deployment paths — install as Cowork/Claude Code plugins, or deploy headlessly via /v1/agents. Same system prompt, same skills.
Architecture
Three-layer split worth copying:
- Agents (
plugins/agent-plugins/<slug>/) — self-contained workflow plugins. Each bundles its own skills. End-user surface. - Vertical plugins (
plugins/vertical-plugins/<vertical>/) — source-of-truth skills, slash commands, and.mcp.jsonconnectors. Authored once, synced into agents. - Managed-agent cookbooks (
managed-agent-cookbooks/<slug>/) —agent.yaml+ depth-1 leaf subagents + steering examples. Same prompt, headless harness.
Sync script (scripts/sync-agent-skills.py) propagates skill edits from verticals into agents. Lint script (scripts/check.py) validates cross-references before push. This is the missing piece in most agent codebases.
The 10 Agents
| Vertical | Agents |
|---|---|
| Coverage & advisory | Pitch Agent, Meeting Prep Agent |
| Research & modeling | Market Researcher, Earnings Reviewer, Model Builder |
| Fund admin & finance ops | Valuation Reviewer, GL Reconciler, Month-End Closer, Statement Auditor |
| Operations & onboarding | KYC Screener |
Each ships with a Managed-Agent counterpart deployable via scripts/deploy-managed-agent.sh <slug>.
Vertical Skill Bundles
- financial-analysis (core, includes
.mcp.json) —/comps,/dcf,/lbo,/3-statement-model,/debug-model,pptx-author,xlsx-author,ppt-template-creator,skill-creator - investment-banking —
/one-pager,/cim,/teaser,/buyer-list,/merger-model,/process-letter,/deal-tracker - equity-research —
/earnings,/earnings-preview,/initiate,/model-update,/morning-note,/sector,/thesis,/catalysts,/screen - private-equity —
/source,/screen-deal,/dd-checklist,/dd-prep,/unit-economics,/returns,/ic-memo,/portfolio,/value-creation,/ai-readiness - wealth-management —
/client-review,/financial-plan,/rebalance,/client-report,/proposal,/tlh - fund-admin — reconciliations, accruals, roll-forwards, NAV tie-out
- operations — KYC document parsing, rules-grid eval
MCP Connectors (financial-analysis/.mcp.json)
11 hosted MCPs, all subscription-gated: Daloopa, Morningstar, S&P Global (Kensho/Capital IQ), FactSet, Moody’s, MT Newswires, Aiera, LSEG, PitchBook, Chronograph, Egnyte. Partner-built plugins from LSEG (bond RV, swap curves, FX carry, vol, macro-rates) and S&P Global (tear sheets, earnings previews, funding digests) live under plugins/partner-built/.
Excel / PowerPoint Mechanics
Two paths:
- Live in M365 add-in —
claude-for-msft-365-install/is an admin plugin that generates the manifest, supports Anthropic API / Vertex / Bedrock / internal gateway as the backend, grants Azure admin consent, configures per-user via Graph. - Headless authoring —
pptx-authorandxlsx-authorskills generate.pptx/.xlsxfiles inside Managed Agents (no Office runtime).ppt-template-creatorteaches Claude branded layouts.
Why This Matters — HIGH
This is the public reference architecture for bespoke FSI agent work. A few concrete takeaways:
1. Skill design pattern is now canonical. The three-layer split (agent bundle → vertical source → managed-agent cookbook) plus sync-agent-skills.py solves the “skill drift across agents” problem cleanly. A strong layout to adopt for any serious agent deliverable — both as a portability story and because it’s now what sophisticated FSI buyers will expect.
2. The PE vertical covers a lot of deal workflow. /source, /screen-deal, /dd-checklist, /ic-memo, /portfolio, /value-creation — Anthropic shipped underwriting and acquisition workflows as installable plugins. The differentiation can’t be “we built a deal screener.” It has to be a proprietary data layer plus the human ops loop.
3. The MCP connector list is a competitive map. Daloopa, FactSet, Capital IQ, Moody’s, PitchBook, LSEG, Morningstar — these are the Bloomberg-replacement seats Anthropic is courting. Notable absences: no Tegus, no AlphaSense, and no CRE-specific data (CoStar, Reonomy, Placer). The CRE gap is an opening — a vertical the official repo does not cover.
4. M365 install plugin matters more than it looks. It targets the IT admin, not the analyst. That’s a distribution wedge into the enterprise — same playbook as Tableau and Notion.
Risks / Tradeoffs
- Sync script + check.py + orchestrate.py is real engineering — not a weekend port. Estimate ~2 weeks to adapt the pattern into another stack.
- Adopting the Anthropic skill format locks you to Anthropic’s plugin/Managed-Agent runtime. Hedge: keep the skill markdown portable; abstract the runtime.
- All connectors are subscription-gated. The agents are useful only with paid data behind them — same Bloomberg dynamic. Confirm the seats exist before promising the workflow.
Source
- Repo:
anthropics/financial-services - Date inspected: 2026-05-14
part of tooling