financial-services Documentation

File: README.md Copy Chapter

Claude for Financial Services

Reference agents, skills, and data connectors for the financial-services workflows we see most — investment banking, equity research, private equity, and wealth management.

Everything here is available two ways from one source: install it 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.

[!IMPORTANT]
Nothing in this repository constitutes investment, legal, tax, or accounting advice. 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. You are responsible for verifying outputs and for compliance with the laws and regulations that apply to your firm.

What's in the repo:

• Agents — named, end-to-end workflow agents (Pitch Agent, Market Researcher, GL Reconciler, …). Each ships as a Cowork plugin and as a Claude Managed Agent template you deploy via /v1/agents.
• Vertical plugins — the underlying skills, slash commands, and data connectors, bundled by FSI vertical. Install these on their own if you just want /comps, /dcf, /earnings and the connectors without a full agent.

Agents Copy Chapter

Each agent is named for the workflow it runs. They're starting points: install the ones that match your work, then tune the prompts, skills, and connectors to how your firm does it.

Each agent plugin is self-contained — it bundles the skills it uses, so installing the agent is all you need.

For Managed Agent deployment — agent.yaml, leaf-worker subagents, steering-event examples, and per-agent security notes — see managed-agent-cookbooks/.

Repository Layout Copy Chapter

plugins/
  agent-plugins/               # Named agents — one self-contained plugin each
  vertical-plugins/            # Skill + command bundles by FSI vertical, plus MCP connectors
  partner-built/               # Partner-authored plugins (LSEG, S&P Global)
managed-agent-cookbooks/       # Claude Managed Agent cookbooks — one dir per agent
claude-for-msft-365-install/   # Admin tooling to provision the Claude Microsoft 365 add-in
scripts/                       # deploy-managed-agent.sh · check.py · validate.py · orchestrate.py · sync-agent-skills.py

Getting Started Copy Chapter

Cowork Copy Section

In Cowork, open Settings → Plugins → Add plugin and either:

• Paste this repo URL — https://github.com/anthropics/financial-services — then pick the agents and verticals you want from the marketplace list, or
• Upload a zip — zip any directory under plugins/ (e.g. plugins/agent-plugins/pitch-agent/) and drop it in.

Claude Code Copy Section

# 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-services

Once installed, agents appear in Cowork dispatch, skills fire automatically when relevant, and slash commands are available in your session (/comps, /dcf, /earnings, /ic-memo, …).

Claude Managed Agents Copy Section

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

Each template under managed-agent-cookbooks/ references the same system prompt and skills as its plugin counterpart. The deploy script resolves file references, uploads skills, creates leaf-worker subagents, and POSTs the orchestrator to /v1/agents. See scripts/orchestrate.py for a reference event loop that routes handoff_request events between agents via your own orchestration layer.

Research Preview: subagent delegation (callable_agents) is a preview capability. See per-agent READMEs for security and handoff guidance.

How It Fits Together Copy Chapter

Everything is file-based — markdown and JSON, no build step.

Vertical Plugins Copy Chapter

Start with financial-analysis — it carries the shared modeling skills and all data connectors. Add verticals for the workflows you need.

MCP Integrations Copy Chapter

All connectors are centralized in the financial-analysis core plugin and shared across the rest.

MCP access may require a subscription or API key from the provider.

Claude for Microsoft 365 — Install Tooling Copy Chapter

If your firm runs Claude inside Excel, PowerPoint, Word, and Outlook via the Microsoft 365 add-in, claude-for-msft-365-install/ is the admin tooling to provision it against your own cloud — Vertex AI, Bedrock, or an internal LLM gateway — instead of Anthropic's API.

It's a Claude Code plugin (not a Cowork plugin) that walks an IT admin through generating the customized add-in manifest, granting Azure admin consent, and writing per-user routing config via Microsoft Graph. Install with:

claude plugin install claude-for-msft-365-install@claude-for-financial-services
/claude-for-msft-365-install:setup

This is separate from the agents and vertical plugins above — it's the on-ramp that gets the add-in deployed in a tenant, after which the agents and skills here are what runs inside it.

Making It Yours Copy Chapter

These are reference templates — they get better when you tune them to how your firm works.

• Swap connectors — point .mcp.json at your data providers and internal systems.
• Add firm context — drop your terminology, processes, and formatting standards into skill files.
• Bring your templates — /ppt-template teaches Claude your branded PowerPoint layouts.
• Adjust agent scope — edit agents/<slug>.md to match how your team actually runs the workflow.
• Add your own — copy the structure for workflows we haven't covered.

Skill & Command Reference Copy Chapter

financial-analysis — core modeling, Excel, deck QC

investment-banking — deal materials and execution

equity-research — coverage and publishing

private-equity — sourcing through portfolio ops

wealth-management — advisor workflows

Contributing Copy Chapter

Everything here is markdown and YAML. Fork, edit, PR. For new content:

• New skill → add it under plugins/vertical-plugins/<vertical>/skills/, then run python3 scripts/sync-agent-skills.py to propagate to any agent that bundles it.
• New agent → plugins/agent-plugins/<slug>/ (with agents/<slug>.md + skills/) and a matching managed-agent-cookbooks/<slug>/.
• Run python3 scripts/check.py before pushing — it lints every manifest, verifies all cross-file references resolve, and fails if any bundled skill has drifted from its vertical source.

License Copy Chapter

Apache License 2.0

File: claude-for-msft-365-install/examples/python-bootstrap/README.md Copy Chapter

Bootstrap endpoint — Python reference

A minimal FastAPI implementation of the Claude in Office /bootstrap endpoint.
It validates the caller's Entra ID token and returns per-employee skills and
mcp_servers based on a simple first-match RBAC table.

Run against your real Entra tenant Copy Chapter

pip install -r requirements.txt
# Find your tenant ID:
python get_tenant_id.py [email protected]
export TENANT_ID=<your-tenant-guid>
python app.py

Run locally with a fake token Copy Chapter

pip install -r requirements.txt
export TENANT_ID=dev-tenant
TOKEN=$(python mint_dev_token.py --oid alice --group investment-banking)
DEV_JWKS_PATH=dev_jwks.json python app.py &
curl -H "Authorization: Bearer $TOKEN" \
     -H "X-Claude-User-Agent: claude-word/1.0.0" \
     http://127.0.0.1:8080/bootstrap

Customize Copy Chapter

Everything you need to change lives in config.py — app.py should not need edits.

• Edit SKILLS and MCP_SERVERS — the full catalog you can hand out.
• Edit RULES — first matching rule wins; the empty when: {} at the bottom is the default.
• Replace the placeholder group/user names in RULES with your real Entra Object IDs (GUIDs).
• Group membership is read from the token's groups claim. If your tenant
  doesn't emit it, swap the groups = ... line in app.py for a lookup against
  your internal directory.
• Rules can be scoped per Office host with "app": "word" | "excel" | "powerpoint",
  parsed from the X-Claude-User-Agent header the add-in sends.
• The groups claim is not in Entra tokens by default. Enable it under
  App registration → Token configuration → Add groups claim for your app.
• Swap the in-memory RULES for your real source of truth (DB, config service, etc.).

Security Copy Chapter

DEV_JWKS_PATH lets the server trust a self-issued signing key instead of
Microsoft's. It refuses to start unless bound to 127.0.0.1. Never set it
in a deployed environment.

File: claude-for-msft-365-install/README.md Copy Chapter

Claude for Office — Direct Cloud Setup

Admin tooling for configuring the Claude Office add-in to call your own cloud
(Vertex AI, Bedrock, or an LLM gateway) instead of Anthropic's API.

Install Copy Chapter

claude plugin marketplace add anthropics/financial-services
claude plugin install claude-for-msft-365-install@claude-for-financial-services

Then inside the session: /claude-for-msft-365-install:setup

Update Copy Chapter

Pull the latest version of the plugin:

claude plugin update claude-for-msft-365-install@claude-for-financial-services

Restart the session to apply. Re-run /claude-for-msft-365-install:setup only
if you need to regenerate the manifest with new options.

Bootstrap Copy Chapter

For per-user MCP servers, skills, or dynamic config, host a bootstrap endpoint
and point the add-in at it:

claude plugin marketplace add anthropics/financial-services   # if not already added
claude plugin install claude-for-msft-365-install@claude-for-financial-services

Then inside the session: /claude-for-msft-365-install:bootstrap

Commands Copy Chapter

Exporting user data Copy Chapter

Chat history, uploaded skills, MCP registrations, memory, and settings live in browser
storage on the user's own machine — there is no server-side copy.

The export scripts are read only: they read Office's storage and write
only to the folder you name. They never modify, move, or delete anything in
Office. With no arguments they just print what they found.

./scripts/export-addin-data.sh                       # macOS: list
./scripts/export-addin-data.sh --out ~/claude-export

.\scripts\export-addin-data.ps1                      # Windows: list
.\scripts\export-addin-data.ps1 -Out C:\claude-export

Changing the add-in does not move or delete the data. Storage is keyed by
the origin the add-in is served from, not by add-in ID, so replacing the
manifest, reinstalling, or moving between a sideloaded manifest and the store
listing all leave it in place. Export before a machine is rebuilt, or — on
Windows especially — before anyone clears the Office add-in cache, since
Wef holds the add-in's storage as well as the manifest cache.

See /claude-for-msft-365-install:export-data for the exact on-disk paths and
what to check when history looks missing.

File: managed-agent-cookbooks/pitch-agent/README.md Copy Chapter

Pitch Agent — managed-agent template

Overview Copy Chapter

Comps, precedents, LBO → branded pitch deck, end to end. Same source as the pitch-agent Cowork plugin — this directory is the Managed Agent cookbook for POST /v1/agents.

Deploy Copy Chapter

export ANTHROPIC_API_KEY=sk-ant-...
export CAPIQ_MCP_URL=... DALOOPA_MCP_URL=...
../../scripts/deploy-managed-agent.sh pitch-agent

Steering events Copy Chapter

See steering-examples.json.

Security & handoffs Copy Chapter

Task-decomposition split — less about untrusted inputs (data comes from CapIQ/Daloopa MCPs), more about parallelism and artifact isolation. Exactly one worker holds Write:

Artifacts land in ./out/pitch-<target>.pptx and ./out/model.xlsx via pptx-author / xlsx-author.

Handoff: to rebuild the model after a thesis change, the orchestrator emits a handoff_request for model-builder; scripts/orchestrate.py (or your workflow engine) routes it as a new steering event. See the script for the allowlist + payload-validation pattern.

File: managed-agent-cookbooks/month-end-closer/README.md Copy Chapter

Month-End Closer — managed-agent template

Overview Copy Chapter

Accruals, roll-forwards, variance commentary. Same source as the month-end-closer Cowork plugin — this directory is the Managed Agent cookbook for POST /v1/agents.

Deploy Copy Chapter

export ANTHROPIC_API_KEY=sk-ant-...
export GL_MCP_URL=...
../../scripts/deploy-managed-agent.sh month-end-closer

Steering events Copy Chapter

See steering-examples.json.

Security & handoffs Copy Chapter

Supporting invoices and vendor statements are untrusted. Three-tier isolation:

poster produces ./out/close-package-<entity>-.xlsx. JE drafts are staged, not posted to the GL.

Handoff: receives handoff_request events from gl-reconciler with verified breaks to fold into close commentary.

File: managed-agent-cookbooks/model-builder/README.md Copy Chapter

Model Builder — managed-agent template

Overview Copy Chapter

DCF, LBO, 3-statement, comps — built as a file artifact. Same source as the model-builder Cowork plugin — this directory is the Managed Agent cookbook for POST /v1/agents.

Deploy Copy Chapter

export ANTHROPIC_API_KEY=sk-ant-...
export CAPIQ_MCP_URL=... DALOOPA_MCP_URL=...
../../scripts/deploy-managed-agent.sh model-builder

Steering events Copy Chapter

See steering-examples.json.

Security & handoffs Copy Chapter

Task-decomposition split — inputs come from trusted MCPs, so the split is about artifact isolation and re-verification. Exactly one worker holds Write:

auditor re-checks ties and balances after builder writes ./out/model.xlsx.

Handoff: when invoked from earnings-reviewer or pitch-agent, the calling agent's handoff_request is routed here by scripts/orchestrate.py.

File: managed-agent-cookbooks/meeting-prep-agent/README.md Copy Chapter

Meeting Prep Agent — managed-agent template

Overview Copy Chapter

Briefing pack before every client meeting. Same source as the meeting-prep-agent Cowork plugin — this directory is the Managed Agent cookbook for POST /v1/agents.

Deploy Copy Chapter

export ANTHROPIC_API_KEY=sk-ant-...
export CRM_MCP_URL=... CAPIQ_MCP_URL=...
../../scripts/deploy-managed-agent.sh meeting-prep-agent

Steering events Copy Chapter

See steering-examples.json. Typically kicked from a calendar event by your workflow engine.

Security & handoffs Copy Chapter

Client-provided documents and inbound emails are untrusted. Three-tier split:

pack-writer produces ./out/briefing-<client>.pptx; it never opens client-provided content directly.

Not guaranteed: this pack is for the advisor, not the client. No client-facing send.

File: managed-agent-cookbooks/market-researcher/README.md Copy Chapter

Market Researcher — managed-agent template

Overview Copy Chapter

Sector or theme → industry overview → competitive landscape → peer comps → ideas shortlist → research note. Same source as the market-researcher Cowork plugin — this directory is the Managed Agent cookbook for POST /v1/agents.

Deploy Copy Chapter

export ANTHROPIC_API_KEY=sk-ant-...
export CAPIQ_MCP_URL=... FACTSET_MCP_URL=...
../../scripts/deploy-managed-agent.sh market-researcher

Steering events Copy Chapter

See steering-examples.json. Kick from a research-queue event or fan out across a coverage map.

Security & handoffs Copy Chapter

Third-party reports and issuer materials are untrusted. Three-tier isolation:

sector-reader returns length-capped, schema-validated JSON. note-writer produces ./out/primer-<sector>.docx (and .pptx if slides requested).

Handoff: to model a single name surfaced in the ideas shortlist, emit a handoff_request for model-builder; scripts/orchestrate.py routes it as a new steering event.

File: managed-agent-cookbooks/kyc-screener/README.md Copy Chapter

KYC Screener — managed-agent template

Overview Copy Chapter

Parses onboarding docs, runs the rules engine, screens sanctions/PEP, flags gaps. Same source as the kyc-screener Cowork plugin — this directory is the Managed Agent cookbook for POST /v1/agents.

Deploy Copy Chapter

export ANTHROPIC_API_KEY=sk-ant-...
export SCREENING_MCP_URL=...
../../scripts/deploy-managed-agent.sh kyc-screener

Steering events Copy Chapter

See steering-examples.json.

Security & handoffs Copy Chapter

Onboarding documents are untrusted. Three-tier isolation:

doc-reader returns length-capped, schema-validated JSON. escalator produces ./out/escalation-.xlsx.

Not guaranteed: this agent recommends a risk rating; the compliance officer decides.

File: managed-agent-cookbooks/gl-reconciler/README.md Copy Chapter

GL Reconciler — managed-agent template

Overview Copy Chapter

Finds breaks between general ledger and subledger for a trade date and set of asset classes, traces root cause, and produces an exception report for controller sign-off.

Same source as the gl-reconciler Cowork plugin — this directory is the Managed Agent cookbook for POST /v1/agents.

Deploy Copy Chapter

export ANTHROPIC_API_KEY=sk-ant-...
export GL_MCP_URL=...           # read-only GL MCP
export SUBLEDGER_MCP_URL=...    # read-only subledger MCP
../../scripts/deploy-managed-agent.sh gl-reconciler

Steering events Copy Chapter

See steering-examples.json. Kick a session with a trade date and asset-class list; follow-up events can re-trace a single break.

Security & handoffs Copy Chapter

This agent reads counterparty/custodian statements — documents authored by outsiders that may carry adversarial instructions. The template is structured so a payload in one of those documents cannot reach a shell, a write tool, or a firm system:

The reader returns length-capped, schema-validated JSON only (validated by scripts/validate.py). The critic independently re-verifies each break against trusted sources before the orchestrator hands the set to resolver. The resolver writes the exception report to ./out/; it never opens an outsider file.

Handoff: to feed verified breaks into Month-End Closer, the orchestrator emits a handoff_request for month-end-closer in its final output; scripts/orchestrate.py (or your Temporal/Airflow worker) routes it as a new steering event. See the script for the allowlist + payload-validation pattern.

Not guaranteed: none of this writes to a system of record. Ledger adjustments require human approval outside the agent.