What Is Claude for Financial Services? Anthropic's 20 Finance Plugins and How to Install Them (2026)

What Is Claude for Financial Services?

Back in May 2026, Anthropic announced ten AI agents for finance work. That was the news; this is the delivery. The code now sits on GitHub — and it does not arrive as shrink-wrapped software you install and forget. It arrives as a reference implementation: task instructions, workflow steps and data wiring written out in plain text for firms to adapt to how they actually work.

What separates it from a general-purpose assistant is the organising principle. Nothing here is grouped by feature; everything is grouped by the job a finance team actually does. The agents are named after the work, not the technology — Month-End Closer, GL Reconciler, Statement Auditor — and the same agent runs in a chat window, in a terminal, or headless behind a workflow engine with nobody watching.

Three surfaces, one agent

The deployment-template documentation puts it plainly: "Same agent, same skills — pick your surface." The agent and its skills do not change; only the place you call them from does. For the headless side, see our deeper piece on Claude Managed Agents.

What Is Inside the Repo?

anthropics/financial-services/
├── plugins/
│   ├── agent-plugins/        # 10 ready-made agents (skills bundled in)
│   ├── vertical-plugins/     # 7 skill bundles by function + data connectors
│   └── partner-built/        # partner plugins (LSEG, S&P Global)
├── managed-agent-cookbooks/  # 10 headless deployment templates
├── claude-for-msft-365-install/  # Microsoft 365 add-in provisioning
└── scripts/                  # deploy-managed-agent.sh, validate.py, orchestrate.py

The marketplace listing comes to exactly 20 entries — 10 agents, 7 function bundles, 2 partner plugins and 1 Microsoft 365 provisioning tool.

All of it is text. A skill is a folder with a Markdown file describing the work step by step; an agent inside a plugin is a Markdown file with a YAML front matter block, while an agent in the deployment templates is plain YAML. Either way it names the model, the tools it may use and the skills it pulls in. There is no compiled code hiding underneath — you can read every instruction before deciding what data it is allowed near.

The 10 Agents — What They Do

These are the closest thing to ready-to-run in the repo, because each one bundles the skills it needs. Install one and start working.

Function Bundles and Slash Commands

The other half of the repo is the vertical plugins. These are not agents; they are toolboxes that extend your session. Five of them add slash commands — type the command and the workflow runs. fund-admin and operations ship skills only, with no commands attached; those fire on their own when Claude judges the task to match.

Install financial-analysis first, always — it carries the 13 shared skills and every data connector. Worth knowing: not all of those skills are finance-specific. xlsx-author and pptx-author write real Excel and PowerPoint files (the documentation scopes them to headless Managed Agent mode), audit-xls traces formulas through a workbook somebody else built, clean-data-xls tidies raw extracts, and skill-creator exists so you can write your own skills in the same shape.

The 12 Data Connectors

The .mcp.json file inside financial-analysis declares twelve MCP servers, all HTTP, all pointing straight at the data provider:

How to Install It

1) On Claude Code (CLI)

# Add the marketplace first
claude plugin marketplace add anthropics/financial-services

# Core skills and connectors — install this one first, always
claude plugin install financial-analysis@claude-for-financial-services

# Then pick the agents you want
claude plugin install month-end-closer@claude-for-financial-services
claude plugin install gl-reconciler@claude-for-financial-services

# Or install a whole function bundle
claude plugin install fund-admin@claude-for-financial-services
claude plugin install equity-research@claude-for-financial-services

The marketplace name after the @ is claude-for-financial-services — spelled exactly like that, and not the same as the repo name. Once installed, the slash commands appear in a new session — if you had a session open in another window while installing, restart it before they show up. Skills fire on their own whenever Claude judges the task to match.

2) On Claude Cowork

No terminal required. Open Settings → Plugins → Add plugin and choose one of two routes:

3) Headless, via the Managed Agents API

For teams that want an agent running behind their own systems — say, at 2am on the last day of every month — the repo ships deployment templates for all ten agents under managed-agent-cookbooks/. This route differs from the other two in one important way: you have to clone the repo first, because marketplace add leaves no checkout for you to run the script from.

# requires jq and python3 with pyyaml, or the script stops immediately
git clone https://github.com/anthropics/financial-services
cd financial-services

export ANTHROPIC_API_KEY=sk-ant-...
export GL_MCP_URL=...          # your ledger endpoint (read-only)
export SUBLEDGER_MCP_URL=...   # your subledger endpoint

# rehearse first — no API key needed, nothing is sent to the API
scripts/deploy-managed-agent.sh gl-reconciler --dry-run

# the real thing
scripts/deploy-managed-agent.sh gl-reconciler

The script resolves the file references in the template, uploads the skills as custom skills, creates the leaf-worker subagents first and POSTs the orchestrator to /v1/agents last. The agent.yaml itself is more readable than you would expect — this is part of gl-reconciler:

name: gl-reconciler
model: claude-opus-4-7

tools:
  - type: agent_toolset_20260401
    default_config:
      enabled: false        # everything off by default
    configs:
      - name: read
        enabled: true
      - name: grep
        enabled: true

mcp_servers:
  - type: url
    name: internal-gl
    url: ${GL_MCP_URL}      # set via environment or a vault

Two details are worth copying. The model: line pins a specific version — change it to whichever model your firm has approved, Opus 5 included. And enabled: false as the default, with tools switched on one at a time, is the posture every agent in your organisation should start from, not just this one.

The Security Architecture — the Part Worth Stealing

The most valuable thing in this repo may not be the finance skills at all. It is the pattern for structuring an agent so that a document from outside cannot do damage. Look at how gl-reconciler splits one job across four roles, each holding a different set of permissions:

One more deliberate constraint: named agents never call each other directly. They emit a handoff_request, and your own orchestration layer validates it against an allowlist before routing it on. When gl-reconciler finishes confirming breaks, that is how the work reaches month-end-closer — through a checkpoint, not through a private back channel.

Limits You Should Know First

The repo is unusually candid about its own boundaries, and they are worth reading before expectations get set:

How Far Can This Reach Into Your ERP?

The question finance teams ask within a minute is whether it can read the data in their ERP. The honest answer is yes, if somebody builds the connection. None of these agents is bound to a particular database; they speak MCP and nothing else. That is why gl-reconciler only ever says ${GL_MCP_URL} — the far end can be any ERP willing to expose a read-only endpoint.

On the Saeree ERP side we hold two lines:

We ran this on ourselves first — connecting MCP to our own procurement, inventory, HR and document workflows, hitting the problems and fixing them — and only opened MCP integration work to clients in September 2026. So there are no client results claimed in this article.

For a team starting out, the sensible order is to begin read-only and offline: export a trial balance, let the agent hunt for unmatched entries, and compare what it finds against what your people find by hand. When the two agree, open a read-only endpoint. Write-back is a much later conversation — and in plenty of organisations the final answer is that it is never needed at all.

Who It Suits — and Who It Does Not

Conclusion

Claude for Financial Services is not a finished product. It is a set of blueprints published in full — how a working finance agent should be structured, how much permission it should hold, and where the human belongs in the process. For firms outside the US, the value may not be the DCF or LBO skills at all; it is the security pattern and the discipline of splitting one job into several narrow roles. That travels to your own internal workflows immediately, without waiting for anyone to localise it.

If you want a real feel for it, five minutes is enough: add the marketplace, install financial-analysis, and run /comps against your own sample data once. That will tell you more than ten pages of description.

"The part worth copying from this repo is not the valuation maths. It is splitting one job into four roles and giving only one of them the right to write."

- The Saeree ERP team

References

• anthropics/financial-services — GitHub repository (Apache 2.0)
• README — Getting started with Cowork, Claude Code and Managed Agents
• managed-agent-cookbooks — headless deployment templates
• GL Reconciler — reader / orchestrator / resolver architecture