Claude Financial Services: what the public plugins actually provide
Claude Financial Services: what the public plugins actually provide
Separate financial workflow files, paid host services and human approval before choosing an agent.
What you will learn
- Start with an analyst work product
- Choose the distribution layer
- Decide what remains outside the trial
Before you start
- Command-line and JSON familiarity
- A disposable workspace with synthetic financial records
A proposed learning project turns validated requests into reviewable records without giving documents control of routing.
Key takeaways
- An agent bundles instructions and skills.
- Published workflow files do not include provider access.
- Human sign-off remains a separate action.
Start with an analyst work product
The financial-services repository contains reference agents, skills and data connectors for Claude. A named agent packages instructions for a task such as reconciling ledger discrepancies or preparing a meeting brief. The deliverable is a draft that a person reviews.
A useful first case is a synthetic reconciliation: two small tables disagree on one invoice. Ask for the discrepancy, its supporting rows and the missing evidence. Do not grant posting permissions or use a real client account to learn the workflow.
Choose the distribution layer
The README describes Cowork plugins and Managed Agent templates that reuse the same underlying prompts and skills. Vertical plugins provide domain methods and commands, while named agents bundle the skills needed for a workflow. Installing both indiscriminately can make ownership of local edits unclear.
These public files are separate from the Claude host and external data subscriptions. An Apache-2.0 declaration for this repository does not mean the model, connector service or provider dataset is included without cost.
Decide what remains outside the trial
The README explicitly reserves investment decisions, transactions, ledger posting and onboarding approval for human control. Treat an attractive spreadsheet or fluent memo as an item requiring evidence checks, not permission to act.
This series inspects commit 574ed36 and two harness helpers. It includes a small offline parser test, but no live account, provider connection or financial-agent deployment was executed. Later chapters distinguish those evidence levels.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Pick one synthetic discrepancy.
- 2
Select its reviewer before installing plugins.
- 3
List the host and data accounts separately.
Copy-ready example
{"trial":"synthetic-invoice-reconciliation","posting_allowed":false,"reviewer":"assigned-human"}Frequently asked questions
Is this a trading bot?
The documented scope is drafting work products for review, not executing trades.
Was a live reconciliation tested?
No. Only the specifically identified offline parser fixture was executed.
Sources
- Financial Services / README.mdSource checked 2026-09-23