Anthropic financial services agents: structure and monitoring

Anthropic financial services agents: explore the Apache 2.0 repository, its installation and document-reader isolation, with no financial advice given.
Anthropic financial services agents : structure et supervision

Anthropic financial services agents refers to the anthropics/financial-services repository: reference agents, skills and data connectors for finance jobs (investment banking, equity research, private equity, wealth management). The repository shows about 40,700 stars in mid-October 2026 under the Apache 2.0 licence, and it opens with a blunt notice: nothing in it constitutes investment, legal, tax or accounting advice.

This article gives no financial advice and does not present these agents as ready for regulated use: the README describes them as reference templates you adapt, every output being staged for human sign-off. Our angle is operations: how the repository is organised, how to install it and what to monitor.

span.observer did not install or run these agents, is not affiliated with Anthropic and does not speak for it. The screenshots are excerpts rendered from the repository files.

Anthropic financial services agents: what the repository says

  • Publisher and licence: Anthropic; Apache License 2.0 (LICENSE file).
  • Figures read on 22 October 2026 on the repository page: about 40,700 stars, about 5,700 forks, 73 commits on the main branch. No release is shown.
  • Two deployment modes from one source: a plugin for Claude Cowork, or deployment through the Claude Managed Agents API behind your own workflow engine. Same system prompt, same skills.
  • Named agents: ten, including Pitch Agent, Meeting Prep Agent, Market Researcher, Earnings Reviewer, Model Builder, Valuation Reviewer, GL Reconciler, Month-End Closer, Statement Auditor and KYC Screener.
  • Vertical plugins: financial-analysis (core, with the connectors), investment-banking, equity-research, private-equity, fund-admin, operations, claude-for-financial-advisors, plus two partner plugins (LSEG, S&P Global).

The repository is not a Dynatrace product. The monitoring part applies to Dynatrace SaaS and Dynatrace Managed alike, through OpenTelemetry.

A tour of the repository: agents, verticals, connectors

The README splits the repository into plugins/agent-plugins (one self-contained plugin per agent), plugins/vertical-plugins (skills and commands per vertical, plus the MCP connectors), plugins/partner-built and managed-agent-cookbooks (one folder per agent for API deployment). The Claude Code install commands are there too.

Excerpt from the financial-services README: repository layout (plugins, managed-agent-cookbooks, scripts) and claude plugin marketplace add and claude plugin install commands for financial-analysis and several agents
Source : github.com

An agent is a markdown file with a header (name, description, allowed tools) and instructions. The GL Reconciler gives itself three deliverables (break list, root-cause trace, exception report) and three guardrails: counterparty statements are untrusted, the orchestrator never writes, and no ledger posting is made by the agent.

Excerpt from the agents/gl-reconciler.md file of financial-services: header with name, description and tools, deliverables, five-step workflow and guardrails
Source : github.com

Step-by-step installation

In Claude Code

According to the README, you add the marketplace, install the core plugin first, then the agents you want.

claude plugin marketplace add anthropics/financial-services
claude plugin install financial-analysis@claude-for-financial-services
claude plugin install gl-reconciler@claude-for-financial-services

The slash commands (/comps, /dcf, /earnings, /ic-memo) become available after installation.

In Cowork

The README says: Settings, Plugins, Add plugin, then paste the repository URL or upload a zip of a folder under plugins/.

Through the Managed Agents API

The README gives a deployment script that reads a key from ANTHROPIC_API_KEY:

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

Here sk-ant-... is only an example: never put a real key in a shell history or a repository. The README states that subagent delegation (callable_agents) is a research-preview capability.

Security: isolating the document readers

The GL Reconciler cookbook README explains the most useful point for an SRE: counterparty statements are written by third parties and may carry hostile instructions. The template therefore separates three tiers: a reader (the only one opening those documents, with Read and Grep and no connector), an orchestrator (no access to raw documents, with read-only MCPs) and a resolver (the only holder of write access, which never sees an outside file).

Excerpt from the gl-reconciler cookbook README in financial-services: table of the reader, orchestrator and resolver tiers with their tools and connectors
Source : github.com

This pattern matches our articles on guardrails for AI agents in production and securing an MCP server. The README also states what is not guaranteed: nothing writes to a system of record, and ledger adjustments need human approval outside the agent.

Limits and open points

  • Not turnkey for regulated use: the README speaks of starting templates to adapt; compliance with your rules remains your responsibility, as it says itself.
  • Third-party connectors: the README lists financial data services (Daloopa, Morningstar, FactSet, PitchBook, LSEG and others) and notes access may need a subscription or an API key from the provider.
  • No release: no version on the repository page; each plugin carries its own (0.1.0 for gl-reconciler when read). Pin a commit.
  • Preview: delegation between subagents is flagged as a preview.
  • Data: the agents read internal financial data. Decide before the first trial what may leave your perimeter.

Observing and monitoring these agents

The repository describes no OpenTelemetry export; monitoring sits on the client and connector side.

  • Plugin agents (Claude Code): the claude_code.tool_result event gives tool_name, success and duration_ms, with tokens and costs per mcp_server.name. See supervising a coding agent in the enterprise.
  • Agents deployed by API: the README ships scripts/orchestrate.py, a reference event loop that routes handoff requests between agents. It is a good place to add your own spans and decision logs.
  • Costs: several subagents per run multiply calls; see the token cost of an AI agent.
  • Business traceability: for each staged report, keep who approved it and when; that is the proof human sign-off happened.

How to check that it works

  1. The core plugin and the agent appear in your client; the README slash commands respond.
  2. On a fictitious dataset, the agent produces a staged report without writing to any real system.
  3. Your traces show each tool call with its status and duration.

Best practices

  • Work first on fictitious data and read-only connectors.
  • Keep the reader, orchestrator, resolver split: never give write access to what reads outside documents.
  • Review every output before use: they are drafts, not decisions.
  • Ask your compliance team what is allowed before any real use.

Frequently asked questions

Do these agents give financial advice?

No: the README says they prepare work product for review by a qualified professional, with no investment recommendations or transaction execution. This article gives no financial advice either.

Cowork or the API?

Both are possible from the same source. Cowork suits an individual trial; the Managed Agents API suits deployment behind your workflow engine.

Same plugin mechanics, see the knowledge work plugins.

Conclusion

The financial-services repository is mostly valuable as an example of well-partitioned agents: isolated write rights, outside documents treated as hostile, explicit human sign-off. Read it for its structure, adapt it with your compliance teams, and monitor it like any agent that touches sensitive data. Source: anthropics/financial-services repository.

span.observer is an independent community blog, not affiliated with Dynatrace LLC, with the projects cited or with Anthropic. Trademarks belong to their owners. Tools change quickly: check commands and options in their official documentation.

Total
0
Shares
Leave a Reply

Your email address will not be published. Required fields are marked *

Previous Post
Knowledge work plugins Claude Cowork : installer et contrôler

Knowledge work plugins Claude Cowork: install and audit

Next Post
AI engineering from scratch parcours IA : le cursus du dépôt

AI engineering from scratch learning path: the repo curriculum

Related Posts