Claude for Financial Services: Reference Agents, Skills, and Connectors for FSI Workflows

What This Repository Actually Solves

The repository addresses a specific pain point: financial-services teams that want to apply Claude to repetitive analyst work without building every prompt, skill, and data connection from scratch. It provides reference implementations for workflows seen in investment banking, equity research, private equity, and wealth management. The target user is an engineer or operations lead inside a bank, fund, or research shop who needs a structured starting point for agents that draft pitch decks, reconcile general ledgers, review earnings calls, or screen KYC documents. The README is explicit about scope: these agents draft analyst work product for review by a qualified professional. They do not make investment recommendations, execute transactions, post to ledgers, or approve onboarding. That boundary is the core value proposition. You get scaffolding for high-volume, low-judgement tasks, not a substitute for human expertise.

Two Deployment Paths, One Source of Truth

The defining architectural choice is that every agent exists in two forms: a Cowork plugin and a Claude Managed Agent template. Both reference the same system prompt and skills, so you pick the runtime without maintaining separate logic. In Cowork, you paste the repository URL into Settings → Plugins → Add plugin, or upload a zip of any directory under plugins/. In Claude Code, you add the marketplace with a command and then install individual plugins. For managed deployment, you export an API key and run a shell script that resolves file references, uploads skills, creates leaf-worker subagents, and POSTs the orchestrator to /v1/agents. The README points to scripts/orchestrate.py as a reference event loop that routes handoff_request events between agents. This dual-path design is practical: you can prototype in Cowork and then move to a managed API without rewriting the agent logic.

Inside the Agent Plugins: Self-Contained Workflows

Each agent plugin under plugins/agent-plugins/ is self-contained, meaning it bundles the skills it uses. The README lists ten agents, from Pitch Agent to KYC Screener. For example, the Pitch Agent handles comps, precedents, and LBO analysis to produce a branded pitch deck. The GL Reconciler finds breaks, traces root cause, and routes for sign-off. The Month-End Closer handles accruals, roll-forwards, and variance commentary. The self-contained design means you can install a single agent without pulling in the entire vertical bundle. That is a deliberate trade-off: it duplicates skills across agents, but it simplifies installation and reduces cross-agent coupling. The repository layout shows vertical plugins as the source for skills and commands, while agent plugins hold synced copies. This duplication is manageable because the sync script (scripts/sync-agent-skills.py) exists to keep copies in line.

Skills, Commands, and Connectors: The Layered Toolset

The repository separates three types of tooling. Skills are domain expertise that Claude draws on automatically when relevant, authored once in vertical plugins and bundled into agents. Commands are explicit slash actions like /comps, /dcf, /earnings, and /ic-memo that you trigger in a session. Connectors are MCP servers that wire Claude to external data: terminals, research platforms, and document stores. The financial-analysis vertical plugin includes a .mcp.json file that defines these connectors. This layering matters because you can install just the vertical bundle if you only want the commands and connectors without a full agent. The README gives an example: install financial-analysis@claude-for-financial-services first for core skills and connectors, then add named agents. That modularity is useful for teams that want to standardize on data access before introducing workflow automation.

Getting Started: Commands and Configuration Keys

The README provides concrete installation steps. For Cowork, you open Settings → Plugins → Add plugin and either paste the repository URL or upload a zip. For Claude Code, the commands are: claude plugin marketplace add anthropics/financial-services, then claude plugin install financial-analysis@claude-for-financial-services, followed by specific agents like pitch-agent@claude-for-financial-services or gl-reconciler@claude-for-financial-services. For Managed Agents, you set the ANTHROPIC_API_KEY environment variable and run scripts/deploy-managed-agent.sh gl-reconciler. The deploy script handles file reference resolution, skill upload, subagent creation, and the POST to /v1/agents. There is also a claude-for-msft-365-install directory for provisioning the Claude Microsoft 365 add-in, which suggests a path for teams using Office tools. The configuration surface is minimal: you pick a runtime, install a plugin, and optionally deploy via the script. No complex YAML editing is required unless you go into managed-agent-cookbooks for customizing agent.yaml.

Limitations and Failure Modes You Should Not Ignore

The most significant limitation is the explicit disclaimer: nothing in the repository constitutes investment, legal, tax, or accounting advice. Every output is staged for human sign-off. That is not a legal nicety; it means the agents are not designed to handle the final judgement calls. A second limitation is that subagent delegation, the callable_agents capability, is a research preview. That implies it may change or be unstable in production. The README directs you to per-agent security notes for handoff guidance, but the preview status is a real risk. A third limitation is the dependency on the Managed Agents API, which is a specific Anthropic service. If your firm cannot use that API or prefers a different orchestration layer, the deploy script and orchestrate.py may not fit. The repository is also a reference set, not a turnkey product. The README says agents are starting points and you need to tune prompts, skills, and connectors to your firm's methods. That is honest but means adoption requires engineering effort.

Alternative Approaches and Where This Fits

The main alternative is building your own agent pipelines from scratch using the Claude API directly, without these reference plugins. That approach gives you full control over prompts, tool selection, and orchestration, but it requires you to author all the domain skills and connectors yourself. Another alternative is using a general-purpose agent framework like LangChain or LlamaIndex, which are model-agnostic but do not come with financial-services-specific skills. This repository sits between those extremes: it provides Anthropic-specific, FSI-tuned skills and a deployment script, but it locks you into the Claude ecosystem and the Managed Agents API. If you need to support multiple model providers or have a custom orchestration engine that is not compatible with the provided scripts, a framework approach may be more flexible. However, if you are already standardized on Claude, the reference agents save significant prompt engineering time.

Maintenance, Upgrade Cost, and License Implications

The repository is licensed under Apache-2.0, which permits commercial use, modification, and distribution, with the caveat that you must preserve license notices. That is permissive, but it does not include any warranty or liability from Anthropic. The maintenance cost is moderate. The repository includes scripts for validation (check.py, validate.py) and synchronization (sync-agent-skills.py), which help keep bundled skills in line with vertical sources. However, the repository publishes no GitHub releases, so there are no version numbers to track; its last push was 2026-10-28. The README mentions partner-built plugins from LSEG and S&P Global, which could introduce external dependencies and update cycles. When you adopt this, you inherit the need to track changes in the Managed Agents API, the MCP protocol, and any connector endpoints. The deploy script is a reference implementation, so you should test it in a staging environment before production use. The presence of a claude-for-msft-365-install directory also implies you may need to manage add-in provisioning separately.

Editorial conclusion

Adopt this repository if you are a financial-services firm already using Claude Cowork or Claude Code and need a starting point for structured workflows like pitch decks, earnings reviews, GL reconciliation, or KYC screening. The agents are self-contained and deployable via the Managed Agents API, which suits teams with an existing orchestration layer. Do not use it if you expect production-ready, compliance-approved automation: the README explicitly states these agents draft work product for human sign-off and offer no investment, legal, tax, or accounting advice. Before adoption, verify that the bundled skills and connectors match your data sources, especially the MCP servers in financial-analysis, and review the per-agent security notes in managed-agent-cookbooks, since subagent delegation is a preview capability. Also confirm that the Apache-2.0 license aligns with your firm's internal policy, and test the deploy script in a sandbox before touching production.

Frequently asked questions

Official sources

1. Official README
2. Project repository
3. README
4. Releases