Claude for Financial Services (repo) — source summary

Anthropic’s reference repository (github.com/anthropics/financial-services) of agents, skills, and data connectors for financial-services workflows. Delivered via Telegram and ingested 2026-05-29. README in raw/claude-financial-services.md.

Topic note: this opens a new cluster largely separate from the wiki’s knowledge-management thesis. Its honest ties to the existing graph: it runs on model-context-protocol (as does qmd), and it shares the “file-based markdown + YAML that an LLM reads and acts on” shape with the llm-wiki pattern.

What it is

Reference templates for IB, equity research, private equity, wealth management, and fund admin. The pitch: “two ways from one source” — install as a claude-cowork plugin, or deploy through claude-managed-agents behind your own workflow engine, sharing the same system prompt and skills. Everything is file-based markdown + YAML, no build step.

Structure

• Agents — 11 self-contained, named end-to-end workflow agents (Pitch Agent, Market Researcher, Earnings Reviewer, Model Builder, GL Reconciler, Month-End Closer, KYC Screener, …). Each bundles the skills it uses.
• Vertical plugins — skill + command bundles by FSI vertical, with financial-analysis as the core (modeling skills + all data connectors). Slash commands like /comps, /dcf, /lbo, /earnings, /ic-memo.
• Partner plugins — LSEG and S&P Global.
• MCP connectors — 10+ providers (Daloopa, Morningstar, FactSet, Moody’s, LSEG, PitchBook, …), centralized in the core plugin. See model-context-protocol.
• Managed-agent cookbooks — agent.yaml + depth-1 subagents + steering examples for headless deployment via /v1/agents. See claude-managed-agents.

Notable design points

• agent-skills: skills are authored once in the verticals and a synced copy is bundled into each agent (sync-agent-skills.py); check.py lints manifests and verifies all cross-file references resolve and that no bundled skill has drifted from source — an explicit maintenance/consistency discipline (cf. the llm-wiki lint pass).
• Strong human-in-the-loop framing: outputs are “staged for human sign-off”; the agents draft work product, they don’t execute transactions or make recommendations.
• Subagent delegation (callable_agents) is flagged as a Research Preview.

Re-verified 2026-08-04 (staleness queue)

The repository URL was named in this page’s prose but never in frontmatter, so nothing could check it. url: added and the repo re-read: 34.0k stars, 5.0k forks, Apache-2.0 — which makes it one of the largest artifacts in this spoke by adoption, not the reference sample the original summary implied.

Current contents: 11 named workflow agents (Pitch, Meeting Prep, Market Researcher, Earnings Reviewer, Model Builder, GL Reconciler, Valuation Reviewer, Month-End Closer, Statement Auditor, KYC Screener), 8 vertical plugin bundles including two partner-built ones (LSEG, S&P Global), and 11 MCP data connectors (Daloopa, Morningstar, S&P Global, FactSet, Moody’s, MT Newswires, Aiera, LSEG, PitchBook, Chronograph, Egnyte, Box).

It confirms the two-surface split this spoke recorded: the same skills install as claude-cowork plugins or deploy as claude-managed-agents via /v1/agents for headless use. Installation is through the claude-code plugin marketplace.