Financial-services work, packaged as a system.

Anthropic’s repository is a reference architecture for turning repeatable analyst workflows into named agents, reusable skills, connected data, and reviewable outputs.

10 named agents6 vertical pluginshuman sign-off

Source: github.com/anthropics/financial-services

The one-line read

It is not a robo-advisor. It drafts analyst work product and stages it for a qualified professional to verify, mark up, approve, or reject.

From “ask an LLM” to “run a controlled workflow”.

The value is orchestration and repeatability, not magic finance answers.

01 / repeatability

Do the same work the same way

Capture the firm’s process in skills, commands, templates, and output rules instead of relying on one-off prompting.

02 / leverage

Bring the right data into the work

Connect research, market-data, document, CRM, and internal systems through MCP servers that the workflow can call.

03 / control

Stop at the review boundary

Agents produce drafts, models, exception reports, or briefing packs. A human remains accountable for the final decision.

Use case examples

Pitch Agent: comps + precedents + DCF → branded pitch deck.
Market Researcher: sector question → overview, peer comps, shortlist.
GL Reconciler: ledger break → root-cause trace → controller sign-off.

Remember this

Think of it as a set of job-shaped starting points. Install the one that matches your work, then adapt prompts, skills, connectors, templates, and approval steps to your firm.

Source: README / Agents · Making It Yours

Four layers, one workflow.

The folder structure is the product story: domain know-how is separated from workflow packaging and deployment.

Vertical pluginsSkills, slash commands, connectors, templates.

Partner pluginsLSEG and S&P Global data workflows.

Named agentsystem prompt + selected skills + tool access

A self-contained workflow for Cowork or the Managed Agents wrapper.

Managed cookbookagent.yaml, leaf workers, steering examples, security notes.

Output boundaryModels, memos, notes, reconciliations, and packs staged for review.

source of truth

Vertical skills are authored once, then synced into agent bundles.

deployment choice

Run interactively in Cowork or headlessly via the Managed Agents API.

operating principle

Separate reading and analysis from writing the final artefact.

Source: README / Repository Layout · Managed cookbooks

Named agents map to real finance jobs.

The repo starts with 10 end-to-end workflows. They are starting points, not immutable products.

Coverage & advisory

Pitch AgentMeeting Prep Agent

Research & modelling

Market ResearcherEarnings ReviewerModel Builder

Fund admin & ops

Valuation ReviewerGL ReconcilerMonth-End CloserStatement Auditor

Onboarding

KYC Screener

The design choice

Each agent owns a workflow end to end: what it reads, which tools it can call, what artefacts it produces, and where it stops.

The product implication

Instead of one giant “finance assistant”, teams can choose narrow, inspectable workers with different permissions and approval paths.

Source: README / Agents table · Agent plugin source

It produces work product, not a final decision.

The key safety pattern is the hand-off between machine-generated artefacts and qualified human judgement.

01 / inputQuestion, file, period

“Review this earnings event” or “reconcile this trade date”.

02 / retrieveRead approved sources

Filings, transcripts, market data, ledgers, CRM, or internal documents.

03 / reasonApply domain skill

Run the method, rules, model, or checklist captured in the plugin.

04 / draftStage artefacts

Model, note, exception report, pack, or escalation packet.

05 / reviewHuman sign-off

Qualified professional verifies, marks up, approves, or rejects.

Hard boundary from the repo

These agents do not make investment recommendations, execute transactions, bind risk, post to a ledger, or approve onboarding. Every output is staged for human sign-off.

Source: README / Important

Look inside the packaging.

The interesting part is not the list of agents. It is how prompts, skills, tools, deployment, and review are composed into one repeatable operating model.

Same system prompt. Different place to run it.

The workflow definition stays portable across interactive and headless contexts.

interactive

Cowork plugin

Add the repo as a plugin marketplace source.
Pick named agents or vertical bundles.
Agents appear in dispatch; skills fire when relevant.

headless

Claude Managed Agents

Deploy an agent.yaml through /v1/agents.
Use leaf workers for parallel or isolated tasks.
Route steering events through your own workflow engine.

Also in the repo

Claude Code marketplace installation is documented, and the Microsoft 365 directory contains admin tooling for tenants routing through Vertex AI, Bedrock, or an internal LLM gateway.

Source: README / Getting Started · Microsoft 365 tooling

Vertical plugins are the reusable knowledge layer.

Install the core first, then add the workflows your team actually needs.

core

financial-analysis

Comps, DCF, LBO, three-statement, deck QC, Excel audit, and shared data connectors.

investment-banking

CIMs, teasers, process letters, buyer lists, merger models, and deal tracking.

equity-research

Earnings notes, initiations, model updates, thesis and catalyst tracking.

private-equity

Sourcing, screening, diligence checklists, IC memos, and portfolio monitoring.

fund-admin

GL reconciliation, break tracing, accruals, roll-forwards, and NAV tie-out.

operations

KYC document parsing and rules-grid evaluation.

The useful abstraction

Agents are workflow products. Vertical plugins are the knowledge and tool substrate they reuse.

Source: README / Vertical Plugins · Vertical plugin tree

Skills teach the method. Commands make it explicit.

The repo supports both automatic relevance and deliberate invocation.

automatic

Skills

Domain expertise, conventions, and step-by-step methods that Claude can draw on when relevant.

comps-analysisdcf-modelaudit-xlspptx-authorclient-review

explicit

Slash commands

Named actions that make an analyst’s intent visible and repeatable in a session.

/comps/dcf/lbo/earnings/ic-memo

Why the split matters

Skills preserve the firm’s way of working.

Commands preserve a visible trigger.

Agents package both into a job-shaped flow.

Source: README / Skill & Command Reference

Connectors are where the workflow meets firm reality.

The core plugin centralises external data connections so multiple agents can reuse them.

Daloopa

Morningstar

S&P Global

FactSet

Moody’s

MT Newswires

Aiera

LSEG

PitchBook

Chronograph

Egnyte

Box

Design opportunity

One connector layer can support multiple agents, while each agent still receives only the tools it needs.

Repo hygiene finding

The README says “all 11 data connectors”, while the current core MCP file lists 12 names and fails JSON parsing at the Egnyte / Box boundary. Treat manifests as code: validate them in CI before deployment.

Sources: README / MCP Integrations · financial-analysis/.mcp.json

Orchestrate broadly. Write narrowly.

The cookbooks make permissions and hand-offs explicit when a workflow fans out to leaf workers.

OrchestratorReads, greps, globs, routes, aggregates. No final write.

Read-only MCPData providers or internal systems scoped to the job.

Agent YAMLmodel + system prompt + tools + skills + callable agents

The manifest is the deployment contract for /v1/agents.

Leaf workersThree depth-1 subagents handle focused tasks.

One write holderAssemble the artefact, then surface it for human review.

Example: Pitch Agent

Researcher pulls verified comps and precedents → Modeler prepares valuation logic → Deck writer creates the branded draft. The single write-holder pattern reduces accidental cross-contamination between data gathering and artefact creation.

Sources: Pitch Agent cookbook · Model Builder cookbook

The system is designed to stop before authority.

That boundary is the repo’s clearest product decision.

What the agents can do

Read approved inputs and connected sources.
Apply a repeatable domain method.
Draft models, memos, notes, reports, reconciliations, and packs.
Flag gaps, exceptions, and evidence for review.

What they do not do

Make investment recommendations.
Execute transactions or bind risk.
Post to a ledger.
Approve onboarding or replace qualified sign-off.

What your implementation still needs

Identity, tenancy, and least-privilege access.

Evidence retention, provenance, and audit trails.

Evaluation datasets for accuracy and escalation behaviour.

Policy mapping, monitoring, incident response, and rollback.

Sources: README / Important · implementation needs are an architectural inference from the repository scope.

Start with one workflow and earn the next one.

Pick a painful, repeatable job, keep the approval boundary visible, and add complexity only when the first path is reliable.

A practical implementation path.

Select a step to see the purpose and done condition.

Source: README / Getting Started · Making It Yours

Choose the runtime that matches the job.

Start interactively and move to headless deployment once the workflow earns it.

Cowork

Fastest for exploration

# Settings → Plugins
Paste the repo URL
Pick an agent or vertical

Best for analyst-led discovery and prompt tuning.

Claude Code

Fastest for local iteration

claude plugin marketplace add
anthropics/financial-services

claude plugin install financial-analysis@claude-for-financial-services

Best for editing skills, adding firm context, and validation.

Managed Agents

Best for orchestration

export ANTHROPIC_API_KEY=…

scripts/deploy-managed-agent.sh
gl-reconciler

Best for steering events, leaf workers, and headless outputs.

Source: README / Getting Started

Copy the shape, then replace the firm-specific truth.

Use the manifests as a contract, not as a black box.

name: earnings-reviewer
model: claude-opus-4-7

system:
  file: agents/earnings-reviewer.md
  append: "Produce files in ./out/"

tools:
  - read
  - grep
  - glob
  - mcp: factset
  - mcp: daloopa

callable_agents:
  - transcript-reader
  - model-updater
  - note-writer

Make the contract explicit

Input: ticker, reporting period, and approved source IDs.
Tools: only the connectors needed for that workflow.
Outputs: named artefacts with a known review state.
Stop conditions: missing evidence or approval required.
Write boundary: one worker assembles the final file.

Done condition

Run one representative case end to end, inspect every citation and calculation, and have a qualified reviewer approve the output shape before widening scope.

Source: earnings-reviewer/agent.yaml · agent prompt

Earnings Reviewer, end to end.

The repo’s workflow logic translated into a simple sequence.

INPUTTicker + period

One covered name and one reporting event.

READTranscript + filings

Pull evidence through FactSet and Daloopa.

UPDATECoverage model

Drop actuals, roll estimates, and flag variance.

DRAFTEarnings note

Headline read, drivers, estimate changes, valuation update.

REVIEWSenior analyst

Mark up and approve before downstream use.

What this teaches

Good agent design is about artefact choreography: each step produces something the next step can inspect, validate, or improve.

What to measure

Evidence coverage, formula integrity, variance accuracy, missing-data escalation, reviewer edits, and time saved per event.

Source: Earnings Reviewer prompt · cookbook

Ship the workflow boundary, not just the prompt.

A useful reference repo becomes a reliable internal product only when the surrounding controls are real.

Data & permissions

Scope connectors per agent.
Validate identities and tenant boundaries.
Keep sensitive data out of prompts and logs where possible.

Quality & evidence

Build golden cases and regression tests.
Require source references for material claims.
Check models, formulas, and tie-outs.

Operations & review

Make approval states visible.
Monitor failure and escalation rates.
Keep a rollback path for skills and connectors.

The smallest credible pilot

One workflow. One data path. One output artefact. One qualified reviewer. Ten representative cases. A clear go / no-go rule.

Source basis: repository structure, agent prompts, Managed Agent cookbooks, and the README’s human-sign-off boundary. Operational controls are implementation recommendations.

The real product is the workflow boundary.

The agents are examples. The architecture is the reusable part.

Package work by job

Named agents make scope, tools, outputs, and review points visible.

Separate knowledge from orchestration

Vertical skills and connectors become reusable infrastructure for many workflows.

Keep authority with people

Drafting and analysis can be automated while approval, risk, and accountability stay explicit.

Best next move

Pick the most repetitive, evidence-rich workflow in your team. Turn it into one agent with a narrow data surface, a measurable output, and an unmistakable human sign-off step.

Primary source: github.com/anthropics/financial-services · snapshot at commit 574ed36