Data, Evidence, and Handoffs: Where Context Crosses a Boundary
In anthropics/financial-services, a connector is more than a URL and a handoff is more than a message. Both are boundaries across which context gains a new audience and may acquire new consequences. This chapter examines what crosses those boundaries: provider data, internal records, structured worker results, and cross-agent events. The practical goal is to distinguish context that can be carried forward from authority that must still be re-established by the receiving workflow or a human reviewer.
A connector is a data-plane contract
The repository centralizes MCP connector definitions in the financial-analysis core plugin so that different workflows can use relevant data providers and internal systems. The provider list includes market-data and research sources as well as document and portfolio systems. Individual Managed Agent manifests then select a narrower set for the job: Model Builder uses CapIQ and Daloopa, Earnings Reviewer uses FactSet and Daloopa, GL Reconciler uses internal GL and subledger endpoints, and KYC Screener uses a screening endpoint.
The important detail is that the repository describes endpoints and configuration patterns, not a complete data service. A connector's value depends on several facts that sit outside the URL. Which provider is it? What subscription or API key gives the firm access? What time period and definitions does the provider use? How fresh is the data? Has the firm's adapter preserved units, identifiers, and read-only behavior? Is the data licensed for the intended use? The repository names these dependencies indirectly through its subscription caveat and its environment-configured MCP URLs, but it does not supply the answers for a particular firm.
That makes the connector part of the evidence contract. A comparable-company skill can prefer structured institutional sources over web search, but the preference only improves analysis if the configured source is entitled, available, and correctly mapped. Model Builder can require historicals and consensus data, but the workflow still needs to know whether a missing field is unavailable, not licensed, or simply not returned by the adapter. GL Reconciler can be given read-only GL and subledger servers, but the organization remains responsible for ensuring that those servers represent the intended books and date scope.
The same connector can therefore have different meaning in different workflows. A research agent may use a provider for market multiples and filings. A reconciliation agent may use an internal endpoint as a source for verification. A screening agent may use a service that returns matches requiring human interpretation. The URL is shared infrastructure; the source hierarchy and review method remain domain-specific.
Freshness and definition are especially important because financial evidence is time-bound. A market multiple as of one date is not interchangeable with a later close, and a ledger balance for one trade date does not answer a month-end question unless the scope is preserved. The repository's workflows express dates, periods, tickers, packets, and asset classes in their requests for a reason: context that is lost at the connector boundary cannot be recovered by a more eloquent summary. The adapter must preserve the identifiers and scope that the method uses to interpret the result.
Why structured sources matter to a domain method
The repository's skills make source choice part of the calculation or classification rather than a preliminary search task. Comparable-company analysis prioritizes institutional MCP sources when available because the analysis needs consistent definitions and an audit trail.