S&P Global - Adaptive Retrieval

Ask one natural-language question and let S&P Global's router pick the datasets.

Opens Agent Studio, where connecting is one click. The connector URL below works in any MCP client.

Paste it into any MCP client. Setup docs

What the S&P Global - Adaptive Retrieval connector does

The S&P Global Adaptive Retrieval MCP server answers natural-language questions from S&P Global's data with citations. Built by Kensho, it exposes one documented tool: you ask, and a router picks which of 24 data agents to search. It is read-only, and every agent the router reaches spends a billed search credit.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on .

Anthropic states this reflects the level of review a connector received, not a security audit.

S&P Global - Adaptive Retrieval tools (1)

Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.

Limits

  • Every call spends, and the caller does not set the price. Usage is counted per data agent search, and the router decides how many agents to search. Do not grant blanket auto-approval to this tool. Use allowed_datasets to cap the fan-out where you know the destination.
  • No idempotency key is documented. A retried timeout can be a second charge, and Kensho's own example notebook retries up to three times on any client error.
  • The agent cannot see its own spend. Usage is reported only on HTTP headers, which MCP clients do not pass to the model. There is no accounting tool.
  • Anthropic's directory lists zero tool names, so the vendor's own documentation is the only enumeration and no cross-check is possible. One tool is a floor, not a verified total.
  • We could not read tool schemas or safety annotations. The endpoint returned 401 to an anonymous request on 2026-08-23, so the read-only claim rests on Kensho's published OpenAPI specification rather than on the server's own readOnlyHint declarations.
  • The documented agent list is not your entitlement. Kensho searches only datasets included in your subscription and its agents endpoint returns those "accessible to the user". The 24 agents are a catalogue; your contract decides your subset, and only an authenticated call enumerates it.
  • Four datasets in the API schema have no documented agent. credit_ratings, macroeconomics, ratings_research and executive_departures appear in Kensho's UnderlyingDataset enumeration and nowhere in the Adaptive Retrieval documentation. We could not establish whether they are reachable, staged for release, or vestigial.
  • min_release_stage can widen the surface to beta agents. It is an ordinary parameter on the one tool, needing no separate grant, and Kensho documents it as controlling access to experimental features.
  • No licensing terms are published in these docs. There is no redistribution, resale or terms-of-use language anywhere in the Adaptive Retrieval documentation — a different position from the sibling Deterministic Retrieval docs, which at least carry a licensing qualifier. S&P Global's own terms-of-use and privacy pages returned HTTP 403 to every client we tried on 2026-08-23, so we could not check them either. Read your contract.
  • The refresh token is documented as permanent. Kensho's example notebook tells users the access token expires in one hour and that they may "alternatively, use the Refresh Token, which does not expire". Treat it as a long-lived credential and store it accordingly.
  • Multi-step reasoning is a known limitation. Kensho names it as such and ships a separate ReAct agent notebook "to address one of the known limitations with multi-step reasoning". The connector answers questions; it does not chain them.
  • Accuracy of the summary is your side of the line. The API returns cited data, but the sentence an agent writes from that data is the agent's. The citations exist so you can check it.
  • The documentation understates its own availability. Kensho's notebooks page says to email for access to both example notebooks; both are in fact public on GitHub under Apache 2.0.

Frequently asked questions

How many tools does the S&P Global Adaptive Retrieval MCP server have?

Kensho documents exactly one, named search. Anthropic's directory lists no tool names at all for this server, so the vendor's MCP guide is the only enumeration and one is a documented floor rather than a verified total. The guide shows a tools/call payload naming search and says the MCP Inspector's tool list contains it.

What S&P Global data can Adaptive Retrieval reach?

Kensho documents 24 data agents covering Capital IQ Pro and Xpressfeed, plus S&P Global Energy AI Ready Data. Its published API schema names 26 underlying datasets. Coverage includes financials, market data, valuation multiples, transactions, estimates, transcripts, SEC filings, news, key developments, business relationships and executive profiles with compensation.

Does each Adaptive Retrieval query cost money?

Yes, and one question can cost several credits. Kensho states it counts each individual data agent search toward your usage, so a query the router fans across four agents spends four. Contracts carry a quota plus optional billed overage. A question whose agent searches all fail is not counted.

Can the Adaptive Retrieval connector change or delete S&P Global data?

No. Kensho's published OpenAPI specification for the service describes 11 paths with 8 GET and 3 POST operations and no PUT, PATCH or DELETE anywhere. Only one path returns data, POST /api/v2/search, and the other two POSTs are token operations. The MCP tool maps onto that single search path.

How is Adaptive Retrieval different from S&P Global Deterministic Retrieval?

Adaptive Retrieval routes for you; Deterministic Retrieval makes the agent choose. Here you send one natural-language query and Kensho's router picks the datasets from their descriptions. On the sibling server the agent picks a named tool and supplies a company identifier. Anthropic lists both, and Kensho says they can be used together.

Can an agent see how much Adaptive Retrieval quota it has spent?

Not through any tool. Kensho returns usage on HTTP response headers including x-usage-remaining and x-overage-remaining, but headers are transport-level and MCP clients do not surface them to the model. The connector exposes no accounting tool, so an agent can spend the quota with no way to check the balance first.

Can I use S&P Global data retrieved through Adaptive Retrieval in my own product?

Check your S&P Global contract, because the documentation does not say. The Adaptive Retrieval docs contain no licensing, redistribution or terms-of-use language at all. Kensho's example notebooks are Apache 2.0, but that licence covers the notebook code, not the S&P Global data it retrieves.

What do I need before connecting S&P Global Adaptive Retrieval to Claude?

A commercial arrangement and S&P Global OAuth credentials. Kensho's MCP guide states you must have S&P Global OAuth credentials and directs you to email commercial@kensho.com for access. There is no self-serve signup. Kensho's overview page points to the same address for trials and pricing.

Sources

Use it in an agent

Put S&P Global - Adaptive Retrieval to work.

Connect S&P Global - Adaptive Retrieval once and your agents call these tools on their own: on a schedule, in a workflow, with nobody at the keyboard.

Call (650) 285-1019Our AI receptionist answers.