The MCP connector race: how Model Context Protocol is becoming the plumbing layer between research vendors and buy-side agents

A quiet standardisation fight is underway in the plumbing beneath buy-side AI, and it will shape vendor economics in financial research for the next five years. The surface story is the agent-runtime arms race: Rogo's funding round, Hebbia's connector push, S&P Global's ChatIQ expansion, every incumbent terminal retrofitting a chat box. The deeper story is the integration contract those agents use to reach third-party data. Our read is that Model Context Protocol, the open specification Anthropic shipped in November 2024, is becoming that contract by default, and the expert networks and transcript vendors that ship native MCP servers early will define the resource shapes the rest of the category has to conform to.

This matters because standards fights in financial data are rarely re-run. The SEC's structured 10-K stack, FIX for order routing, FpML for derivatives: each settled once and then calcified for a decade or more. MCP looks like it is settling now, in public, with surprisingly little coverage inside the research-vendor conversation.

From Anthropic release to default integration contract

When Anthropic open-sourced MCP in late 2024, the framing was modest: a standard way for an LLM-driven agent to reach external tools, files, databases, and APIs through a single protocol rather than a bespoke integration per vendor. The reference documentation describes three primitives (resources, tools, and prompts) exposed by a server and consumed by any compliant client. That is unglamorous engineering. It is also, in practice, the shape of every integration contract that has ever mattered in financial infrastructure.

The critical event was not the launch. It was the cross-lab adoption through 2025. OpenAI shipping MCP support in its Agents SDK in March 2025 broke the usual pattern where each lab tries to own the integration surface. Google Gemini and Microsoft Copilot Studio followed in the same year. For the first time since the LLM boom started, a single protocol runs across the four model providers that enterprises actually buy. From a procurement perspective, that is the moment a protocol stops being a bet and starts being a floor requirement.

Adoption inside developer tooling moved first (GitHub, Linear, Sentry, Replit), then enterprise data (Snowflake and Databricks shipped MCP endpoints to let agents query warehouses directly), and now the pattern is reaching financial research. The direction of travel is clear, and the question for research vendors is no longer whether to ship an MCP server, but what their resource schema looks like when they do.

Why the buy-side agent layer forces the question

The demand side is where this becomes concrete. The last 18 months of buy-side AI product launches share a common architectural tell: they are moving from retrieval-augmented chat over a vendor's own corpus to orchestrated agents that reach across many corpora. Rogo's October 2026 Series D, Hebbia's connector and persistent-memory release, and S&P Global's ChatIQ rollout inside Capital IQ Pro are three versions of the same thesis. The runtime is multi-tool, multi-source, and the analyst's question routes to whichever combination of filings, transcripts, consultant notes, pricing data, and internal memos actually answers it.

That architecture breaks the old integration model. In a world of one chat app per vendor, each research provider could ship its own UI and own the surface. In a world of a few dominant agent runtimes orchestrating across vendors, the research provider is a data source whose shape is defined by whatever contract the runtime speaks. If the runtime speaks MCP (and increasingly all of them do), the vendor that ships a well-formed MCP server is addressable, and the vendor that does not is a tab the analyst opens in a browser.

The second-order effect is a quiet re-pricing of distribution. Historically, a research vendor's distribution moat was its UI, its seat count, and its workflow integrations into Excel and PowerPoint. In an agent-mediated world, the moat is the quality of the structured representation the vendor exposes to any compliant client. That is a very different capability to build, and the vendors that treat it as a 2027 problem will find that the schema has already been set by someone else.

The research-vendor surface that MCP has to represent

This is where the schema fight gets interesting, and where our read is that early movers will carry disproportionate weight. The qualitative-research corpus is not a single resource type. A single expert-network transaction involves at minimum: the consultant profile (with employer history, redactions, and compliance flags), the call transcript (with speaker tags, segment IDs, timestamps, and in some vendors, confidence scores), the request record (who asked, under what entitlement, with what disclosed holdings), and the audit event stream that satisfies the fund's and the vendor's WORM retention obligations.

Expose that naively and the agent cannot reason about it. Expose it well and the agent can answer questions like, give me every passage from a former supply-chain lead at a hyperscaler discussing capex pacing in Q3 2026, restricted to transcripts my fund has consumption rights to, with inline citations and a per-passage compliance flag. That query is the actual product. The MCP server is how it becomes answerable.

There is no settled convention yet for any of this. The transcript vendors (AlphaSense, Tegus, Quartr, Aiera's post-rebrand BlueMatrix franchise, and the transcript libraries bundled inside the major expert networks) each have internal schemas optimised for their own UI. Lifting those schemas into MCP resource shapes is not a port. It is a design decision about what the agent is allowed to see, how it is entitled to see it, and how the audit trail flows back.

A few specific design questions are already live:

Segment identity and speaker modelling

An earnings-call transcript or an expert interview is a sequence of speaker turns, each with a role, a timestamp, and often a confidence score on the diarisation. The MCP question is whether a segment is a first-class resource (addressable, citeable, revocable) or an attribute of a parent transcript. The answer determines whether an agent can cite a single quote with a stable identifier six months later, which in turn determines whether the vendor can be used in a compliance-reviewable research memo. Our view is that per-segment addressability is table stakes, and vendors that only expose transcript-level resources will be reworked within a year.

Entitlement scoping

Research libraries are sold under complex seat, consumption, and firm-level licence structures. An MCP server has to enforce those licences at tool-call time, not at ingestion time, because the agent may be acting on behalf of any user in the firm. The protocol itself is auth-agnostic (OAuth-style flows are the common pattern), but the semantics of what a token entitles the agent to read are vendor-defined. This is where the schema fight gets commercial: a vendor that defines entitlement scoping cleanly can protect its unit economics, and one that does not will discover its content being over-consumed inside an agent runtime it does not control.

Audit emission

Every expert-network transaction and every regulated research consumption event generates a compliance record. In the agent world, the question is who emits the audit event, when, and in what format. If the agent runtime emits it, the vendor loses visibility into how its content is being used. If the MCP server emits it on every tool call, the vendor keeps the audit record but has to design for high-cardinality event streams. Our read is that per-query audit emission from the server side is going to become the compliance-defensible pattern, and it is going to be a differentiator in buy-side RFPs within 12 to 18 months.

Redaction and compliance flags

Expert-network transcripts carry redaction markers (names stripped, non-public figures removed, employer references obscured). An MCP representation has to carry those markers through to the agent in a way the agent cannot bypass, including when the agent summarises or re-expresses the content. That is a non-trivial constraint and will probably require tool-level enforcement rather than resource-level metadata, since metadata is advisory and tools are enforcing.

The SEC-filings parallel, and why first-mover schema wins

The closest historical analogue is the structured filings stack. When the SEC moved to EDGAR and later to inline XBRL, the vendors that defined the early parsing conventions (Edgar Online, then the data layers inside Compustat, S&P Capital IQ, FactSet) effectively set the template that every downstream tool inherited. Later entrants had to conform. The schema was not neutral: it reflected the priorities and blind spots of whoever shipped first, and those priorities became industry-wide defaults.

The MCP race in research data has the same structural property. The first two or three vendors to ship durable, well-documented MCP servers for transcripts and consultant data will not just be addressable to agent runtimes earlier. They will define what the field names are called, how speaker roles are enumerated, how entitlement is scoped, and how audit events are emitted. Vendors that ship later will either conform or build adapters, and the adapter layer is where margin goes to die.

There is one meaningful difference from the EDGAR precedent. The SEC was a central standard-setter, so the schema fight happened once and then stopped. MCP has no equivalent. The protocol is open, the resource shapes are not standardised across domains, and the de facto schema will emerge from whoever ships first and gets consumed by Rogo, Hebbia, ChatIQ, and whatever agent runtimes come next. That is a messier process and favours vendors with strong engineering benches and clear product opinions over those relying on standards-body consensus to arrive.

What this changes for expert networks specifically

The expert-network category sits in an unusual spot. The transcript libraries at GLG, Guidepoint, Third Bridge, Dialectica, and the standalone transcript franchises (AlphaSense's Tegus-derived corpus, Stream from AlphaSense, BlueMatrix's Aiera assets) represent some of the highest-density qualitative data in finance. They are also the most compliance-sensitive, which has historically slowed their integration into third-party tools. MCP changes the calculus in two ways.

First, the integration cost drops. Shipping an MCP server is cheaper than negotiating bespoke API integrations with each buy-side AI vendor, and the resulting surface is reusable across every compliant client. That is a one-to-many distribution model expert networks have never really had.

Second, the compliance surface becomes more defensible, not less, if the schema is designed carefully. A well-formed MCP server can enforce per-call entitlement, emit per-query audit events, carry redaction flags through to the client, and make it structurally harder for an agent to over-consume or mis-cite content than a human analyst with a PDF export can. The compliance team's instinct will be to resist, but the long-run position is probably stronger inside MCP than outside it.

The risk is the mirror image. An expert network that waits, or ships a shallow MCP server (transcript search only, no segment addressability, no audit emission, no entitlement scoping), will find its content exposed to agent runtimes on terms set by someone else's schema, and will spend the following two years negotiating from a weaker position.

The procurement signal: MCP next to SSO and SCIM

The clearest leading indicator that this has moved from a developer-tools curiosity to a procurement requirement is the appearance of MCP-related questions in buy-side vendor RFPs. In the enterprise SaaS stack, SSO (single sign-on) and SCIM (user provisioning) are the two protocol checkboxes that gate every serious deal above a certain size. A vendor without them is not evaluable.

Our view is that MCP connector maturity is becoming the third box on that list for any vendor whose content is intended to be consumed inside an agent runtime, and the first wave of RFPs asking about it in detail (not just, do you have one, but what resources, what tool names, what auth model, what audit emission) is already visible in the market. That is the moment a developer protocol becomes a line item. It is also the moment vendors without a credible MCP roadmap start losing bake-offs they would otherwise have won on content quality alone.

What to watch over the next 12 months

A few specific signals will tell us whether this thesis is playing out:

• Which expert networks and transcript vendors publish MCP server documentation (not just announcements) with per-segment resource addressability and per-query audit emission.
• Whether Rogo, Hebbia, and ChatIQ publish lists of certified MCP integrations, which would effectively create a two-sided marketplace with the agent runtimes as gatekeepers.
• Whether any standards work emerges around a shared vocabulary for financial-research MCP resources (speaker roles, entitlement scopes, redaction taxonomies), or whether the de facto schema calcifies around whichever vendor ships first.
• How compliance teams at tier-one asset managers treat MCP-mediated consumption in their internal research-use policies, since their position will set the ceiling on adoption.