An AI agent that can only reason is useless in production: it needs runtime access to tools and data, and the Model Context Protocol (MCP) collapses the N×M custom-integration problem into a single open standard (N+M). Vendors from Anthropic and Databricks to Red Hat and Atlan now build MCP into their agent tooling, and the protocol reached 110 million monthly SDK downloads by April 2026 across more than 10,000 public servers. Atlan exposes a governed MCP server as the Context Layer for AI beneath the protocol. The catch: MCP moves context, it does not produce or govern it. Connectivity is solved; context quality (lineage, ownership, certification, policy) is the exposed problem.
| Field | Value |
|---|---|
| What it is | An open protocol that standardizes how AI agents connect to tools, data, and prompts at runtime |
| What it solves | N×M custom integrations collapsed to N+M; dynamic tool discovery without reprogramming |
| What it does NOT solve | Whether the data flowing through is certified, current, or correct |
| Governed by | Agentic AI Foundation under the Linux Foundation (since December 2025) |
| Adoption (2026) | 110M monthly SDK downloads; 10,000+ public servers |
| The missing layer | The Context Layer for AI: lineage, ownership, certification, and policy that make agent answers reliable |
Why do AI agents need MCP?
Permalink to “Why do AI agents need MCP?”AI agents need MCP because a model that can reason but cannot reach tools or data cannot do work in production. Before MCP, every connection between an agent and an external system was a bespoke integration: each model paired with each tool meant a custom adapter, a problem that scales as N×M. According to Klavis (2026), MCP collapses N×M custom integrations into N+M by giving every tool and every model one common interface.
The protocol does two things that matter for agents. First, it standardizes runtime tool access, so an agent can call a database, a ticketing system, or a search index through the same contract. Second, it supports dynamic discovery: agents find and invoke tools at runtime without being reprogrammed for each new capability. MCP defines three primitives for this, tools, resources, and prompts, exchanged over JSON-RPC, and Atlan’s explainer on what the Model Context Protocol is covers their mechanics in depth.
Connectivity genuinely was a bottleneck, and MCP unblocked it. The harder question is not whether agents can reach data, but whether the data they reach is trustworthy. That is why teams building on the enterprise context layer treat the protocol as the entry point and the governed context underneath as the actual investment.
How widely adopted is MCP in 2026?
Permalink to “How widely adopted is MCP in 2026?”MCP went from a niche specification to the default agent-integration standard in roughly 18 months. According to DigitalApplied (2026), MCP reached 110 million monthly SDK downloads by April 2026, up from about 2 million at its November 2024 launch, with more than 10,000 active public servers.
The protocol also changed hands. According to Anthropic (2025), Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation in December 2025, and the foundation reached more than 170 member organizations within four months. The handover signals MCP is now neutral, shared infrastructure rather than one vendor’s project.
Adoption is real, not hype. According to a Stacklok survey (2026), 41% of surveyed software organizations run MCP servers in limited or broad production. That production footprint is exactly why the next question carries weight: once connectivity is solved at scale, the binding constraint shifts from access to quality. The trustworthiness of what flows through the pipe is what decides whether agents work.
What MCP solves, and what it leaves unaddressed
Permalink to “What MCP solves, and what it leaves unaddressed”MCP solved the connectivity problem cleanly, and in doing so it exposed a harder one it was never designed to touch. The “why MCP matters” question has two sides. MCP matters because agents need runtime tool access and it delivers that as an open standard. It also matters because, by standardizing the pipe, it makes a question impossible to ignore: is what flows through certified, current, and correct? The cleanest way to see the boundary is to separate what the protocol handles from what it leaves open.
| Dimension | MCP handles | Still unaddressed |
|---|---|---|
| Connectivity | Standard runtime tool and data access (N+M) | Nothing left open here |
| Discovery | Dynamic tool discovery at runtime | Which asset is the right one to trust |
| Access control | Emerging auth and RBAC: who may call a tool | Whether the data the tool returns is certified or current |
| Meaning | Returns schema and table names | Column-level meaning, ownership, lineage |
| Correctness | Moves context to the agent | Does not produce or govern that context |
This is the line the coined anchor captures: MCP moves context. It does not produce it. The protocol is the channel through which an agent reaches data; the Context Layer for AI is what produces and governs the context that travels through it. According to Unwind Data (2026), connected and correct are two completely different things, and a protocol that guarantees the first says nothing about the second.
For teams putting agents into production, the protocol is necessary and now a commodity. The durable investment is the governed context underneath, because that determines whether a connected agent returns a right answer or a confident wrong one.
Why do MCP-only agent projects fail?
Permalink to “Why do MCP-only agent projects fail?”MCP-only agent projects fail because when an agent queries data with no trust signals attached, it guesses, and guessing is where production deployments break down. The protocol delivers the asset; it does not tell the agent whether the asset is certified, current, owned by anyone, or even what its columns mean.
The hard number is forward-looking. According to Gartner (Andres Garcia-Rodeja, Data & Analytics Summit, March 2026, via Unwind Data), by 2028, 60% of agentic analytics projects relying solely on MCP will fail for lack of a consistent semantic layer beneath the protocol. The failure mode is not connectivity. It is the missing layer that gives data meaning.
Practitioners describe the same gap from the ground. According to Scalifi AI (2025), MCP’s open-access default exposes too many tools at once, degrading the agent’s ability to pick the right one, an effect they call “Context Rot.” Most backends return table names without record counts and schema without row-level security state, so the agent fills the gaps with extra queries, retries, and guesses. Forrester (2026) raised the matching concern from the CISO seat, arguing that MCP’s governance gaps are structural, not a configuration detail.
In fairness, the protocol is not standing still. MCP’s authorization working group and a growing layer of gateways are adding access controls. But access control answers who may call a tool, not whether the data it returns is certified, current, or meaningful. The first is necessary; the second decides whether an answer is right.
An internal customer signal shows the gap in practice. In a live evaluation at a medical-device manufacturer, the customer’s CRM assets had no certification, no descriptions, no readmes, and no column-level meaning, so MCP had nothing to rank on. The agent answers anyway: when, for example, a dbt model feeding a revenue dashboard is modified, MCP can fetch the table but cannot tell the agent that the column is uncertified or the lineage is broken. Across external, community, and internal evidence the lesson holds. Connectivity is solved, and context quality is the live problem.
How Atlan approaches MCP for AI agents
Permalink to “How Atlan approaches MCP for AI agents”Atlan treats MCP as the governed runtime interface to the Context Layer for AI, so the context an agent retrieves carries the trust signals it needs to be right. Connected agents still produce wrong answers when the context they pull is uncertified, stale, or stripped of meaning. Access without trust signals is the gap the protocol cannot close on its own.
Atlan exposes a governed MCP server as that runtime interface. The server gives agents search and discovery (search_assets, get_assets_by_dsl), lineage traversal (traverse_lineage), business glossary access, data quality rules, and selective write-back of certification status and descriptions, all backed by the Enterprise Data Graph. These are the exact signals missing in the medical-device evaluation: certification, descriptions, and column-level meaning the agent could rank on. The server is the channel; the Context Layer for AI is what makes the delivered context trustworthy. The difference shows up signal by signal.
| Trust signal an agent needs | Pipe only (raw MCP) | Pipe plus Context Layer for AI |
|---|---|---|
| Certification status | Absent | Travels with the query |
| Column-level meaning | Schema names only | Glossary and lineage attached |
| Ownership | Unknown | Resolved to the owner |
| Freshness and quality | No signal | Data quality state included |
The outcome is the conviction made concrete. One enterprise in HR and finance technology described the result as teaching AI a shared language: the vocabulary it built among people can now be used by AI through Atlan’s MCP server. An electronics distributor described Atlan as a context operating system feeding an MCP server that delivers context to AI models. In both cases lineage is the proof point, because it tells the agent not just what an asset is, but where it came from and whether it can be trusted. For a fuller view of the server’s capabilities, see what Atlan MCP is.
Why the context layer is the durable bet, not the protocol
Permalink to “Why the context layer is the durable bet, not the protocol”MCP matters precisely because it solved a real problem. Agents needed runtime tool access, the integration math was punishing, and the protocol turned N×M into N+M and won near-universal adoption in under two years. That win is also what makes the next problem visible. Once every agent can reach every tool, the question stops being “can it connect” and becomes “is what it retrieves certified, current, and correct.” MCP moves context; it does not produce or govern it, and adding access-control primitives does not change that. The investment that pays off today, and that the Gartner and Forrester evidence points to next, is the Context Layer for AI that produces and governs what the pipe carries. Lineage is the clearest proof: an agent can only trust an answer when it can trace where the data came from. The protocol will keep evolving, and it should. The governed context beneath it is what turns a connected agent into a correct one.
Real stories from real customers: Context delivered through MCP
Permalink to “Real stories from real customers: Context delivered through MCP”"We're excited to build the future of AI governance with Atlan. All of the work that we did to get to a shared language at Workday can be leveraged by AI via Atlan's MCP server…as part of Atlan's AI Labs, we're co-building the semantic layer that AI needs with new constructs, like context products."
— Joe DosSantos, VP of Enterprise Data & Analytics, Workday
"Atlan is much more than a catalog of catalogs. It's more of a context operating system…Atlan enabled us to easily activate metadata for everything from discovery in the marketplace to AI governance to data quality to an MCP server delivering context to AI models."
— Sridher Arumugham, Chief Data & Analytics Officer, DigiKey
The protocol question and the context question are not the same investment. Choosing or adopting MCP is a connectivity decision; building the governed context beneath it is the decision that determines whether agents are trustworthy in production. Teams that sequence this correctly, governed context first, protocol second, make every downstream protocol choice reversible.
FAQs about MCP for AI agents
Permalink to “FAQs about MCP for AI agents”1. What problem does MCP solve for AI agents?
Permalink to “1. What problem does MCP solve for AI agents?”MCP gives agents runtime access to tools and data through one open standard. Before MCP, every model-to-tool connection was a custom integration that scaled as N×M. MCP collapses that into N+M, so an agent can discover and call any compliant tool dynamically without being reprogrammed for each new system.
2. Is MCP secure for production AI agents?
Permalink to “2. Is MCP secure for production AI agents?”MCP is a connectivity standard, not a security model. According to Integrate.io (2026), 53% of MCP integrations use static API keys and only 8.5% use OAuth, so authentication is still maturing. Access control and data governance are separate concerns, solved in the layer beneath the protocol.
3. Does MCP guarantee that agent answers are correct?
Permalink to “3. Does MCP guarantee that agent answers are correct?”No. MCP moves context to the agent; it does not certify that the context is accurate or current. Correctness depends on the governed context layer underneath: the lineage, ownership, certification, and quality signals that tell the agent whether the data it retrieved can be trusted.
4. Who governs MCP now?
Permalink to “4. Who governs MCP now?”The Agentic AI Foundation, under the Linux Foundation, has governed MCP since December 2025, when Anthropic donated the protocol. The handover moved MCP from a single vendor’s project to neutral, shared infrastructure backed by a broad membership of organizations.
5. Can AI agents work without MCP?
Permalink to “5. Can AI agents work without MCP?”Yes, through bespoke integrations, but that re-creates the N×M problem MCP was built to eliminate. Each new tool needs a custom adapter for each model, which does not scale. MCP exists so agents can reach many tools through one standard contract instead.
6. Why do MCP-only agent projects fail?
Permalink to “6. Why do MCP-only agent projects fail?”They fail for lack of governed context, not connectivity. According to Gartner (2026), 60% of agentic analytics projects relying solely on MCP will fail by 2028 without a consistent semantic layer beneath. When trust signals like certification and lineage are absent, agents guess, and guesses do not survive production.
Sources
Permalink to “Sources”- MCP solving the N×M integration problem, Klavis
- MCP adoption statistics 2026, DigitalApplied
- Donating the Model Context Protocol and establishing the Agentic AI Foundation, Anthropic
- MCP report, Stacklok via Zuplo
- MCP and the semantic layer, Unwind Data
- Six Fatal Flaws of the Model Context Protocol, Scalifi AI
- MCP doesn’t stand for many critical problems, but maybe it should for CISOs, Forrester
- Best MCP gateways and AI agent security tools, Integrate.io
