What G2 Reviews Say About the Setup Paradox in MCP Server

MCP Server setup has two parts that are easy to confuse. Connecting an MCP server, so an AI agent can make its first tool call, can take minutes. Preparing it for business use, with the right permissions, policies, and user rollout, takes much longer. MCP server infrastructure platforms are the supporting layer that hosts, governs, and secures those servers, and vendors promise to make both parts fast.

G2's analysis of 190+ verified reviews of these platforms tells a messier story: MCP server setup is praised in 27% of reviews and criticized in exactly 27%. Take out the single largest vendor, and complaints win, 30% to 22%. The gap is not random. It lands hardest on enterprises and on end users who did not configure the platform themselves.

Do MCP server platforms really set up in minutes?

Connecting an MCP server takes minutes, but establishing full operational deployment requires more time. According to G2 Data, 52 verified reviews list setup as a like and 52 list it as a dislike, making it the only theme where praise and complaints are tied. Compare that with ease of use (46% positive, 8% negative) or integrations (36% positive, 19% negative). On the positive side, one verified reviewer said requests that used to sit in the IT queue for weeks now go live in minutes. Another built an end-to-end SQL agent in 30 minutes.

The two sides are not describing the same thing. According to G2 Data, 54% of the setup comments in "like" answers are about speed. In "dislike" answers, 72% are about configuration, planning, or understanding the platform. The complaints are not about installation. They are about decisions buyers did not expect to make. 


One verified administrator said teams "need to think carefully about how to organize MCPs, tools, skills, toolkits, IdP groups, and authentication methods." A verified consultant said setup took time to understand, "especially around governance policies and MCP configuration."

In MCP, setup is not one step. It is four phases, and vendor demos usually show only the first:

• Connect: Register an MCP server, sign in with OAuth, and make the first tool call from an agent. This is where reviewers praise speed.
• Scope: Decide which tools each agent and team may call, map them to identity provider groups, and separate read actions from write actions.
• Govern: Set approval rules, audit logging, and data policies that pass a security review and stream to the SIEM.
• Roll out: Help end users find approved tools, trust them, and know what they still have to build.

Connecting the first agent is fast. Deciding who can do what, across every tool, is the slow part. G2 reviews suggest buyers should budget for phases two and three, not just phase one.

Who Struggles Most With The MCP Server Setup?

According to G2 Data, the MCP learning curve grows with company size. Small business reviewers cite it in 11% of reviews, mid-market in 14%, and enterprise in 22%, double the small business rate. The reason is scale in three places at once:

• Systems: When mid-market and enterprise reviewers name connected business systems such as Jira, GitHub, Salesforce, and Snowflake, they name 2.4 systems on average, compared with 1.1 for small business reviewers. Each system adds its own MCP server, and each server exposes many tools.
• Identity groups: Larger companies run access through identity providers with many groups. Every MCP tool has to be mapped to the right groups, or agents get too much access or too little.
• Policies: Each tool needs rules: read or write, approval or no approval, and what gets logged. Permission decisions multiply as systems × tools × groups. A company with 10 systems, 15 tools each, and 5 groups faces 750 tool-to-group decisions before any policy is written.

G2 Data backs this up. None of the 36 small business reviewers complain about permissions, identity, or access policies. Among mid-market and enterprise reviewers, 12% do.

The bigger gap is by role. Only 4% of verified administrators mention a learning curve. Among end users, it is 19%. This suggests that the challenges of using an MCP platform may differ from those involved in configuring it. For teams evaluating MCP infrastructure, the end-user experience deserves as much attention as the initial administrator setup.

Does it matter if a platform was built for MCP?

G2's category definition says it does. To qualify, MCP must be a product's main interface, "not a REST or GraphQL API with an MCP wrapper added later." Verified reviews back that line up.


According to G2 Data, platforms built around MCP (Willow, Composio, and AdPlug MCP) average 4.92 stars across 106 verified reviews. Data platforms that added MCP to an older product (K2View and CData Connectors) average 4.51 across 70. Learning-curve complaints are three times higher on the adapted side (30% vs. 10%). And not one of those 70 reviews mentions MCP at all. Buyers are still reviewing them as data integration tools. One verified G2 user said, "There are too many elements to use, and it isn't clear what each one is for."

The takeaway for buyers: An "MCP support" checkbox does not tell buyers how a product will behave for an agent team. The difference is operational:

Four questions that settle doubts in a demo:

• Can we allow one MCP tool and block another for the same user?
• Does every tool call log the human identity behind the agent?
• Are tools designed for agent tasks, or copied from API endpoints?
• Does the product support the July 28, 2026, stateless MCP specification?

Connecting an agent is only the beginning. As MCP evolves, how much implementation complexity a platform absorbs versus leaves with the buyer is becoming a key evaluation criterion.

Is MCP replacing APIs, and what does Stateless MCP change?

MCP is not replacing APIs. Most MCP servers still call a REST API underneath. What MCP replaces is the custom connector work on top, so any compliant agent can discover and call a tool without one-off code.

MCP is now governed by the Linux Foundation’s Agentic AI Foundation, co-founded by Anthropic, Block, and OpenAI. For buyers standardizing on MCP, that reduces lock-in risk. Spec changes go through an open proposal process, and the July 2026 revision introduced a formal lifecycle with a minimum 12-month deprecation window, so teams can plan upgrades on a schedule. The four main SDKs, TypeScript, Python, Go, and C#, already support the new revision with backward compatibility.

What changed in the July 28, 2026, stateless MCP specification?

The 2026-07-28 MCP specification removed protocol-level sessions. In plain terms: an MCP server no longer remembers a conversation between requests.  Before, a connection opened with a handshake and kept a session ID for its lifetime. Now every request carries its own context, and any server that needs to remember something must return an explicit ID such as a workflow or cart ID that the agent passes into the next call.

What extra setup work does stateless MCP create, and who handles it?

The changes in the official changelog make hosting simpler but move several responsibilities into the setup phase. Each item lands with a different owner:

For buyers, this connects straight back to the setup question. Stateless MCP makes the connect phase easier and adds work to the scope and govern phases, the same configuration and planning work behind 72% of setup complaints in G2 reviews. The practical test for any MCP server infrastructure platform is how many rows in this table it takes off the application team's plate.

Frequently asked questions (FAQs) about MCP server infrastructure

Q1. What is an MCP server, and how do you set one up?

An MCP server is a program that exposes tools, data, and prompts to AI agents through the Model Context Protocol. A basic MCP server setup has four steps: choose or build a server with an official SDK, run it locally or at a remote HTTPS endpoint, register it in an MCP client such as an AI assistant or IDE, and sign in with OAuth. Business use adds a fifth step, scoping permissions, which is where G2 reviewers report the most friction.

Q2. What are the infrastructure requirements for an MCP server?

A local MCP server needs only the runtime for its SDK, such as Node.js or Python, and an MCP client. A remote server for a team needs an HTTPS endpoint using the Streamable HTTP transport, an OAuth-compatible authorization server, and logging. Since the July 2026 specification, it no longer needs sticky sessions or a shared session store. Business deployments add identity provider integration, per-tool access control, and audit logs, the layer MCP server infrastructure platforms provide.

Q3. What are examples of MCP servers?

The MCP steering group maintains reference servers including Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, and Time. Published servers are listed in the official MCP Registry. Commercial examples on G2 include AdPlug MCP, a hosted MCP server for Google Ads and LinkedIn Ads, and K2View, which provides an MCP server for each data product.

Q4. Where should you host MCP servers?

Host an MCP server locally when one person uses it from a desktop tool. Host it remotely when a team or production agent needs it: in containers, on Kubernetes, on serverless platforms such as AWS Lambda, or on a managed MCP server infrastructure platform. The stateless 2026 specification makes serverless and load-balanced hosting simpler because requests no longer depend on a session.A5. 

Q5. Is there a free MCP server for testing?

Yes. The official reference servers are open source and free, and the Everything server is built for testing prompts, resources, and tools. The maintainers state that reference servers are educational examples, not production-ready solutions.

Q6. Are MCP servers safe?

An MCP server is as safe as its permissions, authentication, and code. The official MCP security best practices name specific risks, including confused deputy attacks, token passthrough, server-side request forgery, state handle hijacking, and compromised local servers. The core defenses are least-privilege scopes, tokens validated for the specific server, sandboxed local servers, state handles bound to the signed-in user, and audit logs. Teams with stricter needs often add AI security solutions.

Q7. How do I secure my MCP server?

Start with minimal OAuth scopes and request more only when a tool needs them, never pass a client's token through to a downstream API, require HTTPS, and log every tool call with the user behind it. For local servers, show users the exact command before it runs and execute it in a sandbox. These steps come from the official MCP security best practices.

Q8. Is an MCP server like an API gateway?

No. An MCP server exposes capabilities to AI agents, while an API gateway sits in front of APIs to route, authenticate, and rate-limit traffic. The two meet in MCP gateways: some MCP server infrastructure platforms sit in front of many MCP servers to enforce policy, a role that overlaps with AI gateways. The 2026 specification's Mcp-Method and Mcp-Name headers make per-tool gateway rules easier to apply.

The bottom line on the MCP server infrastructure

According to G2's analysis of 190+ verified reviews, connecting an MCP server is fast, but getting a team ready to use it is not. MCP server setup is praised and criticized in equal measure. Learning-curve complaints double at enterprise scale and are nearly five times higher among end users than administrators.

The July 2026 stateless MCP specification widens that gap. It makes hosting easier, but it adds idempotency keys for tools that send, pay, or provision; explicit handles in place of sessions; and retry flows for any tool that asks a user for input. Every one of those lands in the same configuration and planning work that G2 reviewers already flag.

Before choosing an MCP server infrastructure platform, buyers should check three things:

• Setup support: Does the vendor provide guided onboarding, templates, and help migrating to the 2026-07-28 specification? Which of the stateless tasks, such as retries, idempotency, and state handles, does the platform handle by default?
• Permissions: Can access be set per MCP tool and per user, mapped to identity provider groups, with least-privilege defaults and every call logged to a named person?
• Rollout: How will end users find approved tools and learn what each one does? Ask to see the end-user view, not just the admin console.

The question for 2026 is no longer how fast an agent can connect. It is how fast a whole team can use it safely, and what happens when the same request runs twice.

Can NLG software deliver quality as fast as it delivers content? Explore G2’s latest research to see what verified buyers say about speed, ROI, and where human editing still matters.