MCP Security Solutions: 12 Best MCP Gateways, Scanners and Access Controls for 2026
MCP security is the set of controls that decide which Model Context Protocol servers an AI agent may connect to, which tools it may call on them, with whose credentials, and what gets logged. The 12 MCP security solutions below do that job in three ways: gateways that sit in the path of every tool call, scanners and registries that vet MCP servers before anyone installs them, and identity layers that handle OAuth and keep tokens away from the agent. We ranked them on how much of that path each one actually controls today, using the vendors' own documentation and trade press, and we note where a feature is still in preview.
MCP lets an AI agent discover and call tools on other systems: a GitHub server, a CRM, a database, a file system. Every connected server adds tool descriptions to the model's context and gets the power to act with someone's credentials. That makes MCP server security a mix of supply chain risk (is this server what it claims to be), access control (should this user's agent call this tool at all) and runtime inspection (is this tool call or tool output carrying an attack).
The tools on this page cover those three jobs in different proportions. Gateways and portals give you one choke point to enforce policy and log every call. Scanners, catalogs and registries vet servers and catch poisoned tool descriptions before an agent loads them. Identity layers implement the OAuth 2.1 flows the MCP specification expects and swap tokens so agents only get scoped, short-lived access. Most teams end up with one product from the first group and one from the second or third.
MCP security sits inside the wider job of AI agent security, which also covers agent inventory, identity and runtime behavior outside MCP. If you are still deciding between a broad agent platform and a focused MCP control, read this ranking next to that one.
MCP gateway
A proxy between agents and MCP servers that authenticates callers, filters tools, inspects calls and logs them, for example Cloudflare MCP server portals, Kong AI Gateway, agentgateway or the Prompt Security MCP gateway.
MCP server scanner
A tool that reads installed MCP configurations and tool descriptions and flags poisoning, shadowing, prompt injection and risky permissions, for example Snyk Agent Scan or the scanner built into Lasso's gateway.
MCP registry or catalog
A curated list of approved MCP servers with provenance data, for example the Docker MCP Catalog or a private registry in Azure API Center.
Agent-to-tool access control
Identity products that issue and exchange tokens for MCP connections and enforce per-user, per-agent and per-tool policy, for example Aembit or Descope.
Compare
MCP security tools: all 12 at a glance
Every company on this page with its focus, who it suits and how it prices. Scores are out of 10.
MCP governance in Azure API Management and API Center, plus Windows containment
Azure-centric organizations exposing internal APIs and MCP servers to agents
Azure API Management tier pricing; Windows features in preview
Redmond, Washington
7.6
Reviews
The top MCP security companies, reviewed
Each review covers what the company sells, who it suits, how it prices and where it falls short.
Rank #1
1. Cloudflare
MCP server portals that put approved MCP servers behind Zero Trust access
9.3
Cloudflare's MCP server portals reached general availability on September 24, 2026, after an open beta that began in August 2025. A portal gives users a single endpoint for the MCP servers an admin has approved, Cloudflare Access logs each tool, prompt and resource request, and portal traffic can be routed through Cloudflare Gateway for HTTP logging and data loss prevention.
Best for
Organizations on Cloudflare One that want one governed endpoint for every remote MCP server
Headquarters
San Francisco, California
Founded
2009
Ownership
Public (NYSE: NET)
Key products
Cloudflare One MCP server portals, Cloudflare Access, Cloudflare Gateway (HTTP logging and DLP), Code Mode
Pricing model
Part of Cloudflare One; Logpush export of portal logs requires an Enterprise plan
The portal model fits the way most companies already handle SaaS access. Admins add remote MCP servers, attach Access policies that decide who sees each one, and choose exactly which tools and prompt templates each portal exposes. Tool aliases let admins rename tools and rewrite descriptions without touching the upstream server, which also helps when a third-party description is vague or noisy. Since the beta, Cloudflare has added service tokens for autonomous agents, static OAuth client credentials for providers that do not support Dynamic Client Registration, session management and Logpush export to a SIEM.
Code Mode is the unusual part. It collapses all upstream tools into a search tool and a code execution tool, and the agent's code runs in an isolated Dynamic Worker, which cuts tool definitions in the context window. The documented limits matter for planning: only remote HTTP servers are supported, so stdio-only servers must be self-hosted behind HTTP; blocked users can still reach a server by its direct URL unless Access is that server's OAuth provider; independent MFA and purpose justification are not enforced for servers authorized through a portal; and each portal holds up to 120 servers with a 60-second tool call timeout.
Strengths
One logged endpoint for all approved remote MCP servers
Per-portal tool curation, aliases and DLP via Gateway
Service tokens for machine-to-machine agent access
Watch out for
Remote HTTP servers only; stdio servers need hosting work
Direct server URLs bypass portal policy unless Access is the OAuth provider
MCP gateway and discovery that risk-scores every MCP server agents touch
9.1
Prompt Security launched its MCP Gateway in May 2025, and SentinelOne signed an agreement to buy Prompt Security in August 2025. The gateway sits between AI applications and MCP servers, intercepts calls, prompt templates and responses, gives each server a dynamic risk score and enforces allow, block, filter or redact actions.
Best for
Security teams that want shadow MCP discovery and enforcement in the same console as endpoint protection
Headquarters
Mountain View, California
Founded
2013
Ownership
Public (NYSE: S); Prompt Security acquired in 2025
Key products
Prompt for Agentic AI Security (MCP Gateway, MCP discovery and risk scoring), Singularity Platform
Pricing model
Quote only
The product's strength is the endpoint view. At launch, Prompt enforced MCP policy through a lightweight endpoint agent or a reverse proxy for homegrown applications, and SentinelOne now pairs the MCP gateway with a browser extension, security plugins and API integrations so agent surfaces that never touch the network perimeter still show up. SentinelOne says the service has risk-assessed more than 15,000 MCP servers, and the console draws a relationship map of which agents connect to which servers and systems.
SentinelOne's May 2026 description of Prompt for Agentic AI Security lists four MCP controls: discovery and governance of sanctioned and shadow servers, risk-based enforcement before agents act, runtime prompt injection blocking on tool calls, and prevention of known malicious servers. Buyers who do not run Singularity can still buy Prompt, but the joined-up reporting is the main reason to pick it over a standalone gateway, so ask which parts need the wider platform.
Strengths
Finds shadow MCP servers on endpoints, sanctioned or not
Per-server risk scores with allow, block, filter and redact
Runs in the same console as SentinelOne endpoint and cloud security
Signed MCP server catalog plus an open-source gateway that runs servers in containers
8.9
Docker runs one of the largest public MCP server catalogs and an open-source MCP Gateway released in July 2025. Docker-built servers come with cryptographic signatures, SBOMs, provenance attestations and continuous vulnerability scanning, and since December 2025 every Docker-built server in the registry is pinned to a specific Git commit.
Best for
Developer and platform teams that want vetted MCP servers isolated in containers by default
MCP Gateway is open source; Catalog and Toolkit ship with Docker Desktop under Docker's subscription terms
Docker treats an MCP server like any other piece of software supply chain. Its December 2025 changes added commit pinning enforced in CI, a nightly workflow that opens a pull request when upstream code changes, an agentic security review of each change using Claude Code and OpenAI Codex with labels such as security-blocked, and publisher trust levels in the catalog so users can tell a known publisher from a community entry. The catalog had passed one million pulls by July 2025.
The gateway's published security model is specific. Signature verification is on by default for images in Docker's mcp namespace, the secret-blocking interceptor scans tool arguments and responses by default, servers get no host environment variables and run with no-new-privileges and CPU and memory limits, host bind mounts are read-only by default, and remote MCP URLs must be public HTTPS with private and metadata IP ranges rejected. Network egress is not blocked by default, so turn on the block-network option or per-server allow lists for anything sensitive.
Strengths
Provenance, SBOMs and commit pinning for Docker-built servers
Container isolation with secret scanning on by default
Open-source gateway usable outside Docker Desktop
Watch out for
Community-built servers get isolation but not Docker's build controls
Enterprise MCP gateway that enforces identity, policy and audit on each tool call
8.7
Snowflake announced its intent to acquire Natoma, an enterprise MCP platform, on May 27, 2026, and the deal closed in June 2026. Natoma's centralized MCP gateway records who requested each action, what permissions they hold and whether the action is allowed, and Snowflake is folding it into Cortex Agents, Snowflake Intelligence and Cortex Code.
Best for
Snowflake customers connecting Cortex agents to SaaS apps, internal APIs and databases
Headquarters
San Francisco, California (Natoma)
Founded
2024
Ownership
Natoma part of Snowflake (NYSE: SNOW) since June 2026
Key products
Natoma MCP gateway, verified MCP server library, Snowflake Cortex Agents, Snowflake Intelligence and Cortex Code integration
Pricing model
Quote only
Natoma was built by a team with identity governance and privileged access management backgrounds, and it shows in the design: the control point is the tool call, tied to the identity of the person behind the agent. Snowflake's plan is to give customers a verified library of MCP servers across SaaS applications, cloud environments, VPCs and on-premises systems, with policy and complete audit for every connection, so agents can pull context from Slack, email, CRM and Jira into Snowflake workflows.
The catch is scope. The Register reported the deal brought about 20 employees, and Snowflake says Natoma's capabilities will be integrated into the AI Data Cloud over time. If you are not a Snowflake customer, ask whether Natoma will still be sold on its own and how long existing standalone contracts are supported. For Snowflake-centric data teams, it is the most natural MCP control on this list.
Strengths
Identity-aware policy and audit at the tool-call level
Verified MCP server library across SaaS and on-premises systems
Native path into Snowflake Cortex agents
Watch out for
Roadmap now tied to Snowflake
Standalone availability for non-Snowflake buyers unclear
Open-source MCP gateway with guardrail plugins and a pre-load server scanner
8.6
Lasso released an open-source MCP Gateway in April 2025 that proxies all MCP interactions, applies guardrails to requests and responses, and scans servers before loading them. A built-in scanner checks server reputation using marketplace and GitHub data and inspects tool descriptions for hidden instructions.
Best for
Teams that want to start with a free, inspectable MCP proxy and add commercial guardrails later
Headquarters
Tel Aviv, Israel and New York, New York
Founded
2023
Ownership
Private
Key products
Lasso MCP Gateway (open source, MIT), Lasso AI Security Platform, Claude Code prompt injection defender hooks
Pricing model
Gateway free and open source; Lasso guardrail plugin and platform by quote
The gateway is a plugin system. The basic plugin masks tokens and secrets, a Presidio plugin masks personal data, and the Lasso plugin calls Lasso's API for custom policy, prompt injection and harmful content checks. The security scanner reads reputation from Smithery, npm and GitHub, looks for hidden instructions and sensitive file patterns in tool descriptions, blocks servers below a reputation threshold and writes the results back into the MCP configuration file. It works with Cursor and Claude Desktop configuration files out of the box.
Lasso also ships open-source Claude Code hooks that scan tool outputs, including MCP tool output, for indirect prompt injection and warn the model. That defender warns rather than blocks by design, so treat it as a signal and not as enforcement. Lasso raised a further $30 million in September 2026 alongside a CPU-only guardrail model, which should help self-hosted deployments.
Strengths
Free MIT-licensed gateway you can read and modify
Reputation and description scanning before a server loads
Commercial guardrails available as a plugin
Watch out for
Advanced guardrails need a Lasso API key and contract
Open-source project depends on community upkeep for enterprise features
MCP server and agent skill scanner born from Invariant Labs' MCP-Scan
8.5
Snyk acquired Invariant Labs, a Swiss AI security research firm, on June 24, 2025. Invariant's researchers named attacks such as tool poisoning and MCP rug pulls, and their MCP-Scan tool now lives on as Snyk Agent Scan, which discovers agents, MCP servers and skills on a machine and scans them for prompt injection and other risks.
Best for
AppSec teams that want to inventory and scan MCP servers and skills across developer machines
Agent Scan CLI free; fleet reporting through Snyk Evo by quote
Agent Scan reads local agent configurations for Claude, Cursor, GitHub Copilot, Windsurf, Gemini CLI, Amp, Amazon Q and others, connects to the MCP servers it finds to fetch tool descriptions, and reports scored risk indicators such as prompt injection, exposure to untrusted or private data, destructive capabilities and malicious skills. The original MCP-Scan added tool pinning, which hashes tool descriptions so a silent change after approval (a rug pull) shows up.
There is an important caveat in Snyk's own README: scanning an MCP configuration can execute the stdio commands in it and make outbound requests, so run scans in a sandbox when the configs are untrusted. Interactive scans ask for consent per server by default. A background mode, deployed through MDM, reports results from many machines into a Snyk Evo instance for central review.
Strengths
Free scanner with deep MCP attack research behind it
API gateway plugins for MCP OAuth, tool-level ACLs and token exchange
8.4
Kong treats MCP as another kind of API traffic. The AI MCP OAuth2 plugin validates that tokens were issued for the target MCP server and, by default, strips them before forwarding, MCP Tool ACLs (January 2026) filter which tools each consumer sees, and Kong AI Gateway 3.14 (April 2026) added RFC 8693 token exchange and scope-based tool filtering.
Best for
Platform teams that already run Kong and want MCP traffic under the same policies as APIs
Headquarters
San Francisco, California
Founded
2017
Ownership
Private
Key products
Kong AI Gateway, AI MCP OAuth2 plugin, AI MCP Proxy plugin, MCP Tool ACLs, Kong Konnect
Pricing model
MCP OAuth2 plugin requires AI Gateway Enterprise; quote only
The OAuth plugin is aimed squarely at the token passthrough and confused deputy problems the MCP specification warns about. Tokens are checked by introspection or, since 3.14, locally against the authorization server's JWKs, and can be mapped to Kong consumers for rate limits and ACLs. Tool ACLs run default-deny and rewrite the tools/list response, so a client never sees a tool it is not allowed to use. With scope-based filtering, an agent carrying a read scope sees read-only tools, with no consumer group management needed.
Kong also publishes recipes for fronting third-party servers such as GitHub's MCP server, including a token swap pattern that keeps a stored service credential at the gateway. Version 3.13 changed the OAuth plugin to treat all incoming traffic as MCP to close a possible authentication bypass, which is a reminder to read release notes closely. The trade-off is that Kong secures access and traffic; it does not scan tool descriptions for poisoning.
MCP authorization server and identity gateway that exchanges credentials for agents
8.2
Aembit announced IAM for Agentic AI in October 2025 and made it generally available in April 2026. It adds a managed MCP Authorization Server that implements the OAuth 2.1 flow in the MCP specification, and an MCP Identity Gateway that validates tokens, enforces policy on every MCP request and exchanges credentials so the agent never holds them.
Best for
Identity and security teams that want per-user, per-agent access policy on MCP servers
Headquarters
Silver Spring, Maryland
Founded
2021
Ownership
Private
Key products
Aembit IAM for Agentic AI (MCP Authorization Server, MCP Identity Gateway, Blended Identity), Aembit Workload IAM Platform
Pricing model
Quote only
Blended Identity is the idea that sets Aembit apart. Each policy decision looks at the agent and the human operating it together, so a rule can say engineers may use Claude Desktop to reach Jira while only the security team may use it to reach the vulnerability scanner. When the gateway connects to a downstream MCP server for two users, each gets credentials scoped to their own identity, and one user's access can be revoked without rotating shared secrets.
The gateway is stateless and evaluates policy on every request. Today it deploys as a Linux virtual machine in your environment, with containerized and fully managed options on the roadmap, and it integrates with Okta, Microsoft Entra ID and Google through OIDC and SAML. Aembit raised a $25 million Series A in 2024, bringing total funding to about $45 million, so ask about support capacity as you would with any smaller vendor.
Strengths
Agent never sees downstream credentials
Policies combine agent identity and human identity
agentgateway, an open-source Rust proxy for MCP, A2A and LLM traffic
8.1
Solo.io built agentgateway and contributed it to the Linux Foundation in August 2025. It is an open-source proxy designed for agent-to-tool, agent-to-agent and agent-to-LLM traffic, with JWT authentication and role-based access control over MCP servers and tools, per-client tool federation and built-in tracing.
Best for
Kubernetes and platform teams that want an open, self-hosted MCP gateway
Headquarters
Cambridge, Massachusetts
Founded
2017
Ownership
Private
Key products
agentgateway (open source, Apache 2.0), Solo Enterprise for agentgateway, kagent
Pricing model
agentgateway free and open source; Solo Enterprise support by quote
MCP sessions are long-lived and stateful, and Solo's argument is that older API gateways were not built for that. agentgateway is written in Rust, multiplexes many backend MCP servers behind one endpoint, virtualizes the tool list per client so each session only sees what it may call, and can expose a REST API as MCP tools from an OpenAPI spec. Contributors listed at the Linux Foundation launch included AWS, Cisco, IBM, Microsoft and Red Hat.
Solo's documentation says the project was accepted into the Agentic AI Foundation in 2026, and Solo sells an enterprise distribution with support. It is infrastructure first: you get policy, audit trails and rate limits, while scanning for poisoned tools or prompt injection needs to come from a guardrail you plug in.
Prisma AIRS scanning of MCP tool calls, outputs and descriptions
8.0
Prisma AIRS checks MCP traffic in two ways. A standalone remote Prisma AIRS MCP server, released in September 2025, scans tool inputs, outputs and tool descriptions for prompt injection, context poisoning and exposed credentials, and MCP threat detection in the AI Runtime scan APIs flags tool description tampering and credential leakage.
Best for
Prisma AIRS and Strata customers adding MCP inspection to existing AI runtime security
The MCP server approach suits low-code platforms where you cannot insert a proxy: you add the Prisma AIRS MCP server to the client with an API key or OAuth token and a security profile, and every external tool call can be checked against prompt injection detection, DLP, malicious code detection, URL categories and custom topic guardrails. Results land in the same logs as other Prisma AIRS scans. Prisma AIRS 3.0 also scans agent artifacts, which covers MCP server configurations as part of agent posture.
Palo Alto's documentation is candid about the main limit: tool descriptions do not guarantee that an agent's model will call the security server, so you need system prompts or integration points that force the check. For hard enforcement, pair it with a gateway. It ranks lower here because it is strongest as one module of a much larger platform rather than a focused MCP control.
Strengths
Scans tool descriptions, inputs and outputs with the full AIRS detection set
Descope's Agentic Identity Hub, launched in April 2025 and updated to version 3.0 in October 2026, is an identity provider for AI agents and MCP servers. It adds OAuth 2.1 with PKCE, Dynamic Client Registration and Client ID Metadata Documents, per-agent and per-tool scopes, a credential vault for downstream connections, and audit logs that can stream to a SIEM.
Best for
Teams publishing their own MCP servers to customers or partners
Self-serve plans and enterprise quotes; check current plans with the vendor
Descope helps the side of MCP security that gets less attention: securing the MCP servers you publish. The January 2026 release covered MCP auth for internal and external servers, including tenant isolation for B2B use. The June 2026 release added step-up approval for sensitive agent actions through CIBA, support for autonomous agents without a delegating user, and a standalone MCP auth option that leaves your existing login system in place.
Descope was founded in 2022 by the team behind Demisto and raised a $53 million seed round in 2023. It fits developers who need spec-compliant MCP authorization without becoming OAuth experts. It does not inspect tool content, so treat it as the identity layer and add a gateway or scanner for runtime threats.
Strengths
Spec-compliant MCP auth without custom OAuth code
Per-tool scopes and step-up approval
Works alongside an existing login system
Watch out for
Identity layer only; no threat inspection
Best for MCP servers you own, less for third-party ones
MCP governance in Azure API Management and API Center, plus Windows containment
7.6
Azure API Management can expose any managed REST API as an MCP server or front an existing one, then apply policies such as Entra ID token validation, rate limits and IP filtering, while Azure API Center serves as a private MCP server registry. On Windows, MCP servers registered with the on-device agent registry run contained in a separate agent session and user account.
Best for
Azure-centric organizations exposing internal APIs and MCP servers to agents
Headquarters
Redmond, Washington
Founded
1975
Ownership
Public (NASDAQ: MSFT)
Key products
Azure API Management MCP servers, Azure API Center MCP registry, Windows on-device agent registry
Pricing model
Azure API Management tier pricing; Windows features in preview
The Azure path is attractive if your APIs already live in API Management. Credential manager injects OAuth tokens for outbound calls so tools reach backends without the agent handling secrets, monitoring flows into Azure Monitor and Application Insights, and Microsoft's own Azure MCP Server guidance recommends putting remote servers behind API Management as the enforcement gateway and keeping an inventory in API Center to spot shadow servers.
There are gaps to plan for. API Management supports MCP tools but not MCP resources or prompts, its MCP features are not available in workspaces, and policies apply to all tools on a server rather than per tool. The Windows containment model is still pre-release, and once a user grants file access to a host app, every MCP server used by that host gets it.
Strengths
Reuses API Management policies and Entra ID
Private MCP registry in API Center
OS-level containment for packaged Windows MCP servers
Watch out for
No support for MCP resources or prompts in API Management
Inventory the MCP servers you already have Before buying enforcement, find out which MCP servers developers and agents use, which are local stdio processes and which are remote, and who approved them. Scanners such as Snyk Agent Scan and discovery from SentinelOne answer that question quickly.
Decide where the choke point lives A network or Zero Trust portal suits remote SaaS servers, an API gateway suits internal APIs exposed as MCP, and an endpoint agent suits local servers on laptops. Most companies need two of these, so check how each product handles the type you cannot route through it.
Demand spec-level OAuth handling Ask whether the product validates token audience, refuses token passthrough, supports scoped tokens per tool and handles servers without Dynamic Client Registration. These are the failure modes the MCP specification calls out by name.
Check tool description and output inspection Access control does not catch a poisoned tool description or a prompt injection hidden in tool output. If the gateway does not inspect content, pair it with a scanner or guardrail that does.
Test with real attack cases Run a tool poisoning sample, a rug pull (a description change after approval) and an injection in a tool response against each shortlisted MCP tool security software product, and confirm what it blocks versus what it only logs.
Confirm logging reaches your SOC You will need per-call records of user, agent, server, tool and decision for incident response. Check export formats, SIEM connectors and whether logs include arguments or only metadata.
Pricing
How MCP security pricing works
MCP security pricing follows the product it is attached to. Network and SASE products such as Cloudflare include MCP portals in their Zero Trust plans, with higher tiers needed for features like log export. API gateways such as Kong and Azure API Management put MCP features in their existing enterprise or tier-based pricing. Security platforms such as SentinelOne, Palo Alto Networks and Snowflake's Natoma sell by quote, usually per user, per protected agent or as part of a wider contract.
There are real free options for the first step. Lasso's MCP Gateway, Docker's MCP Gateway and agentgateway are open source, and Snyk Agent Scan is free to run. Free tools are good for discovery and for a pilot on a few servers, but central policy, fleet-wide reporting and support nearly always need a paid plan. Ask every vendor how pricing changes as the number of MCP servers, agents and tool calls grows, since agent traffic can rise far faster than headcount.
MCP security risks and issues buyers keep running into
The MCP specification's own security best practices page lists the attacks vendors build against: confused deputy attacks through MCP proxy servers, token passthrough, server-side request forgery during OAuth metadata discovery, session and state handle hijacking, compromise of local MCP servers through malicious startup commands, and over-broad scopes. OWASP has since published an MCP Top 10 that adds tool poisoning, shadow MCP servers, supply chain tampering, command injection and missing audit trails.
In practice, three MCP security issues come up most in buyer conversations. Tool poisoning hides instructions in a tool description that the model reads but a human rarely does. Rug pulls change a description after a user has approved the server. Tool shadowing lets one server's description influence how the agent uses another, trusted server. Invariant Labs, now part of Snyk, named and documented several of these attacks in 2025.
Specific server vulnerabilities matter too. Searches for chrome-devtools-mcp security vulnerabilities CVE entries show that teams check individual servers, and the right habit is to look up CVE and advisory records for every server before approval, pin the approved version and rescan when it changes. Treat any MCP server that can launch a browser, run shell commands or touch credentials as high risk by default.
MCP gateway, scanner or registry: which layer to buy first
If agents are already calling remote MCP servers in production, start with a gateway or portal, because it gives you one place to authenticate users, limit tools and log calls. Cloudflare, Kong, agentgateway, Docker's gateway and the Prompt Security gateway from SentinelOne all fit here, and they differ mainly in where they run: at the network edge, in an API gateway, in Kubernetes, in containers on a laptop or on the endpoint.
If MCP use is mostly developers wiring servers into Cursor, Claude Code or VS Code, start with discovery and scanning. A scanner shows which servers are installed, what their tools do and whether descriptions contain hidden instructions. A curated registry, such as the Docker MCP Catalog or a private registry in Azure API Center, then gives developers an approved list so they stop pulling servers from random repositories.
Identity products such as Aembit and Descope come next for most teams, once you need per-user policy and want agents to stop holding long-lived tokens. They complement a gateway rather than replace it, and several gateways on this list can hand off to them.
Securing MCP servers you build versus servers you consume
Securing MCP servers you publish is mostly an authorization job. The specification expects OAuth 2.1 for remote servers, requires that servers only accept tokens issued for them, forbids passing client tokens through to downstream APIs, and recommends a progressive, least-privilege scope model in which servers ask for more scope only when a privileged tool is first used. Descope, Aembit and Kong's MCP OAuth plugin all exist to make that easier, and MCP OAuth 2.1 security is where most homegrown servers fall short.
For third-party servers you consume, the job flips to supply chain and runtime control. Prefer signed or pinned servers, run local servers in containers or a sandbox with no host secrets, restrict network egress, and route remote servers through a gateway that logs every call. A secure MCP tunnel, meaning an authenticated private connection to a server instead of a public endpoint, helps for internal servers, but it does not inspect what the tools do.
Both sides need the same evidence: which identity called which tool with what arguments, and what came back. Decide early whether logs may contain arguments, since tool calls often carry customer data.
How to test MCP integrations before rollout
Many security teams now ask vendors for security testing of MCP integrations as part of procurement, and you can run the same tests yourself. Build a small lab with one benign server, one server with a poisoned tool description, one that changes its description after approval, and one that returns a prompt injection in its tool output. Connect each shortlisted product and record whether it blocks, warns or misses each case.
Then test the boring failures. Point a client at a server through its direct URL and see if the gateway policy still applies. Use a token issued for a different audience. Disconnect the gateway and see if the agent fails closed or quietly connects direct. Many of the gaps we saw in vendor documentation are of this kind, such as Cloudflare noting that blocked users can still use a server's direct URL unless Access is its OAuth provider.
Scanners deserve their own safety check. Snyk warns that scanning an MCP configuration can execute the commands inside it, so run discovery scans of untrusted configurations inside a disposable container or VM.
MCP security for coding agents
Coding agents are where most companies first meet MCP at scale. Claude Code, Cursor, Copilot and Codex load MCP servers from configuration files in the repository or the user's home directory, often with the developer's full permissions. A malicious server there can read files, run commands and leak secrets without passing any network gateway.
That is why MCP controls for developers usually combine an endpoint or hook-based product with a scanner. Our ranking of AI coding agent security solutions groups those products by where they enforce, from IDE hooks to sandboxes and CI, and our AI coding agent security platforms ranking covers the broader platforms that discover coding agents and their MCP servers on laptops.
MCP security best practices, briefly
We keep this short here because our full MCP security best practices guide covers the checklist in depth. The core list is stable: keep an inventory of every MCP server and who approved it, allow only servers from a vetted registry, pin versions and alert on description changes, run local servers sandboxed with no host secrets, use OAuth with audience checks and scoped, short-lived tokens, never pass client tokens through, put remote servers behind a logged gateway, and require human approval for destructive tools.
None of the tools above does all of this alone. Pick the products that cover the steps you cannot do with configuration and policy, and use the rest of the list as the test plan for your proof of concept.
Where MCP security fits with agentic AI security and shadow AI
MCP is one connection type among several. Agents also call APIs directly, browse the web and act through SaaS integrations, so a broad agentic AI security solution covers inventory, identity and runtime controls beyond MCP. If your main worry is unknown agents and servers appearing across the company, our shadow AI tools ranking covers discovery from the SaaS and browser side.
Several vendors here also appear on those lists, which is useful when you want fewer contracts. SentinelOne's Prompt Security is one example; see our Prompt Security review for a closer look. If you would rather have someone else do the first cut, you can get a free shortlist matched to your stack.
FAQ
MCP security tools: FAQ
Answers to the questions buyers ask most about mcp security tools.
MCP security covers the controls around the Model Context Protocol connections that let AI agents call tools: which servers are allowed, how agents and users authenticate to them, which tools each identity may call, how tool descriptions and outputs are checked for attacks, and how every call is logged.
Is MCP secure?
The protocol defines an OAuth-based authorization model and documents known attacks, but security depends on how servers and clients are built and configured. Many servers in public repositories have weak or no authentication, and any server can carry poisoned tool descriptions. Treat MCP servers like third-party code with access to your data.
What are the main MCP security risks?
The most cited MCP security risks are tool poisoning, rug pulls, tool shadowing, prompt injection through tool output, token passthrough and confused deputy attacks, over-broad scopes, compromised local servers that run arbitrary commands, and shadow MCP servers nobody approved.
How do you secure MCP servers?
For servers you build, implement OAuth 2.1, validate token audience, never pass client tokens downstream, use narrow scopes and log every call. For servers you use, install only vetted and pinned versions, run local servers sandboxed, restrict network egress and route remote servers through a gateway.
What is an MCP gateway?
An MCP gateway is a proxy between agents and MCP servers. It authenticates callers, decides which servers and tools each one may use, can inspect tool calls and responses, and records an audit trail. Examples include Cloudflare MCP server portals, Kong AI Gateway, agentgateway, Docker MCP Gateway and the Prompt Security MCP gateway.
What are the best MCP security tools?
For a governed endpoint, Cloudflare. For shadow MCP discovery on endpoints, SentinelOne's Prompt Security. For a signed catalog and container isolation, Docker. For free scanning, Snyk Agent Scan. For identity and token exchange, Aembit or Descope. Most teams combine a gateway with a scanner.
What is an MCP security scanner?
An MCP security scanner reads MCP configurations, connects to the servers, pulls their tool descriptions and flags hidden instructions, prompt injection, risky permissions and description changes. Snyk Agent Scan and the scanner in Lasso's MCP Gateway are two examples. Run scans of untrusted configs in a sandbox.
How does MCP OAuth 2.1 security work?
Remote MCP servers act as OAuth resource servers. Clients obtain tokens through an authorization server using OAuth 2.1 with PKCE, and the server must only accept tokens issued for it. The specification forbids token passthrough and recommends least-privilege scopes that are raised step by step when privileged tools are used.
What is the difference between MCP security and AI agent security?
AI agent security covers everything an agent does: its identity, inventory, permissions, runtime behavior and all of its connections. MCP security is the part that governs agent-to-tool connections over the Model Context Protocol. Many agent security platforms include MCP controls, while focused MCP tools go deeper on that one path.
How much do MCP security solutions cost?
Open-source gateways such as Lasso's, Docker's and agentgateway are free, and Snyk Agent Scan is free to run. Commercial MCP tool security solutions are priced per user, per agent or as part of a platform contract, and most enterprise products are quote only. Cloudflare and Azure fold MCP features into existing plans.
Related
More AI security rankings
Related lists and our overall ranking of AI cybersecurity companies.