Runlayer
Enterprise AI control planeAn enterprise control plane covering MCP gateway, shadow AI discovery, runtime security and agent enablement.
Runlayer sits between the AI clients employees already use (Claude, Cursor, ChatGPT, Codex and agents) and company systems. One gateway carries MCP calls with identity, policy and audit attached; AI Watch discovers unmanaged MCPs, skills and client configs through your MDM; runtime scanners inspect tool calls for intent drift, PII and exfiltration; and an internal catalog lets teams publish approved connectors, skills and agents.
Best for: Companies rolling AI agents out across engineering, security, IT and business teams that want discovery, the gateway, runtime security and enablement in one product instead of several.
What it covers
- MCP gateway across 20+ AI clients, with permissions, authentication and allowed tools defined up front
- Shadow AI discovery of unmanaged MCPs, skills, plugins and client configs through existing MDM, without employee opt-in
- Access scoped by user, group, role, agent, AI client, connector, tool, OAuth state and runtime context
- Agent accounts: each agent can have its own identity, policies, tokens and audit trail
- Runtime scanners for intent drift, PII and data exfiltration, starting in alert-only mode and moving to blocking
- Human approval for risky actions and a full audit trail with user and agent attribution
- Runlayer-hosted single-tenant deployment, or self-hosted in your AWS account or EKS/GKE cluster
- SOC 2, HIPAA and GDPR compliance stated by the vendor
What it doesn't, or doesn't say
- Commercial pricing isn't published; the docs say to ask the Runlayer account team.
- Public pages don't describe mapping evidence to AI frameworks such as the EU AI Act, NIST AI RMF or ISO 42001; ask for this if your audit needs it.
- Self-hosting is documented for AWS (ECS) and EKS or GKE; other clouds and air-gapped installs aren't described.
- On Linux, AI Watch discovery is detect-only; endpoint enforcement modes and sessions are macOS and Windows.
- It governs agent and tool use; it is not a model registry or a vendor-risk questionnaire tool.
On the seven criteria
- Coverage of agents and MCPDocumentedOne gateway for every MCP and LLM call, across Claude, Cursor, ChatGPT, Codex and internal agents. Source: runlayer.com
- Identity and permissionsDocumentedScopes access by user, group, role, agent, AI client, connector, tool, OAuth state and runtime context. Source: Runlayer for IT & Security
- Audit logsDocumentedFull audit trail with user and agent attribution, plus session traces. Source: runlayer.com
- Policy enforcementDocumentedPolicies for users, agents and tool calls; human approval; runtime scanners from alert-only to blocking. Source: runlayer.com
- Shadow AI discoveryDocumentedAI Watch finds unmanaged MCPs, skills, plugins and client configs via MDM; Linux is detect-only. Source: Runlayer docs: Shadow AI
- Compliance mappingPartialVendor states SOC 2, HIPAA and GDPR compliance; AI framework mapping isn't described publicly. Source: runlayer.com/llms.txt
- DeploymentDocumentedSingle-tenant hosted, or self-hosted on AWS ECS, EKS or GKE. Source: Runlayer deployment overview
- PricingNot published. Ask Runlayer's account team. Source: Runlayer deployment overview
Noma Security
Agent security and governanceA security platform for agents across endpoint, SaaS and cloud: discovery, access control, runtime detection and red teaming.
Noma covers agents employees use on their devices, agents running in SaaS platforms and agents engineers build in the cloud. Its products are AI-SPM (discover agents, MCPs, skills and tools), AI-DR (runtime behavior), Agent Access Control and AI Red Teaming. Policy is enforced through control points you already run, such as AI gateways, MCP gateways, agent hooks and SDKs.
Best for: Security teams that need one view of agent risk across endpoint, SaaS and cloud agents, with compliance evidence mapped to AI frameworks.
What it covers
- Discovery of agents, MCP servers, skills and tools (AI-SPM)
- Runtime controls that can stop actions, steer the agent, alert a human, mask data or trigger a killswitch
- Approve, review or block agents, MCP servers and skills; scope actions by user, team or tool sensitivity
- Findings mapped to NIST AI RMF, EU AI Act, ISO 42001, OWASP LLM Top 10 and MITRE ATLAS
- Runtime audit trails covering prompts, tool calls, detections and enforcement actions
- AI red teaming of agents and AI applications
- On-premises and SaaS deployment
What it doesn't, or doesn't say
- Enforcement is designed to plug into gateways, hooks and SDKs you already run, so you may still need a gateway for the request path.
- Public pages don't describe an employee-facing catalog for publishing approved connectors and agents; it is security-first rather than an enablement layer.
- Pricing isn't published; it is sold through demos.
On the seven criteria
- Coverage of agents and MCPDocumentedEndpoint, SaaS and cloud agents, plus MCP servers, skills and tools. Source: noma.security
- Identity and permissionsDocumentedAgent Access Control: approve, review or block; scope by user, team or tool sensitivity. Source: noma.security
- Audit logsDocumentedRuntime audit trails of prompts, tool calls, detections and enforcement actions. Source: Noma: AI Compliance & Regulation
- Policy enforcementDocumentedRuntime actions: stop, steer, alert, mask, killswitch; enforced through your existing control points. Source: noma.security
- Shadow AI discoveryDocumentedAI-SPM and AI Asset Discovery find agents, MCPs, skills and tools across the organization. Source: noma.security
- Compliance mappingDocumentedMaps findings to NIST AI RMF, EU AI Act, ISO 42001, OWASP LLM Top 10 and MITRE ATLAS. Source: Noma: AI Compliance & Regulation
- DeploymentDocumentedOn-premises and SaaS deployment options. Source: Noma: AI Compliance & Regulation
- PricingNot published. Request a demo from Noma. Source: noma.security
Arcade.dev
Agent actions runtimeAn actions runtime for agents: authenticated, policy-checked tool calls to business systems, with a large tool catalog.
Arcade sits between the agents you build and the systems they act on. Authentication runs against your identity provider and authorization is delegated, so an agent acts as its user and never beyond its scope. It ships a catalog of pre-built MCP tools, a framework for building your own, and pre- and post-tool-call hooks for custom policy.
Best for: Engineering teams building their own agents who need delegated auth, credential handling and reliable tool execution, including on-prem or air-gapped.
What it covers
- Delegated authorization through your OAuth and identity provider flows, with no broad service accounts
- Catalog of pre-built MCP tools (Google Workspace, Slack, Microsoft, Salesforce and more) and an open-source framework for custom tools
- Pre- and post-tool-call hooks to block actions, redact PII or stop data leaving a system
- SSO, RBAC and audit logs; SOC 2 compliant per the vendor
- Cloud, on-premises, air-gapped or hybrid deployment
- Framework-agnostic: LangChain, LlamaIndex, Mastra, CrewAI, OpenAI Agents and others
What it doesn't, or doesn't say
- Shadow AI discovery isn't described; it governs the agents and tools routed through it.
- AI framework mapping (EU AI Act, NIST AI RMF, ISO 42001) isn't described.
- Its center of gravity is developers building agents; company-wide governance of employees' own AI clients is less of a focus on its public pages.
On the seven criteria
MintMCP
MCP and LLM gatewayNow Mintive
An MCP gateway and LLM gateway with role-based tool bundles, guardrails and agent monitoring. Now named Mintive.
MintMCP, now Mintive, governs both tool calls and model calls. Admins curate approved MCP servers with pre-configured credentials and group them into role bundles; the LLM gateway adds per-team budgets and allowed models; Mint Guard screens tool calls for prompt injection and sensitive data; Agent Monitor traces activity in claude.ai, Claude Code, Cursor, Copilot and Codex.
Best for: Teams that want a hosted MCP gateway with ready-made role bundles, plus model spend controls, and are comfortable with a managed cloud service.
What it covers
- MCP gateway with an MCP store, role-based bundles and the option to switch off destructive tools
- LLM gateway with per-team budgets and allowed models
- Agent identities with delegated, least-privilege credentials revocable on their own
- Mint Guard screening for prompt injection, exposed secrets and personal data, from monitoring to enforcing
- Audit logs exportable to a SIEM or streamed over OTLP
- SOC 2 Type II audited, BAAs signed, SSO via SAML/OIDC, EU and US data residency (per the vendor)
What it doesn't, or doesn't say
- Managed cloud by default; on-premises or self-hosted is by request, with no published detail.
- Device-wide shadow AI discovery isn't described; Agent Monitor observes specific supported clients.
- AI framework mapping (EU AI Act, NIST AI RMF, ISO 42001) isn't described.
- Pricing is on request.
On the seven criteria
MCP Manager
MCP gatewayA hosted MCP gateway by Usercentrics with a private registry, PII filtering, RBAC, OAuth enforcement and audit logs.
MCP Manager puts one governed gateway between AI clients and MCP servers. It brokers identity per server, applies gateway rules to each request and response, filters PII and PHI, alerts on changed tool descriptions and logs every call. It runs only as a hosted service on Google Cloud in the United States.
Best for: Teams, often in regulated industries, that want a quick-to-start hosted MCP gateway focused on PII filtering and audit, and don't need self-hosting.
What it covers
- Private MCP registry and multiple gateways per team, server or use case
- OAuth enforcement, RBAC and agent identity management; SSO and SCIM on Enterprise
- PII filtering (Microsoft Presidio) and a custom rules engine for DLP on Enterprise
- Alerts when tool descriptions change; tool and server pinning against rug pulls
- Audit logs, with OpenTelemetry export on Enterprise
- Remote, managed and workstation MCP servers, including servers inside your network via an outbound tunnel
What it doesn't, or doesn't say
- Hosted only: no on-premises, customer-deployed or air-gapped build, and none planned, per its docs.
- Data is stored in the US (GCP us-east1); EU data residency isn't available today.
- Covers MCP traffic; model calls and non-MCP agent actions aren't in scope.
- Shadow MCP is handled by routing usage through the gateway with your MDM/EDR, not by its own discovery scanner.
On the seven criteria
Credo AI
AI governance, risk and complianceA governance, risk and compliance platform for AI: registry, risk assessment, policy packs and audit evidence.
Credo AI is a system of record for AI governance. It registers AI systems, agents, models and vendors, assesses risk against an agentic risk and control library, and turns regulations into policy packs with evidence for audits. It integrates with tools like Snowflake, Databricks, ServiceNow and Jira rather than sitting in the tool-call path.
Best for: Risk, compliance and legal teams that need a registry of AI use and framework-ready evidence, often alongside a runtime control plane.
What it covers
- AI registry for use cases, models, agents and vendors, with shadow AI detection
- Policy packs for the EU AI Act, NIST AI RMF, ISO 42001, OMB M-25, Colorado ADMT and NAIC AI
- Risk scoring, an agentic risk and control library, and audit trails and reporting
- Integrations with Snowflake, Databricks, AWS, Azure, ServiceNow, Jira, Slack, GitHub and MLflow
What it doesn't, or doesn't say
- Not an inline gateway: public pages don't describe enforcing policy on each MCP or tool call at runtime. Its Agent Governor is listed as a research preview.
- Runtime identity and credential handling for agents isn't described.
- Deployment options aren't stated on the public site; check its trust center or ask.
On the seven criteria