Skip to main content
Mintive is the new name for MintMCP.Read the announcement
Jiquan Ngiam

10 Best AI, MCP, and Agent Gateways for Self-Hosted Deployments 2026

Self-hosted deployment is an important option for regulated AI workloads that require greater infrastructure control, network isolation, or data sovereignty. A 2024 Tray.ai survey found that more than 86% of enterprises required tech stack upgrades to properly deploy AI agents, making infrastructure and governance important evaluation criteria.

The right MCP gateway should provide centralized authentication, real-time monitoring, and compliance controls without forcing you to send sensitive data through third-party clouds. For enterprises handling sensitive data, air-gapped, VPC-based, and self-hosted installations provide greater control over network access and data flows.

This guide evaluates ten agent gateways across performance, open-source licensing, self-hosted capability, and MCP protocol support to help you identify the right fit for your deployment constraints.

Key takeaways​

  • Mintive (formerly MintMCP) governs tool calls through its MCP Gateway and model calls through its LLM Gateway, while Agent Gateway provides first-class agent identities with scoped permissions and independent audit trails
  • Self-hosted gateways keep the gateway layer within your infrastructure while maintaining centralized authentication, authorization, and observability, but traffic may still reach external model or tool providers
  • Open-source licensing (Apache 2.0 and MIT) dominates the self-hosted category, with varying feature availability between community and enterprise editions
  • Performance characteristics vary significantly by product, configuration, enabled policies, and workload, so teams should compare gateways using equivalent benchmarks
  • Authentication approaches range from basic shared credentials to full OAuth 2.1 brokering with per-agent identity management

1. Mintive Gateway: enterprise MCP infrastructure with Agent Gateway capabilities​

Mintive Gateway provides a platform for running enterprise agents, governing tool calls through the MCP Gateway and model calls through the LLM Gateway. Mint Guard screens gateway tool calls for prompt injection, PII, and secrets, while Agent Monitor covers Claude Enterprise prompts and coding-agent activity, including commands, file access, and MCP tool calls. For organizations requiring infrastructure control, Mintive offers VPC and self-hosted deployment options on request.

What makes Mintive Gateway different​

Mintive bridges two connected categories: MCP Gateway for governed data and tool connections to AI systems like ChatGPT, Claude, Microsoft 365 Copilot, Gemini, and Cursor, and Agent Gateway for identities, permissions, credentials, and monitoring of autonomous agents that work alongside users.

The same 2024 survey found that 42% of enterprises required eight or more data-source connections for AI agent deployment. The architecture wraps stdio, hosted, HTTP-streamable, and SSE MCP servers behind SSO-fronted remote MCP endpoints with OAuth brokering, SCIM-driven membership, and rule-based policy.

Core capabilities​

MCP Gateway Foundation:

  • Virtual MCP Bundles: Create team-specific, per-use-case endpoints that expose only required tools with SCIM-driven membership and role-based access
  • OAuth Brokering: Add enterprise authentication to local and hosted MCP servers, including OAuth 2.x, bearer tokens, and SSO-fronted access
  • Hosted MCP Connectors: Mintive runs connector instances with auto-scaling and sandboxed execution per connector
  • Gateway Middleware: Runs customer-authored logic in a JS sandbox with external DLP and guardrails integrations

Agent Gateway Layer:

  • Agent Bundles: Give autonomous agents first-class identities with M2M auth, scoped tools, independent rotation and revocation, and an "act as agent" flow for connectors requiring per-agent OAuth
  • Two-Layer Monitoring: MCP Gateway tracks tool calls routed through governed connections, while Agent Monitor extends visibility to local activity through Claude Code and Cursor hooks, covering prompts, commands, and file access where supported

Security and compliance​

Mintive is SOC 2 Type II audited with continuous compliance monitoring. Enterprise SSO, complete audit trails, PII detection, and role-based access control are built into every layer. Mintive is compliant with HIPAA standards. Customers handling protected health information can request supporting documentation, and Mintive signs BAAs.

Enterprise integrations​

Deployment options​

  • Managed SaaS-first delivery with US and EU availability
  • VPC and self-hosted options available on request

Pricing​

Contact for enterprise demonstration and pricing

2. Bifrost (Maxim AI)​

Bifrost is an open-source AI gateway built for high-throughput, low-latency workloads. The gateway uses Go for its runtime architecture and documents 11 microsecond latency overhead at 5,000 requests per second in published benchmarks.

Bifrost's primary focus​

Bifrost targets engineering teams that prioritize raw performance and require Apache 2.0 licensing for self-hosted deployment. The gateway supports 25+ LLM providers and includes native MCP gateway capabilities with Code Mode for reducing agent token costs.

Core capabilities​

  • Apache 2.0 open-source licensing
  • Documented 11 microsecond latency overhead at sustained load
  • Air-gapped deployment support for disconnected environments
  • Code Mode for reducing agent token costs at scale
  • Hierarchical budgets at virtual key, team, and customer levels
  • Semantic caching with dual-layer hash and embedding support

Where Bifrost fits​

Performance-critical deployments where latency overhead matters at scale. Organizations requiring full Apache 2.0 licensing and code audit capability. Air-gapped government or defense environments.

Considerations for enterprise teams​

Teams evaluating Bifrost should assess whether they need an operational governance layer around SCIM-driven RBAC, per-use-case tool bundles with membership sync, hosted connector management with auto-scaling, and per-agent identity with M2M auth that purpose-built enterprise gateways provide.

License: Apache 2.0 | Deployment: Self-hosted, air-gapped capable

3. Obot Platform​

Obot Platform provides open-source MCP gateway capabilities as part of a broader AI agent orchestration framework from Acorn Labs. The platform is built for organizations seeking full infrastructure ownership with MIT licensing.

Obot's primary focus​

Obot targets Kubernetes-fluent platform engineering teams who want to operate their own MCP gateway infrastructure. The platform includes a curated MCP catalog with self-service discovery and Virtual MCP endpoints for combining multiple MCP servers into single logical endpoints.

Core capabilities​

  • MIT-licensed open-source deployment with separately available hosted MCP platform
  • Curated MCP catalog with self-service discovery
  • Virtual MCP endpoints combining multiple underlying MCP servers
  • Self-hosted deployment using Docker or Kubernetes
  • Agent workflow orchestration beyond basic MCP routing

Where Obot fits​

Platform engineering teams with DevOps expertise who require full infrastructure ownership. Organizations building custom AI agent platforms. Teams prioritizing open-source tools for transparency.

Considerations for enterprise teams​

An OSS-first, self-hosted model requires the customer to operate the runtime, Kubernetes deployment, scaling, connector lifecycle, and governance stack. Teams should evaluate whether they want self-hosted orchestration or a managed SaaS-first gateway with hosted MCP connectors and pre-configured governance controls.

License: MIT | Deployment: Self-hosted (Docker, Kubernetes)

4. TrueFoundry​

TrueFoundry provides a Kubernetes-native AI platform with unified LLM, MCP, and A2A gateway capabilities in one control plane.

TrueFoundry's primary focus​

TrueFoundry targets platform engineering and ML platform teams who want to consolidate gateway, model serving (vLLM, SGLang, Triton), and agent orchestration in a single platform with physical namespace isolation for multi-tenancy.

Core capabilities​

  • Unified LLM, MCP, and A2A gateway in one control plane
  • Kubernetes-native deployment with namespace isolation
  • 1,600+ models across multiple providers
  • Full self-hosted control plane in customer Kubernetes
  • Air-gapped deployment via forward proxy
  • Latency-based routing with automatic fallback

Where TrueFoundry fits​

Organizations wanting single-vendor consolidation for MLOps and gateway infrastructure. Teams already invested in Kubernetes for AI workloads. Enterprises requiring full data sovereignty with self-hosted control plane.

Considerations for enterprise teams​

Teams should evaluate whether a full platform approach fits their needs, or whether a focused MCP gateway with hosted connector operations, SCIM-driven Virtual MCP Bundles, and Agent Bundles provides faster time to production governance.

Deployment: Self-hosted control plane | Compliance: TrueFoundry documents SOC 2 Type II, HIPAA, and GDPR coverage for managed infrastructure; confirm applicability for self-hosted deployments

5. Agent Router (formerly Envoy AI Gateway)​

Agent Router is a Kubernetes-native gateway formerly known as Envoy AI Gateway and co-created by Tetrate and Bloomberg. It became Agent Router and joined the Agentic AI Foundation in September 2026.

Agent Router's primary focus​

Agent Router targets platform engineers who want an Envoy-based gateway for model and MCP traffic under vendor-neutral foundation governance. It supports CEL-based policy for MCP tool access.

Core capabilities​

  • Apache 2.0 licensing under Agentic AI Foundation governance
  • Envoy-based data plane
  • CEL-based policy engine for tool-level RBAC
  • Service mesh integration with Envoy and Istio
  • InferencePool for self-hosted model routing
  • Token limits per team, app, or model
  • Filtered tool catalogs per agent

Where Agent Router fits​

Organizations requiring foundation governance for procurement. Teams already using Envoy or Istio service mesh. Platform engineers wanting protocol-level control with neutral governance.

Considerations for enterprise teams​

Agent Router uses external guardrail integrations rather than first-party detection. Teams should evaluate whether they need managed deployment, hosted connector operations, and enterprise governance primitives like SCIM-driven bundles alongside the protocol-level gateway.

License: Apache 2.0 (Agentic AI Foundation) | Deployment: Self-hosted, Kubernetes-native

6. LiteLLM​

LiteLLM is an open-source Python gateway with a large documented provider catalog at 100+ LLM integrations. The project uses MIT licensing for its core and an Enterprise tier for advanced features.

LiteLLM's primary focus​

LiteLLM targets developers who need the fastest path to many providers behind a single OpenAI-compatible endpoint. The gateway includes MCP gateway capabilities with per-key/team tool permissions.

Core capabilities​

  • 100+ LLM provider integrations
  • MIT core license with active contributor base
  • OpenAI-compatible proxy with minimal friction
  • MCP gateway with per-key/team tool permissions
  • Virtual keys with budgets and rate limits
  • Load balancing and automatic fallbacks

Security considerations​

In March 2026, LiteLLM experienced a supply chain attack affecting PyPI versions 1.82.7 through 1.82.8. The UK NHS issued national cyber alert CC-4761 in response. The issue was remediated within three hours, but teams should verify they are running patched versions and consider the security implications for gateway infrastructure.

Where LiteLLM fits​

Development teams prioritizing provider breadth and developer experience. Organizations requiring MIT licensing and code audit capability. Teams willing to accept Python performance characteristics for integration simplicity.

Considerations for enterprise teams​

Teams should benchmark LiteLLM under their expected configuration rather than inferring production throughput from its implementation language. The Enterprise tier adds SSO, SCIM, audit logs, and advanced administration, while core observability and metrics capabilities are available outside Enterprise. Teams should evaluate whether core features meet enterprise governance requirements.

License: MIT core, Enterprise tier for advanced features | Deployment: Self-hosted

7. Kong AI Gateway​

Kong AI Gateway extends the Kong Gateway with AI plugins for multi-LLM routing, semantic caching, and prompt transformation. Organizations already standardized on Kong can add MCP support without deploying separate infrastructure.

Kong's primary focus​

Kong targets organizations with existing Kong Gateway deployments who want to extend established infrastructure with AI capabilities rather than introduce new platforms.

Core capabilities​

  • AI Proxy plugin for multi-LLM routing
  • Semantic caching and prompt transformation
  • MCP Registry in Kong Konnect (MCP support shipped in Gateway 3.12+)
  • AI MCP Proxy plugin with REST-to-MCP conversion and per-tool ACLs (Enterprise tier)
  • Mature plugin ecosystem with OIDC, mTLS, rate limiting
  • Hybrid deployment with Konnect SaaS or fully self-hosted

Where Kong fits​

Enterprises with established Kong Gateway investments seeking unified API and MCP management. Organizations preferring to extend existing infrastructure rather than deploy purpose-built solutions.

Considerations for enterprise teams​

Kong AI Gateway is not a standalone product but AI plugins on Kong Gateway. Advanced AI features including semantic caching and MCP proxy require Enterprise licensing. Teams should evaluate whether a plugin-based approach provides MCP-specific governance primitives like Virtual MCP Bundles, Agent Bundles, and hosted connector runtime.

License: Apache 2.0 core, Enterprise for advanced features | Deployment: Konnect SaaS control plane or fully self-hosted

8. Apache APISIX​

Apache APISIX is an Apache Software Foundation API gateway with AI plugins. All AI plugins, including security-oriented ones, are open source with no feature gating.

APISIX's primary focus​

APISIX targets organizations prioritizing ASF governance and open-source licensing where all features remain available in the community edition.

Core capabilities​

  • All AI plugins open source under Apache 2.0
  • ai-proxy-multi for multi-LLM routing
  • ai-cache plugin with Redis-backed semantic option
  • Token-based rate limiting (ai-rate-limiting)
  • Nginx/OpenResty with etcd for dynamic configuration
  • ASF project governance with vendor neutrality
  • Commercial support available via API7

Where APISIX fits​

Organizations requiring all features in open source without Enterprise gating. Teams prioritizing ASF governance and vendor neutrality. Existing APISIX deployments adding AI capabilities.

Considerations for enterprise teams​

APISIX provides MCP-to-HTTP conversion but may not include MCP-specific primitives like per-use-case tool bundles, agent identity management, or hosted connector runtime. Teams should evaluate whether API gateway extension provides sufficient governance depth for MCP workloads.

License: Apache 2.0 (all features) | Deployment: Self-hosted

9. Portkey​

Portkey is an LLM observability and gateway platform now part of Palo Alto Networks following acquisition in May 2026. The platform documents support for 1,600+ models across 45+ providers.

Portkey's primary focus​

Portkey targets organizations with existing Palo Alto Networks investments who want LLM gateway capabilities integrated into their security platform. The MIT-licensed gateway core remains available for self-hosted deployment.

Core capabilities​

  • 1,600+ models across 45+ providers
  • 50+ built-in guardrails including PII redaction
  • MCP Gateway with OAuth 2.1 support
  • MIT-licensed gateway core
  • Integration with Palo Alto Prisma AIRS
  • Flexible deployment via Docker, Node.js, or Cloudflare Workers

Where Portkey fits​

Organizations with Palo Alto Networks security investments seeking integrated AI gateway. Teams prioritizing broad model catalog coverage. Enterprises wanting managed gateway with built-in guardrails.

Considerations for enterprise teams​

Following the Palo Alto Networks acquisition, teams should evaluate how the product roadmap aligns with their strategic direction and whether the self-hosted option continues to receive active development for MCP-specific governance primitives.

License: MIT core plus commercial | Deployment: Self-hosted option available

10. Dify​

Dify is an open-source LLM application platform with agent capabilities and over 100,000 GitHub stars. The platform provides an all-in-one environment for building RAG pipelines and agent workflows.

Dify's primary focus​

Dify targets teams building complete LLM applications rather than teams focused purely on gateway routing. The platform combines workflow orchestration, RAG pipeline management, and agent development in a single self-hosted environment.

Core capabilities​

  • Dify Open Source License, based on Apache 2.0 with additional conditions, with self-hosting available
  • Visual workflow editor for LLM applications
  • Production RAG pipeline with multiple vector store options
  • Support for 100+ model providers including OpenAI, Claude, and Llama
  • Working dashboard with usage analytics
  • Docker Compose deployment with Postgres, Redis, and vector database

Where Dify fits​

Teams building full LLM applications with agents, not just routing layers. Organizations wanting visual no-code workflow builders. Teams prioritizing self-hosted data control with broad model support.

Considerations for enterprise teams​

Dify operates at the application layer rather than the gateway layer. Teams should evaluate whether they need application orchestration or purpose-built MCP gateway infrastructure with enterprise governance primitives like SCIM-driven access control, per-agent identity, and hosted connector management.

License: Dify Open Source License, based on Apache 2.0 with additional conditions | Deployment: Self-hosted via Docker Compose

Selection criteria for self-hosted agent gateways​

Performance vs. governance trade-offs​

Some performance-focused gateways publish microsecond-overhead benchmarks, but results depend on workload, configuration, and enabled governance controls. Bifrost documents 11 microseconds of gateway overhead at 5,000 RPS, while other gateways prioritize governance features over raw throughput. Agent platforms like Mintive govern tool calls and model calls with shared groups, identities, and audit trails, and Mintive offers self-hosted deployment options on request for teams requiring infrastructure control.

Licensing requirements​

Apache 2.0 and MIT dominate the self-hosted gateway category. Evaluate whether all required features are available in the open-source core or gated behind Enterprise tiers that change the effective licensing model.

STDIO server support​

Many community-built MCP servers use STDIO transport. Evaluate whether your gateway handles STDIO-based servers or only supports remote HTTP/SSE endpoints, which limits ecosystem access.

Authentication architecture​

OAuth 2.1 support was added to the MCP specification in 2025, but implementation varies. Some gateways broker OAuth and wrap stdio servers with enterprise SSO, while others require manual configuration per server. Consider whether you need per-user authentication, per-agent identity, or shared service accounts.

Observability depth​

Without comprehensive logging, organizations face visibility gaps where they cannot track which tools agents use or audit data access. Evaluate whether your gateway provides tool call tracking, latency monitoring, error rates, and cost allocation per team.

Deploy AI agents with full governance​

For organizations requiring both self-hosted deployment options and enterprise-grade governance, Mintive Gateway provides the architecture to meet both requirements. Virtual MCP Bundles create team-specific endpoints with SCIM-driven membership. Agent Bundles give each AI agent its own credential set with independent rotation. Agent Monitor extends visibility to local agent activity in Claude Code and Cursor where supported, covering prompts, commands, and file access.

Mintive governs tools, models, and agents together: Mint Guard screens MCP Gateway tool calls for prompt injection, PII, and secrets, while Rules provide declarative matching and Gateway Middleware runs customer-authored logic in a JS sandbox. The LLM Gateway routes model calls and records usage and cost per call. The same groups drive budgets and allowed-model lists, and teams can switch providers without rewiring each agent.

Start with managed SaaS delivery for fast deployment, then move to VPC or self-hosted when your requirements demand infrastructure control. Visit mintive.com to schedule a demonstration.

Frequently asked questions​

What is an agent gateway and why does self-hosted deployment matter?​

An agent gateway governs autonomous agents through first-class identities, scoped permissions, credentials, and auditability. MCP and LLM gateways govern tool connections and model traffic, respectively. Self-hosted deployment matters for organizations in regulated industries or handling sensitive data where sending AI traffic through third-party clouds creates compliance and security risks. Self-hosted gateways provide greater control over where data flows, but keeping all data within your infrastructure also requires private or self-hosted model and tool backends.

How do self-hosted agent gateways handle authentication?​

Authentication approaches vary significantly. Some gateways broker OAuth and wrap STDIO servers with enterprise SSO, while others require manual configuration per server. Purpose-built MCP gateways may offer OAuth 2.x, SAML, bearer tokens, and SCIM group synchronization with identity providers like Okta and Azure AD. Evaluate whether your gateway supports shared service accounts, per-user authentication, and per-agent identity based on your use cases.

Can self-hosted agent gateways detect shadow AI activity?​

Gateway-only solutions provide visibility into MCP traffic routed through the gateway but cannot detect off-gateway agent activity. Solutions like Mintive extend monitoring through Agent Monitor hooks in Claude Code and Cursor to detect local activity including bash commands, file operations, and prompt submissions where supported. This creates two-layer visibility with coverage into supported local agent activity.

What performance overhead should I expect from self-hosted agent gateways?​

Performance varies by product and configuration. Bifrost reports 11 microseconds of gateway overhead at 5,000 requests per second in published benchmarks, but teams should run tests with their own routing, logging, and authentication settings. Evaluate whether your workloads require raw performance or whether governance features justify moderate overhead. High-throughput production deployments should benchmark gateway overhead under realistic load.

How do I evaluate open-source licensing for agent gateways?​

Check whether all features you need are available in the open-source core or gated behind Enterprise tiers. Apache 2.0 and MIT licenses dominate the category. Some vendors maintain fully open cores while others gate SSO, audit logs, and advanced security behind commercial licenses. For self-hosted deployment, ensure the license permits modification, distribution, and commercial use within your organization.