engineering
Best enterprise MCP registries and gateways (2026)
Ten enterprise MCP gateways and registries reviewed on per-user OAuth, access control, audit logs, versioning and EU hosting, plus when an LLM gateway is enough and when you need both.
- #mcp
- #mcp-gateway
- #mcp-registry
- #governance
- #roles
What are the best enterprise MCP gateways and registries in 2026?
For an enterprise MCP gateway with access control and execution logs, the strongest hosted options in late 2026 are Runlayer, MintMCP (now Mintive) and TrueFoundry; Obot is the strongest open-source one. For discovery metadata, the official MCP Registry is the open upstream most private catalogues build on. skilder covers a different need: business-team roles on a governed MCP layer.
These tools sit at different layers, which is why buyer comparisons often go wrong:
- Gateways proxy every MCP call, enforce who can call which tool, and log it.
- Registries list which MCP servers exist and which are approved.
- Roles for business teams bundle MCP tools, skills (the team’s method and expected output) and apps that a team owns and employees run from the assistant they already use. The industry calls these bundles plugins; skilder calls them roles.
Every feature claim below was checked on the vendor’s own site or documentation in October and November 2026. “Not documented” means the page reviewed does not say; it does not mean the feature is missing. skilder publishes this article, includes itself as one option, and states where other tools fit better.
How were the tools selected and scored?
The list keeps tools a European company of 500 to 5,000 people could put in front of employees today, scored on six questions a security review will ask. Does each employee call tools with their own OAuth identity? Who may call which tool? Is every call logged? Can changes be rolled back? Where is data hosted? Is the code open?
Per-user OAuth means a tool call never runs on a shared service account; role-based access control (RBAC) covers vendors’ policy or attribute-based variants; audit logs must record user, tool, arguments and outcome, and export to a security information and event management (SIEM) system; a GDPR claim alone does not count as an EU hosting option. Developer SDKs and listings that neither run calls nor enforce policy are excluded. Most enterprise tiers are quote-only, so no prices appear.
Which 10 enterprise MCP gateways and registries made the list?
Ten tools made the list: four dedicated gateways (Runlayer, Mintive, TrueFoundry, Obot), one role layer for business teams (skilder), three gateways built into wider stacks (Docker, LiteLLM, Azure API Management) and two registries (Kong MCP Registry, official MCP Registry). Each entry gives strengths, limits and the team it fits best.
1. Runlayer: hosted control plane for security-led rollouts
Runlayer is a commercial gateway for MCP and LLM traffic, plus an internal catalogue of approved connectors, skills and agents. Its deployment documentation describes SSO (OAuth 2.0 or SAML), RBAC with policy and attribute-based rules, a security scanner called ToolGuard, per-credential rate limits and audit logging.
- Strengths: runtime scanning for personal data and exfiltration, human approval for risky actions, shadow AI discovery, an optional OAuth broker. The vendor lists SOC 2, HIPAA and GDPR compliance.
- Limits: no public pricing; versioning not documented; a reference architecture centred on AWS, with hybrid designs on request; not open source.
- Best for: security and IT teams that want one policy point for every agent and MCP call, including developer IDEs.
2. MintMCP (now Mintive): managed MCP gateway with SOC 2 Type II
MintMCP now trades as Mintive: the mintmcp.com domain redirects to mintive.com. Its launch release on Business Wire (February 2026) describes one-click deployment of custom and open-source MCP servers, SSO and OAuth, centralised credentials, an agent monitor that traces tool calls, and policy-based guardrails.
- Strengths: SOC 2 Type II audit stated in that release, encryption in transit and at rest, data residency options, enterprise SLAs.
- Limits: product documentation was not reachable after the rebrand, so RBAC wording, versioning and self-hosting are not documented on the pages reviewed; hosted only.
- Best for: teams that want a managed gateway with an audit report ready for procurement.
3. TrueFoundry MCP Gateway: part of a full AI gateway
TrueFoundry’s MCP Gateway sits inside its AI Gateway, next to LLM routing. It offers a curated server registry, two-legged and three-legged OAuth, per-user credential overrides, per-tool enable and disable, guardrails around each tool call, and an audit trail with latency and error metrics.
- Strengths: one gateway for model and tool traffic; virtual servers that combine tools from several servers; control plane and gateway self-hosted on Kubernetes with Helm, or SaaS.
- Limits: access control is per caller, not named roles; versioning not documented; not open source.
- Best for: platform teams that run, or want, a central LLM gateway and want MCP under the same policies.
4. Obot: the open-source MCP gateway and catalogue
Obot is an MIT-licensed MCP gateway with an admin-approved catalogue of MCP servers and skills. Admins set which users, groups and agents can call which tools; every tool call is logged with user, agent, server, arguments and outcome.
- Strengths: full source; runs on Docker or Kubernetes in any cloud or on-premise; admins deploy, version and retire servers from one console; a token broker keeps raw credentials away from agents.
- Limits: according to Obot’s site (2026), the free Community edition is capped at 100 users and 100 devices; unlimited users and SLAs require Obot Cloud or Enterprise.
- Best for: teams that need to self-host and read the code.
5. skilder: roles for business teams on a governed MCP layer
skilder makes a team’s know-how (its method, its tools and the format of its output) available in each employee’s own AI assistant. The container is a role, such as a credit analyst or a site manager role: it bundles the team’s skills, its scoped MCP tools from an internal, curated MCP registry, its access rights and, where useful, an app interface. A role is configured from a blueprint, a starting configuration the team adapts to its own method. The business team owns and maintains its roles; IT governs access, permissions, hosting and audit. Employees use a role from Claude, ChatGPT, Copilot Studio, VS Code or any MCP-compatible agent. skilder is not a pure MCP gateway and can run alongside one.
- Strengths: per-user execution, so a role’s tools act with the requester’s own permissions and never escalate them; RBAC and workspace isolation; enterprise SSO; OAuth for connecting assistants over MCP; an execution trace attributed per user and an exportable audit log; versioning and rollback (documentation); zero retention at the model providers; hosting in Switzerland (Infomaniak) and the European Union, on-premise on request, data processing agreement available.
- Limits: not open source; not a general-purpose proxy for every developer MCP server (pair it with a gateway for that); no SCIM provisioning; Microsoft support covers Copilot Studio, not the Microsoft 365 Copilot chat.
- Best for: business teams that want roles they run themselves (credit analysis, site reports, consultant onboarding) in the assistants they already use. Free self-service is enough for a first role.
6. Docker MCP Catalog, Toolkit and Gateway: containers for MCP servers
Docker’s MCP Catalog documentation (2026) lists more than 300 verified servers packaged as container images with versioning and provenance. The MCP Gateway is MIT-licensed and handles OAuth flows, secrets, logging and call tracing.
- Strengths: isolation by container; custom catalogues of approved servers; profiles shared across Claude Code, Cursor and other clients.
- Limits: enterprise governance is invite-only through Docker Sales; RBAC not documented on the pages reviewed.
- Best for: engineering teams already standardised on Docker Desktop.
7. LiteLLM: open-source proxy with MCP permissions
LiteLLM is an open-source LLM proxy that also routes MCP. It supports per-user OAuth with PKCE, on-behalf-of token exchange, and MCP permissions by key, team, end user, agent and organisation, with the most restrictive level winning.
- Strengths: free to run; tool allow and deny lists; tool-list pinning that alerts on upstream drift.
- Limits: audit logs, SSO and RBAC roles are enterprise-licence features; no version history for registered servers.
- Best for: teams already using LiteLLM for model routing.
8. Azure API Management: MCP for Microsoft shops
Azure API Management can expose a managed REST API as an MCP server, or front an existing MCP server. Policies add Entra ID token validation, rate limits and IP filtering; Azure Monitor captures traffic, and Azure API Center acts as a private MCP registry.
- Strengths: reuses existing API Management policies and subscriptions; self-hosted gateway option; Azure EU regions.
- Limits: tools only (no MCP resources or prompts); per-user delegated OAuth not described on these pages.
- Best for: organisations that already govern APIs in Azure.
9. Kong MCP Registry: API catalogue extended to MCP
Kong’s MCP Registry entered tech preview in Konnect in February 2026, according to Kong’s press release. It links MCP servers to their APIs and owners, returns only the tools an agent may use, and pairs with Kong AI Gateway for rate limits and personal data controls.
- Strengths: fits teams that already run Kong; tracks tool usage and context bloat.
- Limits: tech preview; RBAC, OAuth and versioning not documented in the announcement.
- Best for: existing Kong Konnect customers.
10. Official MCP Registry: the open upstream
The official MCP Registry launched in preview in September 2025 as the public catalogue of MCP servers, according to the Model Context Protocol blog. It is open source, and enterprises are expected to build private sub-registries on top of it.
- Strengths: open, permissively licensed, with a shared API schema reused by other catalogues.
- Limits: metadata only (no gateway, authentication or logs); a preview with no durability guarantees.
- Best for: the discovery layer feeding your own private catalogue.
How do the ten tools compare on access control, logs and hosting?
On the pages reviewed, skilder and Obot document access control, audit logs and versioning together, while TrueFoundry and LiteLLM document per-user OAuth most clearly among the gateways. Features were checked in October and November 2026; vendors ship monthly in this category, so recheck before a purchase.
| Tool | Open source | Per-user OAuth | Access control | Audit logs | Versioning | Hosting and EU options |
|---|---|---|---|---|---|---|
| Runlayer | No | Not documented (OAuth broker, SSO) | RBAC, policy and attribute rules | Yes | Not documented | Hosted or your cloud; EU region not documented |
| MintMCP / Mintive | No | Not documented | Access controls; RBAC not named | Yes | Not documented | Hosted; residency options, regions not listed |
| TrueFoundry | No | Yes (per-user overrides) | Per caller, per tool | Yes | Not documented | SaaS or self-hosted; EU not documented |
| Obot | Yes (MIT) | Token brokering | By user, group, agent, tool | Yes, every call | Yes (servers) | Self-hosted anywhere; Obot Cloud |
| skilder | No | Yes (per-user execution) | RBAC, workspace isolation | Yes, per-user trace, exportable | Yes, with rollback | Switzerland (Infomaniak), EU, on-premise on request |
| Docker MCP | Gateway (MIT) | Not documented | Not documented | Yes, call tracing | Image versions | Local; enterprise via Docker Sales |
| LiteLLM | Yes, plus Enterprise | Yes (PKCE, on-behalf-of) | Key, team, organisation | Enterprise only | Not documented | Self-hosted |
| Azure API Management | No | Not documented | Entra ID, subscriptions | Azure Monitor | Not documented | Azure regions; self-hosted gateway |
| Kong MCP Registry | Not documented | Not documented | Policy controls | Yes, no detail | Not documented | Kong Konnect (preview) |
| Official MCP Registry | Yes | Not applicable | Not applicable | Not applicable | Not documented | Public, community-run |
MCP gateway vs LLM gateway: which do you need?
An LLM gateway sits between your applications and the model providers: it routes prompts, handles fallbacks, caches responses and tracks token cost. An MCP gateway sits between agents and the tools they call: it checks who may call which tool, with whose credentials, and logs each call. Teams whose agents act on internal systems usually need both.
The two answer different audit questions. The LLM gateway tells you which model saw which prompt and what it cost. The MCP gateway tells you which tool acted on which system, and as whom. A prompt filter does little against an agent that calls a tool with broad rights; per-user credentials and tool-level permissions do, as argued in least privilege for AI agents.
| Layer | Sits between | Controls | Examples |
|---|---|---|---|
| LLM gateway | Your applications and model providers | Routing, fallbacks, token budgets, prompt filters | LiteLLM, TrueFoundry, Kong AI Gateway |
| MCP gateway | Agents and MCP servers | Tool access per user, credentials, call logs | Runlayer, Mintive, Obot, Docker |
| MCP registry | Admins and agents, at discovery | Which servers exist and are approved | Official MCP Registry, Kong, Azure API Center |
When employees work in Claude, ChatGPT or Copilot as hosted services, their model traffic goes to the vendor directly and never passes through your LLM gateway. The calls those assistants make to your remote MCP servers do reach your infrastructure, so the MCP layer is where you enforce rights and keep a trace. An LLM gateway earns its place for the applications and agents you build yourself.
TrueFoundry, LiteLLM and Runlayer route both kinds of traffic through one gateway. If LiteLLM already proxies your models, adding MCP there is the shortest path; the comparison of LiteLLM, agentgateway and skilder lays out the trade-offs. For where discovery stops and enforcement starts, see how an MCP registry differs from an MCP gateway.
How do you choose an enterprise MCP gateway or registry?
Start from the problem you must solve this quarter, not from the longest feature list. Control of every MCP call points to a gateway. An existing API stack points to extending it. Business teams running their own work in Claude, ChatGPT and Copilot point to roles.
- You must control and log every MCP call, including developer IDEs. Shortlist Runlayer or Mintive for hosted with a compliance report, TrueFoundry if you also route LLM traffic, Obot if you must self-host. Start with the servers employees already plug in on their own, described in the shadow MCP problem.
- You already run an API stack. Extend it: Azure API Management with API Center, Kong MCP Registry on Konnect, or LiteLLM if it proxies your models.
- Business teams should run their own roles in Claude, ChatGPT and Copilot. You need roles that bundle MCP tools, skills and, where useful, an app, with owners, versions and per-user permissions. skilder targets this and can run alongside a gateway for developer traffic. Runlayer and Obot now list skills too, so compare how each publishes and rolls back what business users run.
- Data must stay in Switzerland or the EU. Ask for the region in writing. Self-hosted Obot, TrueFoundry or LiteLLM give full control; skilder documents Swiss and EU hosting, detailed in the guide to sovereign AI in Switzerland.
- Run a two-week pilot with one team, three tools and your SIEM connected, and check that the logs answer “who ran what, as whom”. To scope that pilot, the free 30-minute exchange with an expert covers which layer to set up first.
The layers combine. A common pattern is the official MCP Registry as upstream, a gateway for traffic policy, and roles for what business teams run. Since skilder governs its own MCP layer, a separate gateway matters most for developer traffic.
To test that pattern against your own assistants, systems and hosting constraints, book a free 30-minute exchange with an expert.
Takeaway: put a gateway on the traffic you must police, and give business teams roles they own for the work they run.
Frequently asked questions
What is the difference between an MCP gateway and an MCP registry?
A registry lists which MCP servers exist and which are approved for use. A gateway sits in the call path: it authenticates the user, enforces which tools that user may call, and logs each call. The official MCP Registry, launched in preview in September 2025, holds metadata only, so it needs a gateway or a governed layer next to it. Many enterprise products, such as Obot and TrueFoundry, bundle a private catalogue with a gateway.
Does an LLM gateway replace an MCP gateway?
No. An LLM gateway controls traffic to model providers: routing, fallbacks, caching and token cost. An MCP gateway controls what agents do with tools: who may call which tool, with whose credentials, and what happened. When employees use hosted assistants such as Claude, ChatGPT or Copilot, model traffic bypasses your LLM gateway, but calls to your remote MCP servers still reach your infrastructure. TrueFoundry, LiteLLM and Runlayer handle both kinds of traffic in one product.
Which MCP gateway has role-based access control and execution logs?
Runlayer documents role-based, policy-based and attribute-based access control with audit logging. Obot, TrueFoundry and skilder document access policies plus call-level logs, and skilder attributes each execution trace to a user and lets admins export the audit log. In LiteLLM, RBAC roles and audit logs require the Enterprise licence, according to its documentation in 2026. Check that the logs record user, tool, arguments and outcome before you sign.
Is there an open-source MCP registry for enterprise use?
Yes. The official MCP Registry is open source and designed as an upstream for private sub-registries, but it holds metadata only, with no authentication or logs. Obot, under the MIT licence, adds an open-source catalogue, gateway and audit log you can self-host on Docker or Kubernetes. Its free Community edition is capped at 100 users and 100 devices, so larger rollouts move to Obot Cloud or the Enterprise edition.
Should we self-host an open-source MCP gateway or buy a hosted one?
Self-host if you have a team to run, patch and secure the gateway, and you need the code on your own infrastructure. Buy hosted if you want SSO, service level agreements and audit reports from day one. Obot offers both paths from one codebase, and TrueFoundry can run as SaaS or on your own Kubernetes cluster. For Swiss or EU data residency, ask each vendor for the hosting region in writing.
What does per-user OAuth mean for MCP tools?
Each employee signs in to the downstream system as themselves, so the tool call carries their own permissions rather than those of a shared service account. A user can never reach data they could not open directly, and revoking one person's access does not break the tool for everyone else. TrueFoundry and LiteLLM document per-user OAuth flows, and skilder runs each role's tools with the requester's own permissions.
Can MCP skills be versioned and rolled back like code?
On some products, yes. skilder documents version history, publishing and rollback for roles, so a change to a skill or tool list can be reverted without touching every assistant. Obot documents versioning of MCP servers, and Docker relies on container image versions. Most other gateways do not document versioning on their public pages, so ask whether a rollback reaches all connected clients at once.
Related articles
-
November 11, 2026
MCP gateway vs skills platform vs Copilot Studio: which one to start with
An MCP gateway governs tool access, Copilot Studio builds Microsoft 365 agents, and what buyers call a skills platform is really agent roles. What each one does, and which one to start with.
-
November 8, 2026
Runlayer and Composio alternatives for distributing MCP tools internally
Runlayer and Composio govern or supply MCP tools for engineers. Six alternatives compared on end users, per-user auth, governance and hosting, with a shortlist by team and a method to reach business users.
-
November 1, 2026
MCP registry vs MCP gateway: what each one does, and why you need both
An MCP registry decides which MCP servers exist and are approved. An MCP gateway enforces policy on every call. What each one does, where it sits in the request path, and what neither of them solves.