mcpgate vs Obot

Both are self-hosted MCP gateways aimed at teams that want AI agents connected to their daily tools — without sending data through a vendor cloud. They overlap in surface; the trade-offs decide which one fits which team. Sources are verified against both projects’ current public docs.

Last reviewed: . We maintain mcpgate. If something on this page is off, tell us and we’ll fix it.

The 30-second verdict

Pick Obot if…

  • You need real OSI-approved open source (MIT).
  • You’re Kubernetes-native and want a Helm chart day one.
  • You want a broader platform — gateway plus MCP server hosting plus a chat client — in one product.
  • You want a tool that has already been pressure-tested by hundreds of users (777 GitHub stars).

Pick mcpgate if…

  • PII handling is a hard requirement and you want it shipped, not as an engineering project. mcpgate pseudonymizes sensitive fields outbound and rehydrates them when the agent calls back. Obot doesn’t ship this layer; you could build it on top, but with mcpgate it’s already there.
  • You want individual users to define their own policy hooks (not just operator-level RBAC).
  • You prefer a focused gateway over a broader platform.
  • You don’t mind a source-available license (BSL 1.1, free for ≤5 users).

Where each one sits

Positioning quadrant: mcpgate vs Obot Focused — single-purpose gateway Broad — multi-product platform (gateway + registry + hosting + chat) Out-of-the-box ready (you configure) Composable OSS toolkit (you compose, harden, extend) mcpgate Obot
Same self-hosted-MCP-gateway use case; opposite engineering trade-offs. mcpgate ships an opinionated layer with PII handling, 22 deep native integrations, and an OpenAPI / MCP-server import path for anything else. Obot is the open platform you compose those pieces onto yourself. Neither axis says one is better — they answer different questions about who does the engineering.

Could you build mcpgate-equivalent on Obot?

Yes. Obot is open and composable — that’s the point of MIT-licensed platform infrastructure. With engineering investment you can:

If your team has the engineering capacity and wants to own every component end-to-end, that’s a legitimate path. Some mature teams prefer it because the stack stays in their codebase, fully visible and customisable, under an OSI-approved license they trust forever.

The trade-off is who does the engineering and who carries the maintenance cost:

The comparison below is then about out-of-the-box parity, not architectural ceiling. Where the table says “Obot doesn’t do this”, read “Obot doesn’t ship this — you’d add it.”

Side by side

mcpgate Obot
License BSL 1.1 (source-available; each release converts to Apache 2.0 after 3 years) MIT (OSI-approved open source)
Free tier Up to 5 users, all features Unlimited (open-source edition); Enterprise edition has separate paid features
Maturity Public since March 2026.. Early stage — community signals (stars, third-party citations) are still catching up. The product is feature-complete for the design target; market validation is the in-progress part. Established. ~777 stars / ~164 forks (verified 2026-05-17). Visible community contributions and a longer track record in the open.
Primary deployment Docker Compose, one command, zero config Docker run + Kubernetes Helm chart
Kubernetes-native Possible with custom manifests; no official chart Yes (Helm chart in repo)
PII pseudonymization with rehydration Built in. Sensitive fields are pseudonymized outbound; mapping stays on-prem, encrypted, 24h TTL; rehydrated when the agent calls back. Not documented as a feature.
Policy hooks Two layers: company-wide hooks set by operator, plus per-user hooks each user defines from their AI client. YAML, hot-reloaded. Gateway-level access policies (RBAC); not exposed as user-configurable hooks.
SSO providers Any OIDC IdP (Google, Microsoft, Okta, Keycloak, …) via OIDC_ISSUER_URL; broker-based Google + Microsoft setup also available. Google + GitHub (base edition); Okta + Microsoft Entra (Enterprise edition).
Pre-configured service integrations 23 built-in (Jira, GitLab, GitHub, Notion, Confluence, Slack, M365, GWS, Grafana, Sentry, Amplitude, Figma, Miro, …). OpenAPI import for any REST API; any MCP server via URL. Curated set of common MCP servers (Outlook, Slack, Notion, GitHub mentioned in marketing copy); the registry approach favors composability over a pre-built catalog.
Scope Gateway only. Gateway + MCP server hosting + MCP registry + multi-model chat client (“Obot Chat”).
RBAC Operator vs. user; not fine-grained per tool. Fine-grained, per skill / per MCP server.
Audit log Yes (admin-gated viewer; tamper-evident hashing). Yes.
Data residency 100% on-prem. Nothing flows through mcpgate.de or any third party. 100% on-prem for self-host; managed cloud trial also available.

The case for Obot

Maturity. Obot has been in the open long enough to accumulate a real contributor base and a visible track record (~777 stars verified 2026-05-17). mcpgate is in early stage — public since March 2026. The external validation surface (stars, third-party citations, “people I can talk to who run this”) is still catching up. If “has at least N production users on record” is on your checklist today, Obot has the head start. We don’t paper over that — it’s a normal early-stage gap, not a product gap.

License clarity. Obot ships under the MIT license, which is OSI-approved and widely understood. If your procurement team has a hard rule against source-available licenses, that’s a fair veto on mcpgate’s BSL 1.1.

Kubernetes day one. Obot’s repo includes a Helm chart and the deployment story is built around K8s. mcpgate works on Kubernetes — it’s containerized — but you build the manifests yourself.

Wider platform. Obot bundles MCP server hosting, a registry, and a chat client around the gateway. If you want a one-shop deployment that covers more than just the gateway layer, Obot reduces the integration surface.

The case for mcpgate

PII pseudonymization. This is the sharpest functional difference. mcpgate replaces sensitive fields (emails, names, phone numbers, customer IDs) with stable pseudonyms before the request reaches the LLM. The mapping stays on-prem, encrypted at rest, with a 24-hour TTL. When the agent calls back into the source system, the real values are rehydrated. The practical consequence: your LLM provider literally never sees the real names. Obot focuses on access control at the gateway layer; pseudonymization is not in its current feature set. If your security review explicitly asks “how do you prevent customer names from leaking to an external LLM?” you don’t have a good answer with Obot today.

User-defined policy hooks. Most gateways let the operator set policies. mcpgate also lets each end-user define their own hooks, hot-reloaded from their AI client. Example: an individual user can say “always set assignee=me on Jira tickets I create through Claude” without filing a request with the operator. This is YAML, no code, and a real productivity unlock for teams where individuals have idiosyncratic preferences. Obot’s policy model is one-way (operator down).

Broader pre-built service catalog. mcpgate ships 23 services pre-configured (Jira, GitLab, GitHub, Notion, Confluence, Slack, M365, Google Workspace, Grafana, Sentry, Amplitude, Figma, Miro, Joan, Transifex, Home Assistant, Phase-6 Content, App Store Connect, Google Play, Bigquery, Metabase, Wordpress, Supernova). Obot leans on its registry: you compose from existing MCP servers, but the out-of-box catalog is smaller.

Focused scope, smaller learning curve. Obot is four products in one. That’s strength if you want the bundle; it’s overhead if you only want the gateway. mcpgate is one thing.

Tradeoffs

Try one — they’re both free to evaluate