Zuplo
API Gateway Comparisons

Zuplo vs
Arcade.dev

  • SOC 2 Type II
  • 99.999% SLA
  • 300+ edge locations

The Independent, Code-First MCP Gateway Teams Choose Over a US-Only, Closed-Runtime Tool Catalog

Talk to an Architect Compare with AI
Feature
Zuplo
Arcade
Product Category and Architecture
Deployment Model and Data Residency
Audit Logs, RBAC, and SSO Tier Gating
Programmable Policies (Code vs. Webhooks)
GitOps and Developer Workflow
Per-User OAuth Brokering and Token Vault
AccuWeather
Read the Case Study
Blockdaemon
Read the Case Study
Lake Michigan Credit Union
Read the Case Study
Finsolutia
Read the Case Study
VNDR
Read the Case Study
Mews
Read the Case Study
Yext
Read the Case Study
Zumiez
Read the Case Study

What's wrong with Arcade

Arcade's key limitations for modern engineering teams

The forces driving enterprises off Arcade in 2026 — operational tax, plugin sprawl, retrofitted AI, and pricing that doesn't predict.

Arcade Cloud is US-only with no multi-region option

All of Arcade Cloud's infrastructure — control plane, MCP server clusters, data storage and processing — is located in the United States, per the deployment docs. There is no EU region, no APAC region, no managed multi-region SaaS. Teams with sovereignty or residency requirements are pushed to self-hosted Kubernetes — which is itself an Enterprise-tier offering, deployed as a closed binary via Helm chart.

Audit logs, RBAC, and SSO are Enterprise-tier-only

Per the published pricing page, Audit Logs, Role-Based Access Control, SSO, and SAML for the Arcade admin console are all Enterprise-tier features. Hobby and Growth tiers share a single tenant with no isolation. For a product whose entire pitch is enterprise authorization, basic enterprise governance starts at a sales call.

Tool training data collection is opt-out by default

Arcade Cloud collects tool queries, tool execution inputs, and tool execution results for ML training by default, with retention up to 5 years. Opt-out is org-level and takes effect immediately — but the default is opt-in. For regulated workloads handling customer payloads through tools like Stripe, Salesforce, or HubSpot, this is a non-trivial default to inherit. (Source: Arcade Cloud docs.)

Engine is a closed binary and tools went closed-source in July 2025

The Arcade Engine is distributed as a closed Docker image and binary — not source-buildable. On July 25, 2025 Arcade transitioned most MCP server source code out of the public arcade-ai monorepo to closed-source, with source "available upon request" only for paying customers (per the public changelog). The OSS framework (arcade-mcp) remains MIT-licensed, but the runtime and tool catalog you actually depend on are not.

Hooks are webhooks, not in-process policies

Arcade's Contextual Access extensions (Access, Pre-Execution, Post-Execution) are external webhooks that you host yourself, with a default 5-second timeout per call. Every PII redaction, allow/deny check, or payload transform adds webhook round-trip latency. There is no first-party JavaScript or WASM sandbox for policy code — policies are arbitrary code in an HTTP endpoint you operate.

No native GitOps or PR preview environments

The documented deployment workflow is arcade deploy from the CLI to push a local MCP server to Arcade Cloud. There is no Git-driven promotion model, no per-branch preview environments, and no first-class CI/CD integration. For teams that expect branch = environment, PR = live preview, push = global deploy, the Arcade workflow is a step back.

Why Zuplo

Built for teams replatforming off Arcade.dev

Managed, modern API management with predictable economics across procurement cycles — no operator overhead, no plugin sprawl, no consumption-pricing surprises.

Product Category and Architecture

Programmable gateway with full API surface vs. MCP-only runtime with curated tool catalog.

Deployment Model and Data Residency

Managed edge, dedicated cloud, and self-hosted with region pinning vs. US-only SaaS with self-host as the only non-US path.

Audit Logs, RBAC, and SSO Tier Gating

OIDC SSO and policy controls on all paid tiers vs. governance basics behind an Enterprise sales call.

A solutions architect can walk you through your current Arcade setup, surface the biggest operational tax, and map a migration path — no slide deck required.

Talk to an Architect Evaluate your migration
"We didn't touch a line of code, it's just plug and play. The results were very surprising, in just a couple of hours we had a great result and a fully working MCP Server."
Finsolutia

Miguel Madeira

CTO & Co-Founder, Finsolutia

Read the case study

Enterprise ready

Production-ready for regulated and high-volume workloads

Compliance & Audit

  • SOC 2 Type II audited annually
  • Third-party penetration test reports available under NDA
  • GDPR-aligned data processing
  • Audit logs across the control plane
  • API governance with policy enforcement

Identity & Access

  • SAML SSO and SCIM provisioning
  • Role-based access control across organizations, projects, and environments
  • Service-account credentials with scoped permissions
  • API key metadata for downstream authorization

Deployment Flexibility

  • Managed edge across 300+ locations — global by default
  • Managed dedicated single-tenant on AWS, Azure, GCP, Akamai, or any major cloud
  • Self-hosted on Kubernetes with full control plane
  • Bring-your-own-cloud for data residency requirements

Support & Success

  • Up to 30-minute response SLA on Enterprise
  • 24/7/365 emergency hotline for critical incidents
  • Named technical account manager
  • Architecture and migration professional services

Built for the AI era

Built as a programmable MCP gateway, not a closed-runtime tool catalog

Arcade ships an opinionated, US-only SaaS runtime with a Python-first SDK, a managed tool catalog that went closed-source in July 2025, and webhook hooks for custom policy. Audit logs, RBAC, SSO, dedicated tenancy, and self-hosting are Enterprise-only. Zuplo gives you the same OAuth 2.1 inbound, per-user upstream OAuth brokering, and tool-level RBAC as first-class policies in a TypeScript-programmable gateway that runs on 300+ edge POPs, your dedicated cloud, or your Kubernetes cluster.

OAuth 2.1 with RFC 8707 resource indicators

Zuplo's MCP Gateway is both an OAuth 2.1 Resource Server and Authorization Server. The <code>resource</code> parameter is required on every authorization and token request, and tokens are bound to a specific virtual MCP server — a token minted for one server is rejected at another. PKCE S256 required, RFC 7009 token revocation published at <code>/oauth/revoke</code>. CIMD (OAuth Client ID Metadata Documents) preferred over Dynamic Client Registration for trusted MCP clients like Claude Desktop, Claude Code, and VS Code.

Per-user OAuth brokering with encrypted token vaulting

Each end-user completes the upstream's OAuth flow themselves; the gateway brokers and refreshes their tokens per session with per-user revocation. Refresh tokens are sealed in the gateway vault, encrypted at rest and keyed to the user's subject ID, and never returned to the MCP client. Token passthrough is explicitly forbidden by the MCP spec and enforced — the inbound token a client presents and the upstream token the gateway forwards are never the same token. Works with Linear, Notion, Stripe, GitHub, Grafana Cloud, Microsoft 365, Slack, PostHog, and any OAuth-compliant MCP server.

Virtual MCP servers and capability filtering

Compose multiple upstream MCP servers behind one Zuplo deployment with a curated tool list per role or team. The <code>mcp-capability-filter-inbound</code> policy filters tools, prompts, resources, and resource templates by exact identifier. Calls to hidden tools return JSON-RPC <code>MethodNotFound</code> before the request is forwarded upstream, so even cached tool names can't be invoked. Finance gets a read-only view of Stripe and QuickBooks; engineering gets GitHub and deployment tools — same governance, different exposure.

OpenAPI-to-MCP and programmable TypeScript policies

The MCP Server handler (GA) auto-generates an MCP server from any OpenAPI specification — Finsolutia stood up a working MCP server in under 24 hours with zero code changes to their REST APIs. Custom policies are TypeScript code, not webhook calls, executing in the same edge runtime as the rest of the gateway. Pre and post hooks compose anywhere in the inbound chain alongside the MCP OAuth, token exchange, and capability filter policies.

See it in action

See Zuplo running on your stack

A 30-minute working session with a Zuplo solutions engineer. Bring an OpenAPI spec or a Kong route definition and walk away with a working preview.

Talk to Sales Compare features

Side by side

Feature-by-feature comparison

Feature
Zuplo
Arcade
Product Category and Architecture
Independent, code-first API and MCP gateway built on a TypeScript runtime. Same programmable model handles MCP traffic, AI/LLM routing, and full API management. MCP Gateway in beta with design partners; MCP Server handler (OpenAPI-to-MCP) generally available on every tier including Free.
MCP runtime focused on per-user OAuth brokering and a managed tool catalog. Explicitly argues against the gateway-as-control-point model. Architecturally a closed-binary Engine, a catalog of ~145 MCP servers covering thousands of tools, an MCP Gateway URL that federates them, and webhook hooks for custom policy. (Source: Arcade blog.)
Deployment Model and Data Residency
Managed Edge across 300+ data centers by default. Managed Dedicated single-tenant on AWS (full VPC integration), Azure (private networking), GCP (custom networking), Akamai Connected Cloud, Equinix, or TerraSwitch. Self-Hosted on any Kubernetes cluster via Helm chart. Region pinning available for sovereignty and data residency requirements, including EU and APAC.
Arcade Cloud runs in US-East (Ohio) only — control plane, MCP server clusters, and data storage are all in the United States. No multi-region managed SaaS option. EU/APAC residency requires self-hosting on Kubernetes via Helm chart — an Enterprise-tier offering, deployed as a closed binary, not source-built. (Source: Arcade Cloud docs.)
Audit Logs, RBAC, and SSO Tier Gating
OIDC SSO via Auth0 presets or any OIDC provider on all paid tiers. Per-user JWT claims drive per-team and per-tenant authorization at the gateway edge. SAML SSO available as an Enterprise add-on. Audit logs and penetration test reports under NDA available as Enterprise add-ons. Per-route capability filtering and JWT-claim-driven authorization on all paid tiers — no sales call required.
Audit Logs, Role-Based Access Control, SSO, and SAML are all Enterprise-tier-only per the published pricing page. Hobby and Growth tiers share a single tenant with no isolation. End-user identity (User Sources) is OIDC-only; SAML is not available for User Sources, only for the Arcade admin console — and only at Enterprise.
Programmable Policies (Code vs. Webhooks)
Custom policies are TypeScript code running in the edge runtime alongside the gateway, with full type-safe access to ZuploRequest / ZuploContext. custom-code-inbound and custom-code-outbound policies compose anywhere in the inbound or outbound chain. Same runtime as the rest of the Zuplo platform. No webhook RTT, no separate service to operate.
Contextual Access "Logic Extensions" (Access, Pre-Execution, Post-Execution hooks) are external webhooks that you host yourself, with a default 5-second timeout per call. Every redaction, allow/deny check, or payload transform adds webhook round-trip latency. No first-party JavaScript or WASM sandbox — policies are arbitrary code in an HTTP endpoint you stand up and run.
GitOps and Developer Workflow
Git is the source of truth. Every branch maps to an environment — push to main and deploy main, push to feature-auth and get a feature-auth environment automatically. PR preview environments let reviewers test changes against a live gateway before merging. Global deploys to 300+ POPs in under 20 seconds. GitHub native integration with auto-deploys; GitLab, BitBucket, and Azure DevOps source-control integration available on Enterprise.
The documented workflow is arcade deploy from the CLI to push a local MCP server to Arcade Cloud. No Git-driven promotion model, no per-branch preview environments, and no first-class CI/CD integration in the public docs. Branch = environment, PR = live preview is not the Arcade pattern.
Tool Catalog and Custom Tools
Auto-generates MCP servers from any OpenAPI specification — Finsolutia stood up a working MCP server in under 24 hours with zero code changes to their existing REST APIs. @zuplo/mcp is published on npm as a fetch-API-based, remote-server-first TypeScript SDK that runs on Zuplo, Node, Deno, and Workerd. Build custom tools in TypeScript with full @zuplo/runtime type safety. Independent of any pre-built catalog — federate whichever upstream MCP servers your team actually uses.
~145 pre-built MCP servers covering thousands of tools, split into "Arcade Optimized" (hand-crafted, LLM-friendly), "Arcade Unoptimized" (auto-generated from raw APIs — Arcade's own docs warn to "thoroughly evaluate each Unoptimized tool with your Agents before using it in production"), Verified, Community, and Auth Provider designations. As of July 25, 2025, most tool source code moved out of the public arcade-ai repo to closed-source. Custom tools built with the Python arcade-mcp-server SDK or Java/Spring AI (TypeScript SDK is early — 2 GitHub stars).
Per-User OAuth Brokering and Token Vault
Per-user OAuth brokering with encrypted token vaulting, keyed to the user's subject ID. Token refresh is transparent — when the upstream returns a 401 mid-request, the gateway refreshes the upstream credential and retries once. Reconsent flow built in — when an upstream policy revokes a refresh token, the next request surfaces reconsent_required and the user re-authorizes through the same browser flow. RFC 7009 token revocation published at /oauth/revoke.
Per-user OAuth brokering with AES-256 application-layer encryption on tokens before persisting to storage. Auth provider catalog includes Asana, Atlassian, Discord, Dropbox, GitHub, Google, HubSpot, Linear, LinkedIn, Microsoft, Notion, Reddit, Slack, Spotify, Twitch, X, Zoom, plus generic OAuth 2.0. Note: Arcade's default OAuth apps cannot be used in production multi-user apps — you must register your own OAuth apps per provider and add them to Arcade. (Source: production auth docs.)
MCP Spec Compliance
Tracks the MCP 2025-11-25 revision for authorization and protocol. Streamable HTTP transport on POST; GET returns 405. Stateless gateway — no MCP session affinity required, scales horizontally on the edge. RFC 8707 resource indicators mandatory. RFC 7636 PKCE S256 required. OAuth Client ID Metadata Documents (CIMD) preferred over Dynamic Client Registration.
Supports the MCP 2025-11-25 spec via the arcade-mcp framework — Arcade co-authored the SEP for URL Elicitation in that revision in collaboration with Anthropic. Streamable HTTP transport. Allowlist mechanism for trusted MCP clients via Client ID Metadata Document URL (Claude Desktop, Claude Code, VS Code). Arcade docs explicitly claim RFC 6750 and RFC 9728; RFC 8707 resource indicators are not mentioned in public docs.
Open Source Footprint
@zuplo/mcp open-source TypeScript SDK published on npm, designed to be "minimum common API" compliant per WinterTC. Inspired by and attributing the official modelcontextprotocol/typescript-sdk. Zudoku developer portal (MIT, ~540 GitHub stars). 81 public repos on the GitHub org. Source-available platform with TypeScript escape hatches.
arcade-mcp framework is MIT-licensed (~900 GitHub stars) and is genuine OSS. The Arcade Engine itself is a closed binary distributed via Brew, APT, and a private Docker image — no source repo published. On July 25, 2025, most MCP server source code was moved out of the public arcade-ai monorepo to closed-source — "source code remains available upon request for our paying customers."
Identity Provider Integration
First-class MCP OAuth presets for Auth0, Amazon Cognito, Clerk, Google, Keycloak, Logto, Microsoft Entra ID, Okta, OneLogin, PingOne, and WorkOS. Generic OIDC for anything else — drop in your issuer URL and your customers click Connect. SAML SSO available as an Enterprise add-on. Per-user JWT claims accessible from any policy via request.user.data.
End-user identity ("User Sources") is OIDC-only — Entra ID, Okta, Auth0, Clerk, Stytch, or "a similar OIDC provider." SAML is not supported for User Sources. SSO and SAML for the Arcade admin console are Enterprise-tier features. Custom subject claim configurable; PKCE required on User Source OIDC flow.
AI Gateway and Guardrails
Full AI Gateway alongside the MCP Gateway. Multi-provider routing across OpenAI, Anthropic, Google (Gemini), and Mistral via a Universal API, with automatic fallback when a provider rate-limits or errors. Token-based rate limits and hierarchical dollar budgets (org → team → sub-team → app) with Enforce or Warn modes. Semantic caching, prompt injection detection, and the Akamai AI Firewall partnership policy (DLP, toxic content filtering, adversarial protection). Secret masking on outbound responses.
No first-class AI Gateway product. Arcade is model-agnostic and integrates with LangChain, CrewAI, Google ADK, AG2, OpenAI Agents, Mastra, Vercel AI, TanStack AI, and Spring AI — but routing, budgets, semantic caching, and provider guardrails are not part of the Arcade platform. You compose Arcade with whichever LLM router and guardrail vendor you already use.
Observability and Audit Trail
Typed analytics events across the MCP request lifecycle — mcp_request, capability_invocation, auth_event — with seven outcome classes and per-event subjectId, upstreamServerId, latency split gateway-vs-upstream, failure origin, and reason code. Structured logs export to Datadog (native plugin), AWS CloudWatch, Dynatrace, Google Cloud Logging, Loki, New Relic, Splunk, Sumo Logic, or any OTLP-compatible destination (Honeycomb, Tempo, self-hosted Jaeger). Audit logs available as an Enterprise add-on.
Audit logs cover administrative actions only (API_KEY, PROJECT, GATEWAY, HOOK, etc.) — the documented audit-log REST API has no TOOL_CALL resource type, so per-tool-call audit trails route through a separate OpenTelemetry pipeline. Audit logs are an Enterprise-tier feature. Execution logs are stored as Application data with user-controlled retention.
Data Handling and Training Defaults
Customer data is encrypted at rest. Zuplo does not train models on customer payloads. SOC 2 Type II audited; penetration test reports available under NDA on Enterprise.
Arcade Cloud collects "tool queries, tool execution inputs, tool execution results" for ML training by default, with retention up to 5 years. Opt-out is org-level and takes effect immediately — but the default is opt-in. For tools that handle Stripe payloads, Salesforce records, HubSpot CRM data, or Microsoft 365 content, that is a non-trivial default to inherit. (Source: Arcade Cloud docs.)
Compliance Posture
SOC 2 Type II via Trust & Compliance Report, third-party penetration test reports available under NDA, audit logs across the control plane (Enterprise add-on), 24/7/365 emergency hotline available on Enterprise. Up to 99.999% uptime SLA.
SOC 2 Type II achieved August 2025 (Source: SOC 2 announcement). HIPAA, FedRAMP, ISO 27001, and PCI-DSS are not currently claimed — framed as "future work." The SOC 2 blog post itself hedges: "Today's SOC 2 Type 2 is table stakes."
SDK Languages and Framework Coverage
TypeScript runtime with @zuplo/runtime, @zuplo/mcp, and @zuplo/cli on npm. Browser-based authoring in the Zuplo portal with Working Copy environments. Local development via the CLI. Drop-in compatibility with the OpenAI SDK baseURL for AI Gateway traffic.
Python is the primary SDK surface (arcade-mcp-server) for authoring custom tools. Java/Spring AI SDK added April 2026. TypeScript MCP server framework (arcade-mcp-ts) exists but is early (~2 GitHub stars). Client SDKs in Python, Node, Go, and Java. Built-in framework integrations for AG2, CrewAI, Google ADK, LangChain, OpenAI Agents, Mastra, Vercel AI, TanStack AI, Spring AI.
Pricing and Commercial Model
Free tier with 100K requests/month, 5 environments, 100 consumers, 1 MCP server, and 1K MCP tool calls/month. Builder $25/month includes 100K requests with $100 per additional 100K up to 1M, plus 2 custom domains, 10 environments, 10K MCP tool calls/month, and 5 MCP users. Enterprise from $1,000/month on annual contract with 99.5% SLA at entry and up to 99.999% available, volume discounts past 1M requests/month. SSO + RBAC, SOC 2 controls, Managed Dedicated, and Self-Hosted are Enterprise add-ons.
Hobby (free): 100 user challenges, 1,000 standard tool executions, 50 pro tool executions, 1 hosted MCP server, shared tenancy. Growth ($25/mo): 600 challenges, 2,000 standard executions, then $0.01 per standard execution + $0.05 per challenge, $0.50 per pro tool execution, $0.05/hour per hosted MCP server. Enterprise: custom pricing for dedicated tenancy, audit logs, RBAC, SSO/SAML, self-hosted Kubernetes, custom SLA. (Source: arcade.dev/pricing.)

Migration path

Moving MCP traffic from Arcade to Zuplo

Migrating from Arcade is a low-disruption change because MCP clients (Claude Desktop, Claude Code, Cursor, ChatGPT, VS Code, MCP Inspector) point at a gateway URL — swap the URL, repoint OAuth to Zuplo, keep your existing IDP. The biggest practical work is mapping the tool catalog: any custom Python tools you wrote with <code>arcade-mcp-server</code> get re-authored as TypeScript MCP tools (or kept on Arcade temporarily and federated through Zuplo until the renewal cycle consolidates the decision). Pre-built upstream MCP servers — GitHub, Linear, Slack, Notion, Stripe — federate into Zuplo natively with one route per upstream and the gateway's standard OAuth brokering.

Migration phases

Typical MCP cut-over in 2–6 weeks for governed deployments

  1. Inventory tools and upstreams in use

    Catalog the tools your agents call today via Arcade — both pre-built MCP servers (GitHub, Linear, Slack, Notion, Stripe, HubSpot, Salesforce, Microsoft 365) and any custom Arcade-MCP-server tools you authored. Capture which require per-user OAuth and which use shared service accounts.

    2 wksPlan locked

  2. Stand up Zuplo MCP Gateway

    Configure your IDP (Auth0 preset or any OIDC provider), import each upstream MCP server as a route, and define virtual MCP servers per role or team using <code>mcp-capability-filter-inbound</code>. Bind a credential model per route — per-user OAuth where the upstream supports it, shared API key from the encrypted vault where it doesn't. Pin <code>compatibilityDate</code> ≥ 2026-03-01 in <code>zuplo.jsonc</code>.

    2 wksFoundation live

  3. Port custom tools to TypeScript or keep on Arcade temporarily

    Custom Python tools authored with <code>arcade-mcp-server</code> re-author as TypeScript using <code>@zuplo/mcp</code>, or expose the existing Arcade Engine as one more upstream in your Zuplo MCP Gateway until renewal forces a consolidation. For APIs you own, use the MCP Server handler to auto-generate from OpenAPI — zero code in many cases.

    4 wksSide-by-side

  4. Apply guardrails and capability filtering

    Add prompt-injection detection (<code>prompt-injection-outbound</code>) and secret-masking (<code>secret-masking-outbound</code>) to the inbound MCP request path. Curate each virtual server's tool list with <code>mcp-capability-filter-inbound</code> so each team sees only the tools they should call. Test against your Claude, Cursor, and VS Code clients via the MCP Inspector.

    2 wksCut-over done
Read the full migration guide

What our customers say

Trusted by engineering teams at scale

Blockdaemon

90%

Hardware footprint reduction at scale

Read the Blockdaemon case study →

"The move to Zuplo from our existing API Management vendor was easy, taking just over 2 months to switch mission critical systems, and we're saving over 70% on costs."

Ryan Waites

Senior Director, Blockdaemon

Case study →

"Zuplo gives us the flexibility to scale efficiently, ensures security and compliance, and reduces operational complexity so we can focus on building new capabilities."

Daryl Benzel

Staff Software Engineer, Yext

Case study →
AccuWeather

1B+

End users served via Zuplo APIs

Read the AccuWeather case study →

Finsolutia

Hours

To launch MCP server on regulated APIs

Read the Finsolutia case study →

"We didn't touch a line of code, it's just plug and play. The results were very surprising, in just a couple of hours we had a great result and a fully working MCP Server."

Miguel Madeira

CTO & Co-Founder, Finsolutia

Case study →

Trusted for regulated and high-volume workloads

SOC 2 Type II Third-party penetration testing GDPR-aligned 24/7/365 emergency hotline
300+ Global edge locations
Billions API requests served / month
Up to 99.999% Enterprise uptime SLA
<20s Global deploy time

Frequently Asked Questions

Common questions about Zuplo vs Arcade.dev.

Ready to talk to an expert?

Book a call with a solutions architect for a tailored walkthrough — SOC 2 controls, dedicated deployment, AI Gateway, and enterprise support. Or start free and explore the platform yourself.

Book a Call Start for Free