New from Zuplo - July 2026

July 2026 takes the MCP Gateway to general availability and adds role-based tool access and Cross-App Access support. New Audit Logs capture a structured, queryable record of every request to your API, and observability is now included on every plan, with a new analytics Overview page. The AI Gateway speaks the Anthropic API natively, and the Dev Portal adds Microsoft Entra ID authentication and guided MCP connection setup.

MCP Gateway Is Now Generally Available

The MCP Gateway, announced in beta last month, is now generally available. The beta label is gone across the product and docs, and July brought a set of access-control capabilities aimed squarely at enterprise rollouts.

Role-based tool access: The MCP Capability Filter policy now supports per-caller access control. In rolesAndGroups mode, the policy matches the caller’s role and group claims against the roles and groups you assign to each capability, so one virtual server can expose different tools, prompts, and resources to different users. Capabilities marked public: true stay visible to everyone, unclassified capabilities are exposed to no one until you classify them, and everything fails closed. A function mode accepts a custom resolver module for authorization logic beyond roles and groups. The MCP Gateway wizard’s Tools step gains matching Roles & groups controls, so you can configure all of this without editing JSON. See Per-Role Tool Catalogs for Your MCP Server for a worked example.

Cross-App Access (ID-JAG): The MCP Gateway now supports the Identity Assertion JWT Authorization Grant, the emerging standard behind Okta’s Cross-App Access, across the generic MCP OAuth policy and all identity provider variants (Auth0, Okta, Microsoft Entra, Google, Amazon Cognito, Clerk, WorkOS, Keycloak, Logto, OneLogin, and Ping). The MCP Token Exchange policy adds an id-jag auth mode that exchanges the caller’s identity token for an upstream token without another round of user consent, so agents move between apps under one enterprise identity. See the MCP Token Exchange documentation and Decouple Agent Auth From Your MCP Server, which includes a video walkthrough.

Also released this month:

• API key upstream auth in the wizard: The Outbound Auth step offers an API key mode that scaffolds a Set Upstream API Key policy sending $env()-sourced secrets to the upstream server
• Visual policy designer: The policy editor gains a Designer tab with a visual capability-filter editor that probes the live upstream catalog
• Consent prompt control: A prompt option on upstream OAuth configuration stops providers like Microsoft Entra from forcing a consent screen on every connection
• Configured scopes on the wire: Scopes set on the token exchange policy are now sent on upstream authorization requests, making refresh-token behavior like offline_access configurable

Get started with the MCP Gateway documentation.

Audit Logs

Zuplo now records a complete audit trail of your API with a single policy: the gateway captures a structured event for every request, stores the events for you, and makes them available in the Zuplo Portal and through the Zuplo API. No database, logging service, or other infrastructure of your own to run.

Add the Audit Logs policy to any route and every request produces a CloudEvents 1.0 event capturing the actor, HTTP method and path, response status, caller IP address, and geolocation:

{
  "name": "audit-logs",
  "policyType": "audit-logs",
  "handler": {
    "export": "AuditLogInboundPolicy",
    "module": "$import(@zuplo/runtime)"
  }
}

The policy’s include options let you turn off individual fields (query parameters, user identity, IP address, geolocation) to satisfy your PII policies, and samplingRate captures a fraction of requests on high-volume routes. You can also emit your own domain events (for example, “account deleted”) from handlers and custom policies with the static AuditLogInboundPolicy.log() method. Audit logging is best-effort by design: it never blocks or fails the request.

In the Portal, a new Audit Logs service under the Services tab lets you search and filter events by time range, event type, actor, and subject, see top event types and most active actors, and inspect any event’s full payload. MCP Gateway traffic is audited too, so tool calls are attributed to the end users who made them.

Through the API, GET /v1/audit-logs/{bucketId}/events searches events and GET /v1/audit-logs/{bucketId}/stats returns top event types and actors. Both operations are also exposed as tools on the Zuplo MCP server, so you can ask an agent what changed. Export to your SIEM from the same endpoints.

Audit Logs are an Enterprise add-on. See the Audit Logs documentation to get started, and the Audit Logs policy reference for all options.

Observability Included on Every Plan

Analytics, logs, and traces are now part of every Zuplo plan, including Free. Plans differ by how much history you keep, not by which tools you get:

Custom retention is available on Enterprise plans, and every new account gets an automatic 14-day trial with extended windows. The in-portal OpenTelemetry tracing views launched in June are included under the same windows, so every account can now drill into span waterfalls. See Analytics access and data retention for details.

The Analytics section also opens on a new Overview page: a product-tailored summary of key metrics, traffic, locations, latency, cache performance, and top consumers. API Gateway projects lead with requests, server errors, and latency; AI Gateway projects lead with tokens, estimated cost, and model usage. The AI analytics section is now called AI & LLMs, and its report also renders on AI Gateway app and team pages with one-click drill-in to full analytics. A Last 30 days preset joins the time-range picker.

AI Gateway: Native Anthropic API Support

The AI Gateway’s /v1/messages endpoint is now a native Anthropic passthrough instead of a translation to an OpenAI-compatible shape. Request bodies are forwarded verbatim, new Anthropic fields and client headers like anthropic-beta pass straight through, and Anthropic’s own error format is returned. Streaming responses reach your client untouched while usage is metered from a background copy. The AI firewall, semantic cache, and Opik and Galileo tracing all understand native Anthropic requests end to end, including streaming.

Also released this month:

• Semantic cache TTL: A new Cache TTL setting (next to Semantic Tolerance) controls how long cached responses are served, up to 30 days
• Standard cache headers: Cache outcomes are reported in the RFC 9211 Cache-Status response header, including the match similarity on hits, and the app Test page shows a matching cache hit badge
• Cache hit rate fixed in analytics: Semantic cache hits now emit analytics events (recorded with zero cost and zero tokens), so the dashboard’s cache hit rate reflects reality
• OpenAI gpt-5 and o-series support: The gateway now sends max_completion_tokens to OpenAI, so client requests using max_tokens keep working against models that reject it
• Provider-name routing: Route requests to a provider by its display name using the providerName/model format, alongside a redesigned AI Providers page
• Redeploy on demand: A Redeploy Gateway button on Settings rolls your gateway onto the latest runtime whenever you choose

Runtime Improvements

AWS Upstream Authentication

Calling AWS from your gateway no longer requires hand-rolled request signing. Two new inbound policies resolve AWS credentials and register them on the request context: Upstream AWS Service Auth uses an IAM access key pair (with optional STS AssumeRole), and Upstream AWS Federated Auth exchanges Zuplo’s ambient OIDC identity for temporary credentials via STS AssumeRoleWithWebIdentity, so no AWS secrets are stored anywhere. The AWS Lambda handler picks up the resolved credentials automatically:

{
  "/lambda/{path}*": {
    "x-zuplo-route": {
      "handler": {
        "export": "awsLambdaHandler",
        "module": "$import(@zuplo/runtime)",
        "options": { "region": "us-east-1", "functionName": "my-function" }
      },
      "policies": { "inbound": ["upstream-aws-federated-auth"] }
    }
  }
}

For custom code, the new @zuplo/runtime/aws module exposes AwsClient, AwsV4Signer, and credential providers, so a handler can sign any AWS request with signRequest(). Both policies are in beta for Enterprise accounts. See the Upstream AWS Federated Auth documentation and Upstream AWS Service Auth documentation.

Claims-Based Authorization Rules

The new Require User Claims policy authorizes requests against the authenticated user’s claims, separating authorization from authentication: any auth policy (JWT, API key, mTLS) proves who the caller is, then a declarative rule decides what they may call. Rules combine and/or logic with exact (eq), allowlist (in), and prefix (startsWith) checks, and reach nested or namespaced claims:

{
  "rule": {
    "or": [
      { "claim": "sub", "in": ["ci-deployer@my-project.iam.gserviceaccount.com"] },
      { "claim": "['https://example.com/roles']", "eq": "admin" }
    ]
  }
}

Checks fail closed: no user means 401, a failed rule means 403, and denied responses never echo claim values. See the Require User Claims documentation.

Also released this month:

• WebSockets are generally available: The WebSocket pipeline handler is out of beta. With no message policies configured it becomes a transparent passthrough that stays out of the data path, upstream connection failures return proper HTTP errors, and a new onOpen hook runs your code once when both sides of the connection are established. See the WebSocket Pipeline Handler documentation
• Upstream request timeouts: The URL Forward and URL Rewrite handlers accept a requestTimeoutSeconds option; when the upstream exceeds it, the gateway returns 504 Gateway Timeout instead of waiting indefinitely
• Data Loss Prevention supersedes Secret Masking: The Secret Masking policy is deprecated (it keeps working). The Data Loss Prevention policy covers the same secret patterns as part of a catalog of 60+ built-in entities, adds block/log actions and per-pattern confidence scoring, and the Secret Masking docs include a migration guide
• Upstream Zuplo JWT claims mapping: The Upstream Zuplo JWT policy gains subject and claimsFromUser options for minting tokens with claims mapped from the authenticated user
• Stable trace identity: OpenTelemetry spans now carry consistent zuplo.policy.type, zuplo.policy.name, zuplo.policy.direction, and zuplo.handler.type attributes, so you can filter traces by policy in any backend
• Fixes: Errors thrown in zuplo.runtime.ts initialization are surfaced as a Gateway Initialization Error response instead of being swallowed, and the rate limit policy’s async mode no longer logs unhandled promise rejections when the rate-limit service is unreachable

Dev Portal Updates

Microsoft Entra ID Authentication

The Dev Portal adds a first-class Entra ID auth provider. Unlike generic OpenID configuration, it handles Entra’s multi-tenant issuer templates, resolves the concrete tenant from each token, and falls back to preferred_username when the mutable email claim is absent. The immutable oid and tid claims are surfaced on the user profile for reliable identity checks:

// zudoku.config.ts
const config: ZudokuConfig = {
  authentication: {
    type: "entra",
    clientId: "<your-application-client-id>",
    tenantId: "common", // or your tenant ID, "organizations", "consumers"
  },
};

CIAM tenants are supported via an issuer override. See the Entra ID setup guide.

Guided MCP Connection Setup

The MCP endpoint page in the Dev Portal replaces the config tabs with a guided onboarding flow: a prominent copyable server URL (with an authentication hint when the server requires a key), a visual client picker, and numbered per-client setup steps for Claude, Codex, Cursor, VS Code, and generic clients. Cursor and VS Code get one-click install buttons using their native deep links; the buttons only appear when no API key header is needed, so secrets are never embedded in a link.

Generated install snippets also now use the MCP server’s real name from the route’s MCP handler options, so the server name the docs tell users to install matches what clients see when they connect. See the Documenting MCP Servers guide.

Server-Side Rendering

The Zuplo-hosted Dev Portal can now render server-side per deployment. Set devPortal.serverSideRendering: true in zuplo.jsonc and protected routes are enforced on the server: protected content leaves the public bundle entirely, and unauthenticated requests cannot fetch it. SSR applies to preview and production deployments and is an Enterprise add-on. See the server-side rendering guide.

Also released this month:

• Search covers protected pages: In statically rendered portals, pages behind protectedRoutes now appear in search results instead of silently missing, and builds log a warning listing any routes excluded from the index
• Clearer pricing cards: Monetization pricing cards now distinguish pay-as-you-go from hard caps, so unlimited usage no longer displays as a “0 / month” quota, and tiered price breakdowns render volume and graduated billing modes correctly
• Multiple subscriptions: When enabled for the environment, the pricing page lets customers subscribe to several plans at once (see Monetization Updates below)
• Slimmer installs: The Dev Portal CLI dropped its error-reporting dependency, removing 20 packages from every install and fixing a build failure mode caused by npm hoisting
• Fixes: The pricing page no longer crashes after logout, and top navigation links honor a configured target such as _blank

Monetization Updates

July’s monetization changes span the plan editor, billing, and the runtime:

• Free trial phases: A plan’s first phase can be marked as a free trial in the plan editor; returning subscribers who already used their trial start on the paid phase automatically
• Stripe Tax: Enable and configure tax collection from the Payment Provider page: toggle Stripe Tax, set your supplier country, choose inclusive or exclusive pricing, and pick a tax code
• Multiple active subscriptions: Consumers can hold several subscriptions at once, enabled per environment on the Payment Provider page
• Usage history: Subscription pages now show historical usage for past billing periods, and the invoice status filter searches all invoices instead of the current page
• Reliable metering: Usage events are now retried with backoff and deduplicated by event ID, so a transient network blip no longer drops billable usage, and the new MonetizationInboundPolicy.flushMeters() API submits accumulated runtime meters immediately instead of at request end

Metering MCP traffic? See Charge Agents for MCP Tool Calls for a pattern that meters tool calls while protocol chatter stays free.

Portal & CLI Updates

Redesigned Route Test Console

The route test console was rebuilt around a two-pane layout: an editable URL bar, a tabbed request builder with body editors (JSON, GraphQL, form, or text), headers, and API-key or bearer auth, and a response panel with the status line, timing, and size. On top of the redesign:

• Request history: Reopen any earlier request with its full setup restored
• Endpoints panel: A collapsible panel lists every route in your API, so you can switch what you’re testing without leaving the console
• MCP request bodies: An MCP body type (auto-selected on MCP server routes) offers one-click standard calls like listing and calling tools, no hand-written JSON-RPC required
• Build-aware testing: When you test while a deploy is in flight, the console waits for the new build to go live before sending, so you always test the code you just saved

Services Navigation Redesign

The Services tab now opens with a sidebar listing every service and its pages (API Keys, Monetization, Audit Logs, and Metering where enabled), with one environment switcher for the whole section. The API Key Service page also got a full redesign with a cleaner consumer list, tidier key controls, dark mode support, and an API Keys tab in the consumer dialog listing every key’s creation and expiry dates.

Also released this month:

• Route Designer: API metadata and files: Edit your API’s title, version, description, contact, and license from the top of the routes list, switch between OpenAPI files or add a new one in place, and imported OpenAPI files now save immediately
• Route Designer: grouping and search: Group routes by handler type or by upstream target, and find any route from the ⌘K palette by path, method, handler, or policy. Read-only environments show a banner offering to switch to Development or check out the branch
• Environment variables in bulk: The create dialog supports multiple rows with a per-row secret toggle, and pasting a .env file fills them in automatically
• Account-level custom domains: Add a custom domain from the account’s Custom Domains page and pick the target project in the dialog
• Programmatic project provisioning: The Developer API’s project create endpoints now support API-key authentication and account-admin callers, so you can provision projects from scripts and CI
• CLI: local mTLS certificates: zuplo dev auto-detects client certificates in .zuplo-local/mtls (or takes --mtls-certs <dir>), so managed-dedicated projects can develop against mTLS backends locally
• Source control: Connecting a repository whose default branch requires pull requests now pushes to a setup branch and opens the PR for you, and pushes no longer fail for GitHub users with private email addresses
• Fixes: Analytics chart timestamps are correct in every time zone, favoriting a project no longer scrambles the projects list, the MCP settings editor no longer discards tool edits, and your last-accessed project follows you across devices

Documentation Updates

• Zuplo Audit Logs: New guide covering the Audit Log policy, custom events, the Portal viewer, the query API, and SIEM export
• Dev Portal Server-Side Rendering: New guide for enabling SSR and protecting content server-side
• Local Development Debugging: Rewritten around attaching to the workerd runtime, with one-click VS Code setup and WebStorm/IntelliJ steps
• Securing Backend mTLS: Expanded with client certificate requirements, generation and validation steps, and the new local development workflow
• Self-Hosted Overview: Rewritten with how the hybrid model works, lifecycle responsibilities, and infrastructure requirements
• Rate Limiting: New section on sub-minute rate limit windows and the accuracy floor imposed by cross-region synchronization
• Analytics Access & Data Retention: New reference for per-plan observability access, retention windows, and the 14-day trial