Skip to main content

Changelog

November 6, 2026​

Choose who can use a skill​

A skill can now be Private to you, shared with specific teams, or shared with everyone in your organization. Choose the audience when you create the skill, or change it later under Access. The registry groups skills into Private and Shared, with a toggle between Available to me and Managed by me, and people only see the skills shared with them.

A skill also has to reach everyone the persona reaches: a skill scoped to two teams can go on a persona scoped to those teams, but not on a company-wide one. The skill picker only offers skills that qualify and tells you how many it left out.

See Who can use a skill.

Send a private pool's traffic through your proxy​

A private pool can now sit behind your corporate HTTP, HTTPS or SOCKS5 proxy. Point the pool at it under Configuration → Outbound Network, then choose which traffic uses it: the Operator's Control plane connection, Sign-in and token checks, and SIEM export. Each MCP server picks its own Network path — through the proxy, or straight from the cluster for internal services the proxy can't reach. Always direct keeps named hosts, domains and CIDR ranges off the proxy, and Fail open stops a broken proxy cutting the pool off from management.

See Send pool traffic through your proxy.

Skills are ready to attach as soon as you create them​

Skills no longer have Draft and Ready states. Previously a new skill had to be marked Ready before any persona could use it; now you create a skill and attach it. Editing the instructions, references and scripts of an attached skill works as before — agents pick up changes the next time they load it.

See Skill Registry.

October 30, 2026​

One URL for every agent​

Everyone in your organization can now add a single URL to their AI agent and ask for the work they need. The gateway finds the right persona or MCP server and opens only the tools that person is allowed to use. Administrators choose what the URL offers from Operations → Agent Endpoint, and everyone finds the URL, setup steps for their agent and ideas for what to ask on My Agent Access.

See Agent Endpoint and My Agent Access.

Connect apps from the portal​

Connect your own account to an app before your agent needs it. MCP servers that act as you show a Connect button in the MCP Registry and on their Overview tab, and a persona's Connections section has Connect all.

See Connect your apps.

Simpler OAuth setup​

The OAuth form now reads the app's published OAuth details and fills itself in. New MCP servers use the AI Gateway's latest callback, and the gateway can identify itself to the app with a client ID metadata document or register itself, so many apps need no OAuth app from your IT team. You can still use your own client, or the v1 callback for an OAuth app that is already registered.

See OAuth 2.1 Authorization.

Cross App Access​

Let people reach an app through the trust your identity provider, such as Okta, already has with it. Set up a Cross App Connection under Settings → Auth Profiles, choose how accounts are matched, and choose Cross App Access as the MCP server's app authentication.

See Cross App Access.

Exclude exact values from custom detectors​

List known safe values, such as test IDs, that a custom regex detector should never report. The detector builder now also checks your pattern as you type and tells you when the gateway can't run it, instead of the detector being skipped.

See Custom and composite detectors.

Field name format for OTLP destinations​

Send logs to destinations that reject dots in field names, such as Azure Data Explorer. Choose Underscore (service_name) under Field name format on an OTLP destination.

See Field names for OTLP destinations.

DLP engine metrics​

The DLP engine now publishes Prometheus metrics, so you can watch scans, findings, blocks and redactions in your own monitoring tools, or export them through SIEM.

See Monitoring and metrics.

October 16, 2026​

OAuth 2.1 Client Assertion​

Connect MCP servers and APIs to apps whose identity provider requires a certificate instead of a client secret, such as Microsoft Entra ID. Choose OAuth 2.1 Client Assertion under App Authentication and provide your private key with the certificate or its thumbprint. The gateway signs in on its own, and the key is stored encrypted and never shown again.

See OAuth 2.1 Client Assertion.

Custom endpoints for OpenAI and Anthropic​

OpenAI and Anthropic entries in the LLM Registry can now send requests to an endpoint you choose, such as your company's own proxy in front of the provider, instead of the provider's default. Choose Custom under Endpoint when you add the provider, or change it later from the entry's Configuration tab.

See LLM Registry terminology.

October 8, 2026​

Agent Persona Builder​

Describe what an agent needs to do in plain language, and the AI Assistant drafts a persona with the right apps and tools. Before you save, you see everything the agent will be able to do, and anything that changes data is called out. Choose Read only to make sure an agent can never change, send or delete anything.

See Agent personas.

LLM Continuity​

Keep your AI working when a provider has an outage. Add backup providers to an AI connection, and requests are answered by a backup when the main provider is down or overloaded. Models your company has blocked are never used as backups. Each persona can follow your settings, narrow them or turn continuity off.

See LLM Continuity.

Budgets & AI Spend​

Set spending limits per person, with team limits and named exceptions, plus shared team pools and agent budgets. Use Check a person to see which limit applies to someone and why. Monitor mode shows who would have been stopped, without blocking anyone. AI Spend reflects your actual usage at your contracted rates.

See Budgets and AI Spend.

Isolation Tiers (private-cloud gateways)​

Split a gateway into priority tiers, each with its own dedicated capacity, so critical agents aren't slowed down by heavy workloads elsewhere. You can turn tiers on or off and change the default tier without interrupting live traffic.

See Isolation tiers.