For the complete documentation index, see llms.txt. This page is also available as Markdown.

AI Gateway

This page lists the features, changes and fixes in each release of the P0 AI Gateway.

1.4.1 (2026-11-05)

Security

  • Prevent denial-of-service of OAuth credentials. Previously, any gateway user could register an OAuth credential for a client's MCP connection. This could deny service to the client, since the gateway validates credential ownership against the client's owner. Now, only a client's owner can register an OAuth credential for a client's MCP connection. Anyone else gets the error: "You can only sign in to this MCP server through a client you own. Ask the client's owner to complete this sign-in." The gateway also refuses the registration when the client has no owner.

  • Prevent cross-account use of impersonation credentials. Previously, a malicious container MCP server configured to use service-account impersonation credentials (GCP Service Account Creator) could gain access to any session's impersonation credentials. Now, impersonation credentials work only for the MCP server and sign-in session they were provisioned for, like all other credential sources. Access granted before this release stops working, so agents must request access again after you upgrade.

1.4.0 (2026-11-03)

Changed

  • Chart value for allowed claims in native credentials (Beta). Set nativeCredentialAllowedClaims to the claim names, separated by commas, that a native credential's Claims template may set. In the batteries-included chart this is ai-gateway.nativeCredentialAllowedClaims. The chart applies this to both the Auth Server and the Access Proxy. If you leave it empty (default) the allowed claims are scopes and exp. A list you set replaces that default, so include scopes and exp if you still need them, for example "scopes,exp,groups". The Auth Server doesn't start if the list includes sub, iss, act, aud or may_act. Before, you set NATIVE_CREDENTIAL_ALLOWED_CLAIMS on each service yourself. See Native credentials.

Fixed

  • A new gateway completes its first sync with P0 faster. If the gateway starts before P0 accepts it, such as right after you install it, it now keeps trying for about an hour. It syncs within about 30 seconds of P0 accepting it. Before, it stopped trying after about 5 seconds and waited for its next periodic sync, up to 15 minutes by default.

Security

  • OAuth credentials work only for the client's owner. The gateway now uses a stored OAuth credential only while the user who authorized it is the client's owner. This applies both to credentials that users authorize when they sign in to an MCP server, and to credentials that a client's owner registers for headless agents. The gateway also refuses the credential when the client has no owner, or when it can't confirm the owner with P0. Credentials authorized on 1.3.0-beta or earlier stop working. See the steps below.

  • Access requests are tied to the agent and session that made them. For External and Container MCP servers, a tool call can now use only an access request made in the same session by the same agent. Any other call fails with "Access not granted for this session." Before, the gateway didn't check who made the request. See Requesting access.

Upgrade actions

  1. Plan to authorize OAuth credentials again if they date from 1.3.0-beta or earlier. After the upgrade, users sign in to their MCP servers that use OAuth again, and client owners register headless agent credentials again. Credentials authorized on 1.4.0-beta keep working if the person who authorized them owns the client.

  2. Give every client that uses MCP servers with OAuth an owner. An admin who manages clients sets the owner, and then the owner authorizes the credential. If you change a client's owner later, its credentials stop working once the gateway receives the change from P0, until the new owner authorizes them again.

  3. Upgrade the gateway as described in Upgrade the gateway.

1.4.0-beta (2026-10-28)

Added

  • Vaulted credentials in AWS Secrets Manager (Beta). The gateway can now read a vaulted credential that P0 stores in AWS Secrets Manager, as well as one in Google Secret Manager. It reads the secret as the AWS OIDC role that P0 sets up for the grant, so install AWS OIDC for the account that holds the secret. Before, these credentials failed with "Vaults in aws are not supported yet". MCP servers whose connector runs on AWS, such as ServiceNow and Tableau, need this release. See MCP server.

  • TLS and passwords for your own Valkey. If you use your own Valkey or Redis-compatible cache, the valkey.url in authServer.caches can now use rediss:// for TLS, user:password@ for a password, and /<n> for a database number. Before, the Auth Server ignored all three and connected to database 0 without TLS or a password. The URL, password included, shows in the Auth Server's Deployment, so restrict who can read it. See Gateway chart values.

Changed

  • Valkey 9 in the batteries-included chart. The bundled Valkey now runs valkey/valkey:9-alpine and reads its existing data. After a failed OAuth refresh, the gateway now deletes the stored credential only if no other refresh has replaced it, and this needs Valkey 9.0 or later. If you use your own cache, we recommend Valkey 9.0 or later. On older versions, a credential whose refresh failed stays stored until the user signs in again, an admin revokes it, or it expires. See Valkey.

  • Stricter cache settings. The Auth Server no longer starts if authServer.caches sets a url on a single Valkey tier, which never worked. Set the shared valkey.url instead. It also refuses a valkey.url with a username but no password, a query string, a path other than a database number, or a scheme other than redis://, rediss://, valkey:// or valkeys://. See the steps below.

Fixed

  • Long tool calls cut off by Envoy's stream idle timeout. For the MCP servers you add to the gateway, it now streams each tool call's response from the start and sends a heartbeat every 15 seconds while the tool works. Envoy's stream idle timeout, 5 minutes unless you set gateway.streamIdleTimeout, no longer ends a slow call. Change the interval with accessProxy.sseHeartbeatIntervalMs. Tool call results now arrive as an event stream rather than a single JSON body. See Stream timeouts for long-running calls.

  • Sign-in to SaaS MCP servers that limit client authentication. When the gateway registers itself with a SaaS MCP server's authorization server, it now asks for a client authentication method that the server advertises. Before, it always asked for client_secret_basic, even from servers that accept only other methods, such as LaunchDarkly's, which accepts only none. Servers the gateway already registered with keep their method. See SaaS.

Security

  • Only a client's owner can register its OAuth credentials. When someone registers an OAuth credential for a headless agent, the gateway now checks that the email they signed in with matches the client's owner in P0. Anyone else gets "You can only register credentials for a client you own. Ask the client's owner to complete this registration." A client with no owner, such as one registered before P0 recorded client owners, can't have credentials registered until an admin who manages clients sets its owner. Before, any P0 user could register a credential for any client. See Register an OAuth credential.

  • Cloud credentials encrypted in the cache. The Auth Server now encrypts the AWS and Google Cloud credentials it caches in Valkey, and hashes their cache keys, which used to hold user and role identifiers in plain text. Existing entries are fetched again once after the upgrade. There's nothing to configure.

  • Only the Auth Server can reach the bundled Valkey. The batteries-included chart adds a NetworkPolicy that lets only the Auth Server's pods connect to Valkey. The policy has an effect only if your cluster's network plugin enforces NetworkPolicy. See the steps below.

Before you upgrade

  1. Check authServer.caches if you use your own cache. In the batteries-included chart this is ai-gateway.authServer.caches. Remove any url on a single tier. The Auth Server now uses the scheme and any username and password in valkey.url, so make sure they match your server. If valkey.url ends in /<n> with n other than 0, the Auth Server now uses database n instead of 0, so users sign in to their SaaS MCP servers again and admins register headless agent credentials again. If you set overrides for the caches gcp-wif, aws-wif or gcp-impersonation, rename them to wif.

  2. Let other pods reach the bundled Valkey if they need to. If anything besides the Auth Server connects to the bundled Valkey, add your own NetworkPolicy for pods labeled app: valkey-primary that lists it under from, before you upgrade.

  3. Upgrade the gateway as described in Upgrade the gateway.

If you roll back the batteries-included chart to 1.3.0-beta, Valkey 8 may fail to start because it can't read data written by Valkey 9. Delete the Valkey volume claim data-valkey-primary-0 and its pod, then roll back again. Users sign in again, and admins register headless agent credentials again. See Going back.

1.3.0-beta (2026-10-19)

Added

  • MCP protocol revision 2026-07-28 (Preview). The gateway now accepts MCP clients that use protocol revision 2026-07-28, alongside clients on 2025-11-25. This applies to external HTTP and Docker MCP servers. There's nothing to configure, and existing clients keep working as before.

Changed

  • Bundled Valkey keeps its data across restarts. In the batteries-included chart, Valkey now stores its data on a persistent volume. Before, a restart emptied it. Set valkey.storageClass to an encrypted StorageClass before you upgrade, because the chart won't render without it. See the upgrade steps below.

Upgrade steps

These steps apply to existing gateways that run the bundled Valkey from the batteries-included chart. If you install the gateway chart on its own, or use your own Redis-compatible cache, you don't need to change any values.

  1. Set valkey.storageClass to an encrypted class. The Valkey volume holds users' credentials for SaaS MCP servers. Reuse your postgresql.storageClass only if that class is encrypted. On EKS, use a class with encrypted: "true". The stock gp2 class isn't encrypted. GKE and AKS encrypt disks by default. If you leave it out, the upgrade stops before it changes anything in your cluster.

  2. Upgrade the gateway as described in Upgrade the gateway. The upgrade creates the Valkey volume for you. Valkey is down for up to a minute and comes back empty.

  3. Register headless agent credentials again. Because Valkey comes back empty, the upgrade erases all OAuth credentials. Users sign in to their SaaS MCP servers again, and admins register their headless agent credentials again. See Register an OAuth credential.

Proof-of-concept installs on a cluster with no storage class turn off the bundled Valkey and keep the cache in memory instead. Before you upgrade, set these values:

With the memory cache, the gateway loses all OAuth credentials whenever the Auth Server pod restarts. See Evaluating without a storage class.

If you roll back to 1.2.0-beta, Helm deletes the Valkey volume.

1.2.0-beta (2026-10-15)

Added

  • Token exchange credential source (Preview). The gateway can trade its own token for an upstream access token at any token endpoint that supports RFC 8693. Certain SaaS MCP servers use it to call the service's hosted MCP server as each session's own user. See SaaS.

  • Access token lifetime setting. Set authServer.accessTokenTtlSeconds in the gateway chart to choose how long Auth Server access tokens last. The same lifetime applies to credentials the gateway mints for MCP servers. Leave it empty to keep the default of 1024 seconds. Keep it at 300 or more. The gateway won't start an MCP server whose credential has less than two minutes left. See Gateway chart values.

Changed

  • Workload identity federation settings. If you write MCP server configurations by hand, GCP workload identity federation credentials must now set audience. P0 rejects a GCP credential without it. MCP servers that P0 manages for you need no change.

Fixed

  • Requests with underscores in header names. On clusters with Envoy Gateway v1.1.0 or later, the chart now drops these headers, so Envoy Gateway no longer answers the request with a bare 400 error. This error stopped certain SaaS MCP servers from reading the gateway's signing keys. If you already have a ClientTrafficPolicy that targets the Gateway, it takes priority and the error comes back. Point your policy at one listener with sectionName, and keep withUnderscoresAction: DropHeader.

1.1.0-beta (2026-10-13)

Added

  • Revoke a headless agent's OAuth credentials (Preview). See the OAuth credentials stored for a headless agent's client, and revoke one if it leaks or is no longer needed. Before, such a credential stayed usable until it expired. If you deploy the Auth Server and the gateway separately, upgrade the Auth Server first. See See and revoke credentials.

Changed

  • Renamed Helm values keys. In the gateway chart, agenticAuthServer is now authServer and agenticGatewayServer is now accessProxy. The old keys still work, but if an old key and its new key disagree, the chart refuses to render. Check your values before you run helm upgrade. You no longer need to set accessProxy.authServerUrl, because it points at the release's own Auth Server by default. See Renamed values keys and Upgrade the gateway.

  • JWT bearer requests with client credentials. A JWT bearer token request (RFC 7523) that also includes client credentials is now checked and accepted, where before it was rejected. The agent's identity still comes from the assertion. See JWT-SVID.

Fixed

  • AWS federated sign-in. AWS no longer rejects the gateway's tokens with InvalidIdentityToken ("Couldn't retrieve verification key from your identity provider"). This error broke the AWS and RDS PostgreSQL MCP servers in 1.0.0-beta. See AWS and PostgreSQL on AWS RDS.

1.0.0-beta (2026-10-08)

The first release of the P0 AI Gateway, part of P0 Runtime Access for AI. It sits between your AI agents and the systems they use, and gives agents just-in-time access through P0.

  • MCP servers with just-in-time access. The gateway puts each of your MCP servers at its own path and checks identity on every tool call. Agents get built-in tools to request access through P0, wait for approval, and give access back when done. A tool call fails unless it carries an approved request, so an agent can't skip the request. See Using the P0 AI Gateway and Requesting access.

  • Tool-level access control. Choose how each MCP server decides what an agent may do. Native mode leaves it to the upstream system's own permissions. Tool mode grants named tools. Annotation mode grants groups of read-only, write or destructive tools. Tool and Annotation modes are in preview. See Access control method.

  • User and agent identity together. A person signs in through your identity provider and approves a consent page that names both the MCP server and the agent. The token they get ties the user to the agent and works for one server only. MCP clients such as Claude Code find the sign-in endpoints on their own. See Connect an MCP client.

  • Sign-in options. People can sign in with any OIDC provider, such as Google, Okta, Entra ID or Ping, and you can limit sign-in to your email domains. If you can't create an OAuth client, use password sign-in through Google Cloud Identity Platform. See Gateway chart values.

  • Headless and workload agents. Agents with no person at the keyboard can sign in with a signed JWT from an issuer you trust, including SPIFFE JWT-SVIDs. Add each issuer to P0 as an identity provider. See JWT-SVID and Identity provider.

  • Short-lived credentials, no standing secrets. For each approved request, the gateway gets a short-lived credential for the upstream system, and the agent never sees it. Sources include AWS IAM federation, GCP workload identity federation, OAuth, secrets P0 stores in your vault, and RDS IAM. See MCP server.

  • Predefined MCP servers. Connect agents to the AWS, GCP (BigQuery, Compute Engine, Cloud Monitoring, Cloud Storage and IAM), Google Drive, and PostgreSQL on AWS RDS. The PostgreSQL server is in preview.

  • Custom MCP servers. Point the gateway at an MCP server's URL, or have it run a Docker image for you. Each container serves one user, session and request, and stops when its credential expires. Container servers are in preview. See Custom.

  • SaaS MCP servers that use OAuth. The gateway signs each person in to the upstream service with their own account, then stores and refreshes the credential for them. For headless agents, an admin can register one OAuth credential for an agent. See SaaS and OAuth servers from a headless agent.

  • Agent-to-agent calls. Put an agent host behind the gateway so calls between agents are checked and logged. An agent requests access to a specific agent, and the called agent can act for the original user. See A2A bridge.

  • Activity logs. The gateway records every tool call and agent call with the user, agent, session and request, and exports them over OpenTelemetry. P0 reads them from Google Cloud Logging or VictoriaLogs to fill its event log. See Event log and Log provider.

  • Access ends right away. When an agent gives access back, or P0 revokes a request or it expires, the agent loses access at once. If P0 is briefly unreachable, agents keep the access they already have. See Session behavior.

  • Runs in your Kubernetes cluster. Install the gateway with a Helm chart, from the P0 console or by hand. It works with any Gateway API class, such as Envoy Gateway or GKE, and supports private image registries and proxies that inspect TLS. A batteries-included chart also installs what the gateway needs to run. See Deploying the gateway.

Last updated