Release Notes

Agent conversations no longer fail when they exceed the model’s context limit.

The ServiceNow connector now lets you choose which field on an incident tracks its state, so completing or transitioning a task updates the column your instance actually uses instead of a fixed one.

The Pylon connector can now fetch the full reply and internal-note thread on an issue as its own tool, with pagination for long threads.

A new help button explains the graph’s controls, and keyboard shortcuts let you clear the current selection or zoom in and out.

This release includes general reliability and stability improvements to how agent activity is collected and stored.

Token-optimized output now applies to a remote MCP server, not only to the built-in connectors; previously the setting had no effect there.

The ServiceNow (dynamic) connector now starts reliably against an instance with many tables or columns, instead of timing out during setup.

The ServiceNow (dynamic) connector’s catalog item form lookup no longer fails with a permission error for a user who can already see the item.

A record the ServiceNow (dynamic) connector creates is now always attributed to the person who made the request, instead of sometimes ending up unattributed or owned by a shared account.

Panning the map and selecting a node near a connection line now works reliably; the lines no longer intercept the click.

A conversation whose sender could not be verified now shows as unattributed instead of displaying the wrong user.

This release includes general reliability and stability improvements to how agent activity is collected and stored.

The new Agent Network page, available as a preview under Agents, shows a live map of your agents and the MCP servers and LLM providers they call, with each part’s health, errors, token usage, and cost, so you can see at a glance which piece has a problem.

The new ServiceNow (dynamic) connector generates its tools from your own ServiceNow instance, so agents work with your actual tables, fields, and choice lists instead of a generic fixed set; the existing ServiceNow connector is unchanged.

The Audit Log page is now available to every organization; Redpanda previously had to turn it on for you.

Transcript recording is now on by default when you create an LLM provider: Record inputs and Record outputs start enabled, either can still be turned off, and existing providers are unchanged.

A conversation that fails now shows a clear notice with recovery actions and keeps the messages already exchanged, with the technical cause available under Error details.

This release includes general security and stability improvements.

Long responses from an agent using a Claude model on Amazon Bedrock are no longer silently cut off partway; affected agents now return the complete answer and may use more output tokens than before.

Saving an agent with a model that its LLM provider does not offer now fails with a clear error, instead of the agent breaking later when someone uses it.

The Pylon connector now returns the real text, author, and time of each message on an issue instead of blank entries, and messages are ordered oldest first.

The ServiceNow connector now submits a service catalog request or a record producer form, and returns the requested item number an employee tracks it by.

The ServiceNow table schema now reports whether a field is read-only or filtered by a reference qualifier, so an agent knows a write would be dropped before it makes it.

This release includes general reliability and stability improvements to how agent activity is collected and stored.

Pause and resume a cron trigger without deleting it, so its schedule and run history stay intact while it is temporarily turned off.

The Slack connector now has a read-channel tool, so an agent can read a channel’s message history instead of only a thread it was handed directly.

Search the ServiceNow service catalog by name or by what an item’s request form asks for, instead of working from a hardcoded list of items.

Posting a message through the Slack connector now returns its timestamp and channel, so an agent can react to or reply in a thread under its own post.

Page content is now capped at a consistent maximum width across the product, so forms and lists no longer stretch edge to edge on wide screens.

Transitioning a ServiceNow incident now uses the state name, such as Resolved or Canceled, so a customized instance’s renumbered states can no longer send it to the wrong one.

The ServiceNow connector now verifies what a write actually saved instead of trusting a success response, and correctly saves explicit values such as zero or a negative state.

Streaming from an OpenAI-compatible provider that sends keepalive comments, such as OpenRouter, no longer fails immediately, and a transient decode or connection failure now retries automatically instead of ending the agent’s turn.

A transcript for an agent that delegates to a subagent no longer splits into a separate, unreachable record or shows duplicated or fabricated messages.

Activity now keeps updating after a brief database restart, instead of getting stuck until the service is restarted by hand.

The rpk ai proxy command now routes every service in a private cluster through the tunnel, including Redpanda Console, instead of only two of them.

The Setup tab now shows the install and run commands for the language and framework you picked, and the install command lists every dependency the sample code imports, so you can copy it and the sample runs without adding packages by hand.

When Redpanda cannot list a server’s tools, the Overview tab now shows why, with a button to retry, instead of showing the server with an empty tool list. A managed server also keeps showing the tools from the last successful check beside the warning.

The rpk ai auth login command no longer skips sign-in when a session is already on disk, so a revoked session can no longer leave you with a command that insists you are already logged in. Every run is a full interactive sign-in, so a script that needs a conditional login can check rpk ai auth status -o json first.

Connecting your own third-party account for an MCP server that requires one now works for any user with access to the server, because Redpanda ships the permission for it by default, scoped to each user’s own connections. It previously failed with a permission error until an administrator added a policy by hand.

The command the Setup tab gives you to check your setup now calls each provider’s own API instead of its OpenAI-compatible endpoint, so the token counts it reports match the sample program, and a reasoning model returns a real answer instead of an empty response.

Revoking a client secret now removes it from the environment variables shown in the Setup tab, which used to keep offering a value that no longer works.

Signing in with rpk ai auth login now works on a machine with no browser, such as an SSH session or a container: the sign-in URL is printed for you to open on another device instead of the login failing.

The new Audit Log page, available as a preview, lists who did what in your ADP environment, with filters for actor, resource, subsystem, outcome, and time range, and a detail view that shows the permission decision behind an event and any configuration it changed.

A tool call is labeled with the name the model called, which includes the MCP server it came from, and its duration covers the whole tool execution.

Tool calls in a transcript now show the arguments sent and the result returned, tool calls that do not go through an MCP server such as delegation to a subagent now appear at all, and an MCP tool call is no longer listed twice.

The model catalog no longer lists models the providers have retired: Claude Opus 4.1, GPT-5.2 Instant, and GPT-5.3 Chat Latest.

The home page keeps its layout while the Budget governance and Top spenders panels load, so content below them no longer shifts.

This release includes general security and stability improvements.

The active and expired credential counts on a self-managed agent update as credentials expire, without reloading the page.

Sentry is available as a provider preset. Selecting it pre-fills the Sentry endpoints and read-only scopes, so connecting Sentry needs only your client ID and secret, and users can then connect their own Sentry accounts for the Sentry MCP server.

The Workday MCP server adds a read-only list_reference_values tool that returns the codes Workday expects for a field, for example office locations or citizenship statuses, so agents enter real codes instead of guesses.

Set custom per-model pricing in dollars per million tokens with the repeatable --pricing flag on rpk ai llm create and rpk ai llm update, so spend reports use your negotiated rates. Hand-written pricing JSON keeps working.

Tool lists on the MCP server pages are split into pages with a selectable page size, so servers with hundreds of tools stay quick to browse.

Custom date ranges keep the exact window you picked, and the date and time pickers behave the same on every page.

Lists, charts, and pagination controls fit small screens, so the UI works on a phone.

The sign-in page remembers which login method you last used, and tells you when you pick a different method than your account uses.

The provider page shows its connection status in the page header, matching other detail pages.

This release includes general security and stability improvements.

Traffic charts on the provider page and Home no longer render at the wrong size.

This release includes general stability improvements.

This release includes general security and stability improvements.

The provider Overview tab shows which agents and users call which models, with requests, errors, and spend per caller, over a preset period or an exact window.

All provider settings are on one Settings tab. Old edit-page links still work.

The SQL MCP server can return rows as named fields, which a data policy can target column by column, so a policy can mask a salary column and leave the rest readable. Set row_format to ROW_FORMAT_OBJECT. Positional rows stay the default.

Report on an exact UTC window, to the minute, not only a relative period such as the last 7 days.

Spending keys on each agent’s permanent identifier, so reusing a deleted agent’s name no longer mixes the two agents' costs. Past periods are corrected too.

The top spenders card counts agents that shared a name as one, so its top five are the real top five.

Managed MCP servers are listed on the connections they use, with any scopes those servers need that the connection is missing.

Code blocks are readable in dark mode, and opening a menu no longer shifts the page.

The self-managed agent Setup tab is corrected: the X-Redpanda-Genai-Conversation step is gone, conversations group on gen_ai.conversation.id, the Java sample sends its token request, and the Vercel AI SDK, CrewAI, and Mastra samples match current SDKs. Take the current version of any sample you copied.

The Data Policies tab now counts a server’s tools by protection state: protected for a tool a data policy covers, findings for a tool with no policy that returns fields whose names look sensitive, and no policy for the rest. Every tool falls in exactly one group. The tool lists on Data Policies and the MCP server Inspector tab are also searchable, and long tool names wrap in full instead of being cut off, so tools whose names differ only at the end are easier to tell apart.

The agent Inspector tab is now called Playground.

Editing a provider’s connection settings in the ADP UI, for example pointing it at a different API key secret, reported success and kept the old value. Those edits now save. The display name and model selection were unaffected, as were changes made with the rpk ai plugin or the API. Re-open any provider whose connection settings you edited in the UI and confirm it uses the secret you intended.

The agent Playground now tells you when an agent has no subagents attached instead of opening an empty panel.

All granted scopes are readable on provider and client detail pages. A long scope ran outside its popup, and a long list of scopes ran off the bottom of the screen with no way to scroll, so not every scope was visible.

Side panels close when you click outside them.

This release includes general security and stability improvements.

The new Access page, available as a preview, controls what people can do with the resources in your ADP environment. Write your own permit and forbid rules on Policies, or start from a Templates preset: Read only, Sandboxed to use agents and MCP servers without changing them, Standard to also manage resources, or Full access. Roles shows what each of your organization’s roles can do, and System policies shows who already has access through those roles.

The agent prompt editor now formats your prompt as you write it. Live mode renders headings, lists, and emphasis, and shows the raw Markdown only on the line your cursor is on. Switch between Live, Markdown, and Preview without losing your place, or start from a template for a support agent, data analyst, coding assistant, incident responder, or research coordinator.

The agent Setup tab splits MCP servers into Managed, run by Redpanda, and Self-managed, run by you, and shows each managed server’s logo, so a long list is easier to scan.

The app opens faster on a first visit.

Claude can now connect to an MCP server on ADP with a Client ID Metadata Document (CIMD). Retry the connection from Claude, with nothing to change on your side. Connecting failed with unable to retrieve client by id because ADP refused an entire document over a single OAuth grant type that the document lists and ADP does not issue. ADP now ignores those grant types and honors the rest of the document, for any client, not only Claude.

The create-agent wizard previewed a self-managed agent’s client ID in a format that rpk ai agent credential create no longer returns. Only the preview was wrong, so credentials you already issued are unaffected.

Notifications are readable in dark mode instead of washing out against a dark background, and they now appear at the top right of the screen.

An MCP client can now reach a protected server with a Client ID Metadata Document (CIMD), the MCP spec’s successor to Dynamic Client Registration. The app hosts its own metadata at an https URL and presents that URL as its client ID, so nobody creates a client by hand. Turn on Accept client metadata documents (CIMD) on the Integrations setup page, choose whether to accept documents from Any domain or Only these domains, and list the MCP servers those clients can reach.

GPT-5.6 Sol, Terra, and Luna are now in the Amazon Bedrock catalog. Enable openai.gpt-5.6-sol, openai.gpt-5.6-terra, or openai.gpt-5.6-luna on a Bedrock provider.

This release includes general security and stability improvements.

Creating an LLM provider now works on Google Cloud environments. Saving the provider’s API key failed because the form attached an internal label that Google Secret Manager rejects.

Cost reporting overstated GPT-5.6 Luna and Terra spend. OpenAI reduced the rates for both models on July 30, so reported cost now matches the current rates. Figures recorded before this release keep their earlier, higher numbers.

The server details page groups setup, tools, and connection details more clearly, and tool schemas stay readable when you expand a nested object.

Switching to another environment now takes you to the home page of that environment, instead of leaving you on a page that belongs to the previous one.

This release includes general security and stability improvements.

An MCP client that registers itself with Dynamic Client Registration (DCR, RFC 7591) can now connect to a server on ADP. Claude.ai, Claude Code, and Cursor were rejected during registration, and the error named the MCP server’s vendor rather than the real cause.

Confirmation dialogs now place their buttons in the same order as the rest of the product.

Connecting a remote MCP server that uses User OAuth now takes a URL and one sign-in per person, with no app created by hand in a vendor console. Select Set up automatically and ADP reads the server’s OAuth configuration and registers itself as a client using Dynamic Client Registration (DCR, RFC 7591). For a server that does not support DCR, you can still enter a client ID and secret yourself.

Claude Opus 5, Gemini 3.6 Flash, and Gemini 3.5 Flash Lite are now in the model catalog. Enable claude-opus-5 on an Anthropic provider, or gemini-3.6-flash and gemini-3.5-flash-lite on a Google provider. Claude Opus 5 is also in the Amazon Bedrock catalog.

The rpk ai connection commands bring your own OAuth connections to the terminal, matching what My Connections shows in the UI. Run rpk ai connection list to see the services you are signed in to, and rpk ai connection revoke <provider> to disconnect your own account from one.

Creating an LLM provider is now a single form: paste model IDs one per line, and create the provider’s API key secret without leaving the page. For Amazon Bedrock you can browse the catalog by model family and enable or disable models in bulk.

OpenAI MCP servers now transcribe audio files with gpt-transcribe instead of whisper-1. Servers that set a transcription model explicitly keep it.

When the inspector cannot reach a server, it now explains the failure in readable text and offers Try again and Edit server. The raw response stays available in a collapsed panel for network debugging.

Remote MCP servers set to OAuth (Service Account) now attach the service account’s access token to their upstream requests. They previously sent no credentials at all, so a server requiring authentication rejected every request.

Cost reporting understated large Anthropic requests: above 200,000 input tokens, Anthropic charges a higher long-context rate that ADP was not applying. Those requests now report their real cost, so expect reported spend to step up with no change in usage. Figures recorded before this release keep their earlier, lower numbers.

Selecting Contact support on an error now files a support request with Redpanda. The error was previously logged for diagnostics without a request being created.

When your organization has no Agentic Data Plane environment yet, ADP now shows which of your clusters can host one, with each cluster’s cloud region and city.

Conversation transcripts are recorded again after new conversations stopped appearing in the Activity views on LLM Providers and Cost & Usage in the previous release.

Filtering and paging through lists now work correctly across Connections, Integrations, and Secrets. Filter checkboxes show and clear what you selected, and lists no longer jump to the wrong page while data reloads.

This release includes general security and stability improvements.

Agents now support scheduled triggers. Give an agent a cron schedule and it runs automatically at the times you set.

You can now pick a token-optimized output format for each MCP server. Set a server’s Output format to JSON, TOON, or JTON to control how tool results are encoded before an agent reads them.

Connecting a remote MCP server now streams its responses reliably, so its tools load instead of appearing empty.

As a preview, MCP servers now have a Data Policies tab where you can define rules that mask, redact, or filter fields out of a tool’s results before the model sees them. Author and preview policies per server.

OpenAPI MCP servers can now authenticate to a downstream API with an OAuth client-credentials (service-account) flow. All requests share one service-account identity instead of each user’s own credentials.

OpenAPI MCP servers now return tool results as structured output with the response’s status code, headers, and parsed body, so agents can read them reliably.

You can now find a provider in the list by its display name, not just its ID.

Deriving an environment with rpk ai env add --from <base> now inherits the base environment’s ADP API URL and authentication mode instead of resetting them to defaults.

The rpk ai env list output no longer shows duplicate or stale local-only environments.

This release includes general security and stability improvements.

Connecting a remote MCP server that uses OAuth no longer requires you to paste in authorization and token endpoints or a client ID and secret. Redpanda reads the server’s OAuth configuration from its URL and registers a client automatically through Dynamic Client Registration (DCR).

The ServiceNow MCP server can now search CMDB configuration items, so agents can look up servers, laptops, applications, and services.

The setup guides for connecting an AI client now cover Claude, ChatGPT, Zed, and the Codex CLI, with corrected steps.

The copy-paste code examples for connecting to MCP servers and LLM providers now use the latest SDK versions across Node, Python, Go, Java, and curl.

List rows are now fully clickable to open the detail view, instead of only the name link.

When something fails, the interface now offers clear next steps such as retry, reload, and copying the error details.

Connecting a remote MCP server that signs you in with your own account (user-delegated OAuth) now works. Sign-in no longer fails before completing.

Every LLM provider now has a Models tab where you can browse the models it serves, turn individual models on or off, and filter them by capability or status.

Each provider now has an Activity tab that shows the agents, requests, and conversations routed through it over a time range you choose.

The providers list shows each provider’s models with small trend charts for recent requests and spend, and you can filter the list to providers that serve a specific model.

Each provider’s Settings tab shows which Secret Store entries authenticate it, with a shortcut to manage those keys.

The Connect setup guide lets you pick your coding agent and operating system, and gives copy-paste rpk ai steps tailored to your environment.

Running a tool in the server inspector now shows clear, actionable validation errors for invalid inputs instead of a generic failure.

Connecting an AI client to a Code Mode MCP server over OAuth now succeeds; Code Mode adds tools that let agents run sandboxed code against the server. Clients such as Claude Code and the MCP Inspector now register automatically through Dynamic Client Registration (DCR) instead of failing with a missing-scope error.

Using the browser Back button after starting a connection no longer leaves the Connect button stuck; the connection state resets cleanly.

A brand-new environment no longer shows a misleading all-clear health status before you have set anything up.

OpenAI providers now offer the GPT-5.6 model family: GPT-5.6 Luna, GPT-5.6 Terra, and GPT-5.6 Sol.

Each model now has a detail page showing its overview, pricing, and capabilities.

Model usage now shows a clear error when a usage query fails, instead of misleading zero values.

Cost and usage reports now count cached tokens accurately even when a provider returns incomplete usage data.

The inspector shows upstream HTML and XML errors as readable text instead of raw markup.

The Connect and Reconnect buttons now show a loading state while the provider’s consent screen opens.

You can now use the Google Gemma 4 model family with your Amazon Bedrock providers, in three sizes from largest to smallest: Gemma 4 31B, Gemma 4 26B-A4B, and Gemma 4 E2B.

Manage OAuth providers and MCP client access in one tabbed Integrations setup area, replacing the separate pages.

The My Connections page groups your account connections by status and lets you connect or reconnect without leaving the page.

This release includes general security and stability improvements.

Connections that renew automatically, for example GitHub, no longer show as expired shortly after connecting. You are prompted to reconnect only when a connection can no longer renew itself.

Amazon Bedrock providers now offer two new models: Mistral Large 3 and Amazon Nova 2 Lite.

The Amazon Bedrock MCP server can now generate embeddings with Cohere Embed v4.

Creating or updating an agent now rejects an invalid Bedrock model name immediately, instead of accepting it and failing later.

When an agent task is still running after five minutes, the inspector stops the live view and points you to the Activity page to follow it; the task keeps running.

This release includes general security and stability improvements.

A reply that was cut off by the model’s token limit no longer causes the next message in the conversation to fail.

Tag your agents and break down cost and usage by tag: group the report by a tag key, filter by tag values, and see untagged traffic in its own bucket.

This release includes general security and stability improvements.

Download your full cost and usage report as a CSV, broken down by provider, model, user, or agent and by hour, day, or month.

The home dashboard is rebuilt around an overall health status, an hourly chart of token use and spend, and your top spending agents and users at a glance.

The Connect tab gives you ready-to-run rpk ai setup steps and client code snippets, including running Claude Code against a Bedrock provider without local AWS credentials.

When a request fails, a clearer error page lets you retry, open full error details, and copy the failing request as a curl command to reproduce or report it.

This release includes general security and stability improvements.

In the agent editor, the model picker now stays anchored to its field instead of jumping to the side when you open it.

After a new release is deployed, the app now reloads cleanly instead of getting stuck on a stale page.

Text across the interface is now larger and easier to read.

This release includes general security and stability improvements.

Collapsible and accordion panels now animate smoothly.

Launch Claude Code against one of your Amazon Bedrock LLM providers with rpk ai run claude -L <provider>; the gateway authenticates to AWS for you, so your machine needs no AWS credentials.

Give each subagent its own model and LLM provider in the agent editor.

Updating an agent no longer interrupts it: the new version starts and takes over before the old one stops.

Each model now reports its max_input_tokens and max_output_tokens limits through the API.

Filters in the metric detail panel on provider and model pages now apply to the chart, and the panel no longer hides its filter menus or covers the sidebar.

The spend total in the metric detail panel was inflated when grouping by cost type; only the display was wrong, recorded usage was unaffected.

In period comparisons grouped by model or user, each line now compares against its own numbers from the previous period, not another line’s.

MCP clients such as Claude Code, Cursor, and ChatGPT can now register themselves through OAuth Dynamic Client Registration (DCR), instead of needing an OAuth client created by hand.

Creating policies is now faster, since resource options load as you need them instead of all at once.

Claude Code users can now set up managed MCP servers with a single command, or use manual configuration from a separate tab.

Claude Sonnet 5 is now available to select on the Anthropic and Bedrock providers.

A completed tool call in the agent inspector now switches to Completed the moment it finishes, instead of showing Working until the next step.

Signing in with rpk ai auth login now clears the expired-token errors that kept rpk ai commands failing even after a successful login.

When an agent produces a chart, the agent inspector now displays it: switch between the chart, its data, and its definition, and zoom in or download it.

Capture a failed request for support and get a reference ID to quote.

Failed requests now show the full error message and type, with expandable technical details, instead of a generic internal error.

The MCP servers, LLM providers, OAuth providers, OAuth clients, and Secrets Store lists now have consistent search, filtering, and multi-row selection.

The Bedrock provider rejected valid model IDs whose version contains a dot, in all regions; these now route correctly.

Connect to Sentry through a new managed MCP server.

Set your own per-model pricing on a provider, so usage and cost reflect the rates you actually pay.

Apply or diff your LLM providers, MCP servers, and OAuth providers from config files, now against an Agentic Data Plane environment.

Filter the server list by type and status.

Filter providers and clients by status and how they were created.

Export charts to CSV and download a full usage report.

The environment switcher now shows each environment’s region, so you don’t make changes in the wrong one.

Fixed hangs when switching environments, and the app now recovers on its own after a new version ships instead of showing errors.

Fixed filters showing wrong results, and lists now refresh immediately after you add or edit an item.