Skip to content

What's new

Recent additions and fixes to Myra AI Workspace, organised by feature area.

Product updates — September–October 2026

Chat

  • One-click sign-in link. The sign-in email now carries a one-click link alongside the 6-digit code; opening it on the same device signs you in without typing the code. See Signing in.
  • PDFs attach their embedded figures. When a PDF is read as extracted text, the figures and charts embedded in it are also attached to the message, with a notice when some are left out. See Importing a file.
  • Add budget from a blocked send. When the workspace prepaid balance is exhausted, a wallet owner sees an Add budget action in place of the upgrade prompt. See Chat.
  • Ask the docs. The documentation button in the icon row below the logo now opens an in-app assistant that answers questions about the product in the language of the question, with links to the documentation. See Interface overview.

Agents and workflows

  • Delivery-failure alerts. The Runs tab shows a banner when recent runs failed to deliver their output, and the Schedules tab marks a failing schedule with a consecutive-failure badge. See Agents.
  • Connector workflow step. A workflow can call a saved MCP or API connector mid-run and reuse its scrubbed result in later steps. See Workflows.

Administration

  • View as user protects private data. An administrator opening View as user can no longer read a member's conversations, memories, billing, tokens, or other private content — those are refused, and the member's email is masked. See Users.
  • Internal unmasked delivery toggle. A tenant admin can control whether a scheduled or workflow run that masked PII emails the full result to internal recipients (on by default). See Tenants.
  • Code interpreter on by default. A new gateway offers the code-interpreter tool unless the toggle is cleared. See Gateways.
  • Brand-stripped model names. The model Display Name drops the vendor prefix and keeps the tier and version (for example Sonnet 4.6). See Provider costs.

Billing and lifecycle

  • Partner payouts. Partner payouts are produced as a SEPA credit-transfer batch and a self-billing Gutschrift per partner. See Partner.
  • Signup-abandonment reminder active. The one-off reminder to unverified, consent-bearing self-serve signups is enabled platform-wide. See Authentication.
  • Enterprise budget tops up the wallet. On an Enterprise tenant, raising the organisation budget credits the increase to the tenant's prepaid wallet as an administrative grant, with no payment taken. See Budgets.

Observability

  • Member-reactivation card. The Dashboard (platform admin) and Cost analytics (per tenant) show a member-reactivation roll-up — mails sent, reactivations, and opt-outs. See Dashboard.
  • Plan-aware taglines health. The Health view gained a counter that reports whether the generated plan-aware model copy is current. See Health dashboard.
  • My feedback. Every user can track their own submitted feedback and its status at My feedback. See Feedback review.
  • Costs in your currency. The Dashboard shows costs in your Display currency preference, matching the Currency control already on Cost analytics and the Live monitor. See Dashboard.

Product updates — late August 2026

Chat

  • Auto shows the resolved model. On an Auto conversation the model picker now shows which model Auto chose for the conversation (for example Auto · Sonnet 5), with the matching privacy indicator.
  • Read aloud in five languages. A finished reply is spoken in the language of the answer — German, English, French, Dutch, or Spanish — and numbers, dates, currency, ranges, and abbreviations are first rewritten into fully spoken form, entirely inside the EU boundary. See Reading a reply aloud.
  • Collapsible spreadsheet attachments. A sent spreadsheet or CSV attachment appears as a collapsible chip; the full extracted content is still sent to the model.
  • Jump to the latest message. A button returns a scrolled-up thread to the newest message.

Code interpreter

  • More input formats. Markdown, HTML, XML, and YAML files can now be used as inputs and edited across turns. See Code interpreter.

Configuration & routing

  • Route on headers and metadata. Routing-rule conditions can match a request header (header:<name>) or request metadata (meta:<key>). See Routing rules.
  • Tabbed tenant editor. The tenant edit dialog is organised into tabs with a single Save Changes and an unsaved-changes guard. See Tenants.
  • Add a user from a tenant. An administrator can add a user directly from the tenant detail view.
  • Compliance reports per tenant. The monthly compliance-report job is enabled per tenant. See Governance.

Account & access

  • Edit your own name. The Profile page lets any user change their account name. See Profile.
  • My access. A read-only page explains your own roles, permissions, and group membership.
  • Install as an app. Add Myra AI Workspace to a phone or tablet home screen and open it as a standalone app. See Interface overview.
  • Resend the sign-in code. The email-code step carries a Resend control with a short cool-down.
  • Premium features. Agents, Workflows, Playground, and Scheduled tasks are always visible; when the plan does not include one, it shows a Premium badge. Clicking Workflows, Agents, or Playground opens a shared upgrade pop-up whose call to action leads to the contact form; Scheduled tasks links to its upgrade page.

Product updates — August 2026

Security & governance

  • Moderation queue for flagged interactions. When a guardrail flags a top-level prompt or output, the interaction is queued for human review on the new Tool reviews tab of Settings › Approvals — filter by kind, then approve, deny, or annotate. See Tool reviews.
  • Human-in-the-loop agent tool gate. An unattended agent run can be held at a gated built-in tool call until a reviewer approves it; an un-reviewed hold expires fail-closed and the tool is never called. See Agents and Tool reviews.
  • Tool/RAG/MCP results are now scanned for injection at the tool-result fusion seam (a jailbreak-phrase tripwire), and response-phase guardrails now run on streaming replies too. See Threat model.
  • Guardrail config-as-code. Export a tenant's guardrail configuration, review a diff, and dry-run a policy before importing it. See Guardrails API.
  • Per-role guardrail & model-access policy. Guardrail and model-access rules can now be scoped per role. See Guardrails API.
  • OWASP-LLM Top-10 / MITRE-ATLAS taxonomy. Guardrail events now carry a standardized threat taxonomy, surfaced in compliance exports and SIEM records. See Compliance.

Agents & access

  • Per-agent tokens. Issue a gateway token bound to a single agent for clean attribution and fail-closed confinement. See Users & tokens.
  • "View as User" impersonation. An admin can view the product as a specific user for support, server-scoped and dual-audited, revocable with a stop control. See Impersonation.
  • Per-tenant feature toggles. Scheduled Tasks and Workflows can each be turned off per tenant; the nav entry is hidden and the routes refuse access when disabled.

Chat & documents

  • Edit an answer. Correct a settled reply in place — saved for display only, not sent to the model, with an edited marker and Revert to original. See Reply actions.
  • Document generation in the code interpreter. The assistant can author Word, PowerPoint, PDF, ODT/ODS/ODP, and HTML deliverables. See Code interpreter.
  • iPhone photos. Chat image attachments now decode HEIC/HEIF, AVIF, and DNG (Apple ProRAW), plus TIFF and BMP. See Importing a file.

Models & billing

  • The default local model became a dense, multimodal Qwen. The canonical on-prem model moved to the dense, multimodal qwen3.6-27b (used by Auto and for image understanding). It was superseded on 2026-09-13 by qwen3.8-27b (qwen3.6-* are now hidden aliases). See Myra provider.
  • Per-user display currency. Choose EUR or USD for how costs are shown; enforcement stays in the stored currency. See My preferences.

New workspace layout. The navigation was reorganised into two modes on a single left rail: a Workspace mode (the Workspace and Overview clusters) for everyday work, and a Settings mode (the Settings, User Management, and Account sections) for configuration and personal settings. Switch to Settings/User Management/Account from the user block at the bottom of the left sidebar; a Back to workspace link returns you. Several surfaces moved or were renamed:

  • Model prices → Provider Costs (now Settings › Costs › Provider Costs; table refreshed to eight columns).
  • Governance moved to Settings › Security.
  • The approval inboxes are now one Approvals page with Personal, Agents, Tool reviews, and Configurations tabs.
  • Personal pages (Profile, My preferences, My commands, My PII, My MCP connectors) regrouped under Settings and Account; member Plan & Billing moved out of the chat overlay into the Account section.
  • My MCP Connectors is visible to admins again. Settings › Connections now shows the personal My MCP Connectors tab to every role — admins manage the tenant's connectors on the MCP Connectors tab and activate their own personal credentials here (an earlier iteration of this layout had hidden the personal tab for admins).
  • The Prompt library tab now reads Prompts (Workspace › Library) and has a Use in chat row action.

List rows now open via an explicit action icon (open/edit/run/more) — the whole row is no longer clickable — and data tables support optional keyboard row navigation.

Session feedback rating changed direction to 5 = best (was inverted).

Capabilities catalog — July 2026

New: Capabilities page — A single feature map groups every capability by what it lets you do and links into each detail page. Image and scanned-document ingestion via OCR is now discoverable from the top level. See Capabilities.

Scheduled tasks — July 2026

New: Failure notifications — The task owner is notified when a run fails, at most once per task per day, with no run content in the notice. See Reliability.

Context window — July 2026

New: Long and data-file turn handling — Documented incremental streamed compaction, the partial-answer synthesis when a turn runs out of time, and the automatic handling of long and data-file turns. See Context window management.

Agents — July 2026

New: Agents chapter — Members build reusable agents that combine a system prompt, a model, tools, and bound knowledge. Each agent runs on demand from the chat, on a schedule, or from a webhook. See Agents.

New: Agent versions — Every publish creates a numbered version. Compare two versions side by side and restore an earlier version at any time. See Managing agent versions.

New: Organisation agent catalog — Publish an agent to the organisation catalog so every member can clone it. See Publishing an agent to the organisation catalog.

New: Publication governance — A KI-Manager reviews and approves an agent before it reaches the catalog. See Governing agent publication.

New: Per-run cost — Each agent run records its own cost, attributed to the run owner.

Workflows — July 2026

New: Workflows chapter — The workflow builder chains agents and delivery steps into a single automated run. The chapter is now a top-level part next to Gateways. See Workflows.

New: Human-approval step — A workflow can pause at an approval step and wait for a nominated approver before it continues.

New: Schedule and webhook triggers — A workflow starts on a fixed schedule or from an inbound webhook, in addition to a manual start. See Triggers.

New: Per-step test — Test a single step against sample input from the builder before publishing the workflow. See Testing a step.

New: Runs monitoring — A runs list and a per-run detail view show step status, per-step cost, PII state, and the option to cancel a running workflow. See Monitoring runs.

New: My approvals inbox — A dedicated inbox collects the workflow approvals and held scheduled-run deliveries assigned to you. See My approvals.

Chat bridge — July 2026

New: Chat bridge — Connect an agent to Mattermost, Microsoft Teams, or Slack. Members talk to the agent from the chat platform through a bot account. See Chat bridge.

New: Bot token rotation — Rotate a bridge bot token without recreating the connection. See Rotating a bot token.

New: Web-search opt-in — The connection owner opts a bridge bot in to web search. Every other egress tool stays blocked on the bridge path. See Allowing web search.

Groups — July 2026

New: Groups — Admins and KI-Managers group users and grant access to agents and knowledge spaces at group level rather than per user. See Groups.

Data analysis — July 2026

New: Code interpreter — A gateway-hosted sandbox runs Python on an attached file, produces charts, and returns downloadable result files. Code interpreter works on EU and open models, not only Anthropic models. Myra enables it per organisation. See Code interpreter.

Privacy & residency — July 2026

New: EU data residency — A gateway restricts every commercial model to an EU region and blocks non-EU routes fail-closed. See EU data residency.

New: Enforced PII masking — A tenant admin makes PII masking mandatory. Members cannot switch off the masking that runs before a message reaches an external model or web search. The privacy pill shows the locked state. See Easy chat — Privacy pill.

Identity & access — July 2026

New: Single sign-on — A tenant signs in through a generic OIDC provider or a SAML 2.0 identity provider, in addition to email OTP. The login page detects the sign-in method from the email domain. See Signing in.

New: SCIM provisioning — An identity provider creates, updates, and deactivates users through the SCIM 2.0 endpoint. See Tenants.

New: Department-to-group mapping — A department attribute from the identity provider maps a user to an internal group automatically, without a SAML configuration row.

New: Restrict processing — An admin restricts processing for a user account under Article 18 of the GDPR. See Users.

New: Organisation-wide system instruction — An admin sets a system instruction that applies to every chat and every agent across the tenant. See Tenants.

Appearance — July 2026

New: Tenant white-label branding — A tenant admin sets a product name, a logo pair, a favicon, an accent colour, a home greeting, a chat disclaimer, and up to four sidebar links. The branding applies at runtime with no rebuild, and login pages and notification emails carry it. See Tenants.

Chat — July 2026

New: Guardrail-block bubble — A request that a guardrail blocks now renders as a distinct policy-block bubble rather than an ordinary reply. The bubble survives a conversation reload, and a transient guardrail-unavailable block offers a Retry action. See Easy chat — Guardrail block.

New: PII pill mask and unmask — The privacy pill shows the masking state per conversation and lets you mask or unmask a detected item, unless the tenant enforces masking. See Easy chat — Privacy pill.

New: Inline citations — A web-search answer carries numbered [N] citations, a freshness picker, and a process timeline that shows each search step. See Easy chat — Citations.

New: Message comments — Add a comment to any reply in a conversation. See Easy chat — Commenting on a reply.

New: CSV and TXT export — Download a conversation as a CSV or a plain-text file, alongside the existing Markdown and PDF exports. See Easy chat — Exporting a conversation.

Improved: Regenerate provenance chip — A steered regenerate shows a subtle chip on the regenerated reply that names the adjustment. The chip persists across reload and share. See Easy chat — Regenerating a reply.

Improved: Generated-image actions — A generated image card carries its own Copy and Download actions, so the copy places the image on the clipboard and the download writes a correctly named file. See Easy chat — Generated-image actions.

Shared conversations — July 2026

New: Live rooms — A shared conversation runs as a live room with a presence roster and single-driver turn-taking. A reader with access takes the wheel to become the driver. See Shared conversations — Live rooms.

Account & billing — July 2026

New: Self-serve sign-up — A new customer signs up, picks a plan, and starts a tenant without an invitation. See Signing up and managing billing.

New: In-app billing — A billing panel inside the chat shows the plan, the allowance, and the renewal date. Cap states appear in the chat when the allowance is spent, and a cancellation page ends the subscription. See Signing up and managing billing.

New: Allowance alerts — A self-serve tenant receives an email at 80 percent and at 100 percent of its allowance, once per billing period. See When the allowance runs out.

New: Provider-key rotation — A self-serve tenant that hits a blocked provider key fails over to a healthy key automatically, so a revoked key no longer stops inference.

Observability — July 2026

New: By Agent and By Channel analytics — The analytics view breaks activity down by agent and by chat-bridge channel, with per-agent and per-user cost columns. See Cost analytics — Analytics tabs.

New: Per-tenant request-log disabler — An admin turns off request logging for a specific tenant. See Request logging — Disabling request logging for a tenant.

Web search — July 2026

Improved: PII scrub on web search — A web-search query is scrubbed of PII before it leaves Myra, and only the tokenised query egresses to the search provider. See Web search.

Chat — June 2026

New: Office and OpenDocument export — Generated content can be downloaded as Word (.docx), Excel (.xlsx), PowerPoint (.pptx), and OpenDocument (.odt, .ods, .odp) files. A single Download menu offers every format; the gateway reports inline when a format does not fit the content.

New: Voice input — Record a voice message in the chat input. The gateway transcribes the recording to text before sending.

New: Generated-file artifact cards — Files the model generates appear as artifact cards in the conversation and persist to the conversation or project file list.

New: Conversation file space — A chat that is not part of a project owns its own file space. Generated files stay with the conversation. See Conversation file space.

New: Personal style toggle — A per-conversation toggle controls whether the personal style from the profile applies to the conversation.

Improved: PII restore — Masked PII tokens are reliably restored in the model response so the final answer reads naturally.

Privacy & PII — June 2026

New: My PII configuration — List personal words and phrases that the gateway masks in your messages before the messages leave Myra, in addition to the gateway's PII detectors. See My PII configuration.

Account — June 2026

New: My MCP Connectors — Connect to the per-user MCP tool servers your admin made available, each with your own personal credential. See My MCP Connectors.

MCP connectors — June 2026

Improved: Tools sent in the request body — Tool definitions are sent in the request body rather than a header, removing a case where large tool sets were silently dropped.

Fixed: Tool load errors are visible — When a connector fails to load its tools, the chat now shows an explicit message instead of silently continuing without the tools.

Activity and analytics — June 2026

New: Health dashboard — Platform admins can review the operational state of the gateway across its failure surfaces for the last 24 hours. See Health dashboard.

Appearance — June 2026

Changed: Myra theme — The interface uses the Myra navy and cyan colour scheme across every view.

Projects — June 2026

New: Project access tiers — Projects can be set to Local only, PII protection required, or User configurable, controlling which models are available and whether PII masking is forced. See Projects — Access tier.

Chat — April 2026

New: Ghost mode — Toggle ghost mode using the ghost icon in the conversation list header. While active, no messages are saved to the database, no request logs are created, and conversations exist only in memory. Use it for exploratory or sensitive conversations.

New: Conversation sharing — Generate a public share link for any conversation. Recipients can view the conversation without logging in. Share links can be revoked at any time. See Easy chat — Sharing a conversation.

New: Starring and archiving — Star conversations to pin them at the top of the list. Archive conversations to remove them from the main view. A toggle reveals archived conversations when needed.

New: Memory system — The model can remember facts, preferences, and instructions across conversations. Memories are injected into the system prompt automatically. Create memories manually or let the model learn them from conversation context. Memory can be disabled per conversation. See Easy chat — Memory.

Removed: Chat presets — The tenant-level chat-preset buttons and the per-user saved-preset list in the Settings drawer have been removed. Users now pick gateway and model directly through the unified Curator/model selector. Routing decisions are driven by Curator tasks or by an explicit model pick — no separate preset layer.

New: Web search in Chat — A globe icon in the chat toolbar toggles web search on or off for the current conversation. Previously web search was available only via API header or in the Playground.

New: PowerPoint (.pptx) attachment — .pptx files can be attached to chat messages alongside existing Word, Excel, and PDF support.

New: Extended thinking display — When extended thinking is enabled, the model's reasoning appears in a collapsible block above the response with a duration indicator.

New: Automatic context management — Long conversations are automatically summarised when they approach the model's context limit, allowing uninterrupted multi-turn sessions.

New: Conversation URL sync — The active conversation ID is reflected in the browser URL as a ?conv= parameter, enabling direct links to specific conversations.

New: Copy to Markdown — A clipboard button in the configuration bar copies the full conversation as Markdown text.

New: MCP connectors — Register external Model Context Protocol servers to provide tools that models can call during chat. See MCP Connectors.

Projects — April 2026

New: Binary knowledge files — Upload PDFs, Word documents, spreadsheets, presentations, and images to project knowledge bases. Server-side text extraction makes the content available in conversations.

New: Search, filter, and sort — The projects list includes a search bar, a role filter (all, your, team, shared), and a sort control (recent activity, last edited, date created).

New: Batch save-to-project — Save multiple conversations to a project at once from the Chat view.

Renamed: "Knowledge" tab is now "Files" — The project detail tab that manages knowledge files has been renamed from Knowledge to Files.

Chat

New: Session feedback — Rate any conversation on a 1–5 star scale and add an optional comment using the flag (🚩) icon on each conversation row. Ratings are saved per conversation. See Easy chat — Rating a conversation.

New: Drag-and-drop file upload — Drag any supported file from your desktop and drop it anywhere on the message area. A blue drop target appears while the file is dragged over the panel. The file is attached exactly as if selected via the paperclip button. See Easy chat — Importing a file.

New: Processing status indicator — A spinner and status label appear in the message area between the moment you send a message and the moment the first response token arrives. The label shows what is happening — for example, extracting text from a document or waiting for the model.

New: Spreadsheet file upload — Attach spreadsheet files (.csv, .tsv, .xlsx, .xlsm, .ods) to any chat message. Claude reads and analyses the content and can answer questions about it. Spreadsheet files require an Anthropic provider key on the selected gateway. See Easy chat — Importing a file.

New: Conversation export — Download a conversation as a Markdown or PDF file using the export buttons in the chat configuration bar. Both buttons are disabled when no conversation is active. See Easy chat — Exporting a conversation.

Improved: Long response handling — Long responses complete automatically without any action required.

Improved: Conversation auto-title — Conversation titles generated after the first exchange are more accurate and consistently formatted.

New: Default system prompt — The settings drawer opens with a default system prompt already filled in. Edit or clear it as needed.

New: Code-block copy button — Code blocks in assistant messages show a language label and a Copy button.

Improved: Chat UI — Updated message layout, avatar styling, and Markdown typography.

Authentication

New: Stay logged in — A Stay logged in for 30 days on this device checkbox is available on the Email OTP login step. When selected, the session remains active for 30 days instead of the default 8 hours. See Authentication — Session duration.

Model prices

Updated: Anthropic model support — Pricing is now included for the Claude 4.5 and Claude 4.6 model families, including all dated model aliases. Prices for existing models have been updated to current rates. See Model Prices.

Chat

New: Persistent multi-turn chat — The Chat page provides a full conversation UI that routes every message through the gateway. Conversations are saved per user, support multi-turn history, and are completely isolated — no user can access another user's conversations. See Chat.

New: File attachments in chat — Attach images (JPEG, PNG, GIF, WebP), PDFs, and plain text files to any message. Each file type is sent as the appropriate Anthropic content block (image, document, or text).

New: Word document (.docx) support — .docx files are uploaded to the Anthropic Files API and processed by the Anthropic docx Agent Skill. Claude reads and analyses the document content server-side. The skill header is automatically re-sent on follow-up turns in the same conversation.

New: Unsupported file error — Attaching a file type that the gateway cannot forward now shows an explicit error message listing supported formats. Previously, unsupported files were silently ignored.

New: Chat localStorage persistence — Tenant, gateway, and model selections are saved to local storage and restored when you return to the Chat page or navigate away and back.

Gateway detail view

New: Collapsible cards — Each card on the gateway detail page (Gateway config, Provider Keys, Auth Tokens, Guardrails, Routing Rules, Circuit Breaker) can be individually collapsed and expanded using the ▼/▶ toggle in the card header. Collapsed state is persisted per gateway in local storage.

SIEM integration

New: SIEM event streaming — Security events can now be forwarded asynchronously to an external SIEM. Supported backends: Splunk HEC, Elasticsearch / OpenSearch, Vector HTTP source, and Syslog (CEF or RFC 5424). SIEM config can be set at tenant level (default for all gateways) or overridden per gateway. Delivery is fire-and-forget and never adds latency to inference requests. See SIEM Integration.

Admin UI — role-based navigation

New: Management section hidden from member and viewer roles — The Management sidebar section (Tenants, Gateways, Users) is now visible only to admin and tenant_admin users. member and viewer users see only the Observability, Config, and Account sections.

Users

New: Sortable user table — All columns in the Users view are now sortable. Click any column header to sort ascending; click again to sort descending. Sorting is applied server-side. See User management.

New: Tenant reassignment — admin users can now reassign a user to a different tenant from the Edit User dialog. tenant_admin users cannot change another user's tenant.

Analytics dashboard

New: Analytics tabs — The analytics view now breaks down activity across five tabs: By Tenant, By Gateway, By Provider, By Model, and By User. Each tab has a filter bar for searching by name or ID.

New: Overview chart — A 30-day cost and request chart appears above the analytics tabs. Cost is shown as bars (left axis) and request volume as a line (right axis).

New: Latency percentile strip — p50, p95, and p99 latency chips are shown below the overview chart for the selected analytics window.

New: Expanded hero cards — The dashboard now shows six cards: Total Spend, Cache Savings, Total Requests, Error Rate, Top Spender, and Budget Warnings (previously three cards).

New: By Provider tab — Provider-level breakdown aggregated client-side from top-model data, showing request share, cost, model count, and average latency per provider.

New: By User tab — Per-user breakdown (up to 50 users) showing cost, cache rate, error rate, blocked count, and average latency. Only requests with a user_id on the auth token are included.

New: Error rate — By Tenant and By Gateway tables now show an Error% column counting upstream 4xx/5xx responses per entity.

Tracing

New: Gateway request tracing — Gateways can now record step-by-step execution traces for inference requests. Enable with "tracing": {"enabled": true} in the gateway config. Steps include request normalisation, routing decisions, guardrail results, upstream calls, and response delivery. See Request Tracing.

New: Tracing config UI — The gateway Config tab includes a Tracing section to enable tracing and optionally capture request bodies (include_bodies).

New: Traces API — GET /gateways/{id}/traces lists recent traces for a gateway. GET /traces/{id} returns the full step list for any trace. See Traces API.

Budgets & quota

⚠️ Caution: The spend ledger is now persisted in a database table rather than in-memory counters. Spend survives process restarts and worker crashes, but existing in-memory counters are not migrated on upgrade.

New: Persistent spend ledger — Spend is now tracked in a spend_ledger table rather than shared-dict counters. Spend survives process restarts and worker crashes.

New: total budget period — A new "total" period accumulates spend over the lifetime of the budget without ever auto-resetting. Useful for one-time allowances and trial accounts. Valid period values are now "daily", "monthly", and "total" ("weekly" is not a valid value).

New: Actionable QUOTA_EXCEEDED messages — When a budget is exhausted, the 429 response now includes the configured budget, the current spend, and the exact API endpoint needed to either increase the budget or reset spend for the current period.

Observability

New: Timeseries stats — Time-bucketed request counts, block counts, and cost are now available via the Stats API. Supports six bucket sizes (5m, 15m, 30m, 1h, 6h, 1d) with up to 168 buckets per query. See Stats API.

New: Multi-day dashboard views — The dashboard now shows Yesterday and Last 7 days alongside the existing Today, Last hour, and Last minute views.

Providers & routing

Improved: Streaming reliability — The gateway now uses a 5-minute read timeout for streaming responses, preventing premature connection drops on long model outputs. Stop-reason normalisation ensures consistent finish_reason values (stop, max_tokens, tool_calls) across all providers.

Fixed: Ollama model prefix stripping — Requests using the ollama/ model prefix (e.g. ollama/llama3.2) now correctly strip the prefix before forwarding to the Ollama API. Previously the prefix was forwarded verbatim, causing model-not-found errors.

New: Weighted load-balancing — Routing rules support weighted distribution across multiple providers. Traffic that exceeds a provider's weight is automatically routed to the next entry in the fallback chain.

New: Per-provider circuit breaker — After a configurable failure threshold is crossed, the gateway stops routing to the failing provider for a cooldown period before retrying.

Playground

New: Web search — The Playground includes a web search toggle. When enabled, the gateway performs a live search and injects results into the model context before responding. A "searched" badge is shown on responses that used web search.

Fixed: Web search empty results — Search snippets that were empty or whitespace-only are now skipped; the page description is used as fallback.

New: Filter non-runnable models — The model picker can be filtered to show only models with a configured API key, reducing noise in large setups.

New: Gemini native grounding — Gemini models use the built-in grounding feature of Google when web search is enabled, rather than the Brave Search path.

Security & guardrails

New: Human-readable block messages — Guardrail block responses now include the human-readable harm category name (e.g. "Violent Crimes") alongside the category code (e.g. S2).

New: Anthropic tool use on compat endpoint — OpenAI-format tool_calls sent via the compat endpoint to Anthropic are now automatically converted to the native tool_use format of Anthropic.

See also