CKEditor AI On-Premises

• Added new AI models to the default configuration

  The default configuration now also includes: Claude Sonnet 5, Claude Fable 5, GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, Gemini 3.6 Flash.

• Added Azure OpenAI configuration options

  Azure OpenAI can now authenticate with tokens instead of API keys. An optional useDeploymentBasedUrls flag routes requests through the deployment-based Chat Completions path, which unblocks Foundry-style endpoints that expose a custom base URL.

• Added a custom content moderation adapter

  Deployments can route content moderation to their own service instead of the built-in provider moderation.

• Added guardrail rule overrides

  Each guardrail operation accepts rules, which replaces the built-in classification rules, and extraRules, which is appended to them. Guardrail checks now also cover context library prompts and take the editor configuration into account.

• Added an MCP server for external agents

  The service exposes its own MCP endpoint, so external agents can process documents, discover and use context items.

• Added deployment-level MCP server configuration

  MCP servers can be configured once for the whole deployment instead of per request, including servers behind OAuth. MCP servers can also back context items, so context content is pulled from external systems such as wikis, document stores and file servers instead of being uploaded.

• Added context library

  Contexts are reusable buckets for prompts and files, defined once in the context library and applied across the different features. A context can hold tone of voice, brand and style guidelines, domain knowledge, or any instruction that would otherwise be repeated in every request. Once saved, a context can be applied to chat messages, actions, and reviews, so the AI follows the same rules consistently without them being restated.

• Improved AI editing capabilities

  Structural editing tools are enabled, including a dedicated table tool, and they can operate on a range of a document instead of its whole content.

• Added per-request headers for image analysis

  Requests can pass extraHttpHeaders, which the assistant forwards when fetching images, so images behind authentication can be analyzed.

• Added retention cleanup for actions and reviews

  Stored actions and reviews are now removed by the scheduled cleanup job after one year, the same retention window conversations already use.

• Enhanced the document processing endpoint

  The endpoint can stream its response, with HTML as the default streaming format, process multiple documents and selections in a single request, return changes in the ARF format, accept the editor configuration, use skills, and call external MCP servers.

• Fixed partial translation of long documents

  In specific cases, our service failed to translate some paragraphs in long documents. We have made a number of improvements to our tools to alleviate this problem.

• Fixed web resources being returned with web search disabled

  GPT models no longer return online resources when web search is turned off.

• Fixed unsupported file errors being hidden

  The error is returned in the API response instead of appearing only in the container logs.

• Fixed the status code for malformed request bodies

  A malformed JSON body now returns 400.

• Fixed conversation titles for chat commands

  Titles are generated in the right language and describe the conversation more accurately.

• Fixed context identifier validation

  Identifiers built only from symbols are rejected, and the denylist matches substrings.

• Updated dependencies

  Selected dependencies have been updated to the latest versions, improving performance and security.

• Added new AI models to the default configuration

  The default configuration now also includes: Claude Opus 4.8, Gemini 3.5 Flash.

• Added AI image recognition

  The assistant can now analyze images in a document and generate descriptions, captions, and alt text for them.

• Added configurable guardrails

  Content-safety guardrails can be configured per operation (chat messages, custom actions, reviews, document processing) and for content uploads (files, web resources, documents), with configurable classification prompts and a deterministic phrase blocklist as a first-pass check.

• Added configurable model resolution with fallback

  Define “agent” models that resolve to a priority-ordered chain of concrete models, allowing graceful model rotation and deprecation without per-instance configuration changes.

• Added an endpoint for updating conversation file metadata

  A new PATCH endpoint updates mutable file metadata in place, without re-uploading the file.

• Improved AI editing capabilities

  Including more reliable handling of structural changes and better support for large and multi-root documents.

• [EXPERIMENTAL] Enhanced the document processing endpoint

  Now supports model selection, reasoning, and web-search options. AI also produces CKEditor 5-specific HTML markup more reliably.

• Enforced permission checks on authenticated endpoints

  File and web-resource read/delete, file upload now require the appropriate permissions.

• Hardened URL validation

  Only HTTP(S) URLs are now accepted for web resources and file references; other URL schemes are rejected with a 400 error.

• Improved intent classification

  Fixing responses in the wrong language, missed document-modification queries, and false-positive edit intent on chat-only queries.

• Reduced response times for listing conversations and for content-safety checks

• Updated dependencies

  Selected dependencies have been updated to the latest versions, improving performance and security.

• Added new AI models to the default configuration

  The default AI models configuration now includes:

  • Claude Opus 4.7
  • GPT-5.5

• Added support for OAuth-authenticated MCP servers

  MCP servers that require user-level OAuth can now be connected to the assistant. The full flow is included: an OAuth provider, public OAuth callback and management endpoints, persistent token storage, and authenticated MCP tools wired into conversations.

• Improved CKEditor 5 feature awareness in AI Chat

  The AI Chat is now more aware of CKEditor 5 features and returns feature-specific markup in more scenarios.

• Added an endpoint for updating message attributes

  A new PATCH endpoint allows updating client-side metadata on a conversation message.

• Added an endpoint for updating document attributes

  Attributes of a conversation document can now be updated via a dedicated endpoint.

• Enabled reasoning by default

• Added per-request reasoning effort control

  Clients can now override the reasoning effort on a per-request basis via the API.

• [EXPERIMENTAL] Introduced a stateless document processing endpoint

  A new stateless, one-shot endpoint for processing a document without creating a conversation is now part of the public API.

• [EXPERIMENTAL] Added multi-root document handling

  Initial support for documents with multiple content roots has been introduced, with iterative stability improvements. Multi-root processing activates only when the request contains more than one document. Requires CKEditor 5 v48.1.0 or later.

• Reduced chat response verbosity

  Default chat responses are now more concise, especially for GPT analysis and critique responses.

• Fixed the status code returned for an invalid compatibilityVersion

  Requests to /v1/models/* with an unknown compatibilityVersion now return 400 instead of 500.

• Fixed review endpoints rejecting empty content

  Submitting a review with empty content is no longer rejected.

• Stabilized the ordering of documents in batch responses

  Documents in batch responses are now returned in a stable, deterministic order.

• Suppressed false-positive MCP warnings in logs

  Spurious MCP-related warnings no longer appear in on-prem container logs.

• Fixed web scraping cancellation

  In-flight web scraping HTTP requests are now canceled when the parent operation is aborted.

• Improved processing of long documents

  Previously, the service would sometimes stop generating changes partway through a very long document. Generation now reliably completes for long documents.

• Updated dependencies

  Selected dependencies have been updated to the latest versions, improving performance and security.

• Dedicated support for cloud providers

  Improved support for leading cloud providers. It is now possible to easily connect models from:

  • Vertex AI
  • Amazon Bedrock
  • Azure OpenAI

• Added OpenTelemetry observability support

  The service now supports OpenTelemetry instrumentation, enabling monitoring of AI interactions, token usage, and response quality, as well as HTTP calls and database operations, and is compatible with observability platforms such as Langfuse.

• New AI models support added

  • Claude 4.6 Sonnet (replaced 4.5 Sonnet in the recommended models list)
  • GPT-5.4 (replaced GPT-5.2 in the recommended models list)
  • GPT-5.4 Mini
  • Gemini 3.1 Pro

• New endpoint for listing documents of a conversation

  Added a new, paginated endpoint (GET /v1/conversations/:conversationId/documents) which allows listing all documents belonging to the conversation.

• New endpoint for getting file metadata

  Added a new endpoint that allows retrieving metadata for uploaded files without downloading the full content.

• Conversation sorting by pinned status

  Conversations can now be sorted by their pinned status, with pinned conversations appearing first.

• Increased maximum non-image file size to 25 MB

  The maximum size for a single non-image file has been increased from 5 MB to 25 MB. The Anthropic-specific 5 MB restriction now applies only to images, not all files. The total size limit for all files remains at 30 MB.

• Improved AI model failover reliability

  The service now recovers faster and more gracefully when an AI provider experiences outages.

• Improved reasoning and web search capabilities

  The AI Service has better guidance on when and how to use reasoning and web search capabilities, resulting in more effective responses.

• Improved AI Quick Actions suggestions to produce more reliable results

  In some cases, AI Quick Actions were producing incorrect suggestions and were falling into repetitive stream loops. This has been fixed.

• Improved PDF parsing performance

  Resolved performance issues with PDF file parsing that could cause slow processing.

• Updated dependencies

  Selected dependencies have been updated to the latest versions, improving performance and security.

Initial release of the CKEditor AI On-Premises.
The version was synchronized with the CKEditor Collaboration Server On-Premises to
ensure and underline the compatibility of the two applications.

It is highly recommended to synchronize the version of the CKEditor AI On-Premises with
the CKEditor Collaboration Server On-Premises to ensure the compatibility of the two applications.