From setting up your default chat engine and deciding what kind of permissions you give it when searching or working with your Mac, the AI view is where you set these options. For creating AI-based images, see the Image Generation view. And if you need to detect or convert speech to text in images and media files, see the Transcription view.
Chat
Choose your AI model and settings specific to it, as needed. Also set from where the model can get information, if it can effect changes to your database, and what kind of summaries you'd like it to return.
Chat Setup: Specify what large language model (LLM) you want to use and set up any required parameters for it. Note several of the controls here are dynamic and the options will change depending on what LLM you've chosen.
Provider: Choose from the list of supported chat engines, e.g, ChatGPT, Anthropic's Claude, or even one you are running locally.
API Key: Enter the personalized key you were provided by your AI service provider.
URL: Enter the URL of a locally running LLM server. This option will only appear as needed.
Model: Choose from the list of models for a specific LLM, e.g., Gemini Flash 8B. Each model may show one of several icons showing its capabilities: for reasoning, for vision, for tooling, for deep search, for coding support, and for cost, with the icon's boldness indicating higher or lower costs.
Context Window: This is the number of tokens the LLM can process and "remember" in a conversation. A larger context window means more data is passed or remembered. However, if you're trying to run a local LLM, larger context windows use more RAM. This also displays how many tokens have been used, sending and receiving responses. Bear in mind, this setting can be set per model as each has its own context window size.
Usage: Choose an option to balance cost and quality of results, from fewer tokens with less precision to more tokens but a higher chance of useful results.
Role: Define an optional default "persona" or instructions for the AI, e.g., "You are an undergraduate professor presenting to your class. Use Markdown formatting with sections and subsections but no lists. Include links to your sources." This is used in automation, like AI assisted scripting.
Assistant: Certain AI models have access to "tooling" and may be able to accept DEVONthink-related commands. You need to decide what behaviors you will allow it to use on your Mac and with your databases.
Allow property & content changes: Decide whether the chat assistant can make changes to your database, e.g., add tags to a document or create a new one for you.
Allow MCP tools: Give the Chat assistant access to DEVONthink's own MCP server.
Allow download of web pages: Allow AI to download the content of web pages to examine as part of its querying process. This may be considered an insecure option.
Allow screenshots of window: Allow the AI assistant to capture and examine a screen capture of DEVONthink's window for use in queries. Requires a compatible AI model, e.g, Claude Sonnet.
Load remote images automatically: Control whether linked images in web content is downloaded for examination. This may be considered an insecure option.
Max. Recent Chats: Set the number of remembered chats you can revisit in the Recent Chats popup in the Chat assistant.
Library
This pane is where you can define your own custom repository of roles, prompts, and skills to use with external AI. From creating "personalities" like subject matter experts to having a ready-made set of prompts you often use to going beyond simple inquiries with skills built from more complex instruction sets, there are things to explore and opportunities to push the limits with AI and DEVONthink. The interface is divided into a "bookshelf" of several built-in options and – where you'll add your own – and the controls to create and modify them.
Function List: On the left is the list of all your AI functions. Select one to see its controls and parameters on the right. Add or remove functions with the plus and minus buttons at the bottom of the pane. The button has commands for creating new functions of each type: Prompt, Role, or Skill. There are also these commands: Duplicate, Rename, Import, Export, and Delete. If you delete some of the built-in functions, you can restore them via Add Default Items.
Controls: When editing an AI function, you will use the controls on the righthand side. These let you define what you want to happen, when you want it to happen, and what level of autonomy you give AI in accomplishing the tasks. Here are the other available controls:
Keywords: Also known as slash commands, are abbreviations for quickly using one of your library items. When you type a forward slash, e.g., in the Chat assistant, a popup will show possible function matches so you can quickly run one with a few taps.
Command Type: Choose the type of function, Prompt, Role, or Skill
Use when: This is a description of the AI function, the "when" or "why" is used.
Instructions: Define the commands, parameters, and requirements sent to and used by AI.
Context: Define what the function applies to. This can be nothing, selection, or items the Chat assistant is handling. A selection can be selected text or an item. Selected items can have their text, text and images, or you can let the Chat assistant decide the appropriate use.
Tools: Choose how much AI can do with your data, either Nothing, Read only, Read & organize, Read & edit contents, or Read, organize & automate. The tooltip under the popup shows specific information about the scope of its action.
Model: Choose a function-specific model, helpful if you have lighter tasks requiring less expensive models, or use the default chosen in the AI > Chat settings.
Options: One last suite of settings are specific options for the Chat assistant. Let the assistant use this on its own determines if AI can determine and use this as it seems appropriate. Ask before automatic use requires AI to ask permission before using a function. And Isolated run doesn't use the chat history when using the function.
The MCP settings provide controls for setting up DEVONthink's MCP server, allowing connecting to and working with your databases while outside our application.
Privacy: Enable Redact sensitive content to remove or obscure recognizable private information like credit card numbers, email addresses, phone numbers, etc. before the content is sent to the AI model.
MCP Server: If you are integrating with Claude Code or other HTTP-based MCP clients, enable Launch at login. Note that Claude Desktop and Cowork launch their own instance of the server as they need it.
Port: The port the server is using, even when running locally. A default port is assigned but can be changed to a different unused port, as needed.
URL: Select and copy this URL to use in other applications or on other devices accessing the MCP server.
Accept connections: Choose From the Mac only to restrict access only to local use, From the local network to connect to devices on your network, or Any device to allow even remote connections. The last two options require a valid TLS certificate and bearer token.
TLS Certificate: Choose an existing valid TLS certificate you have installed. Note that self-signed certificates aren't supported.
Bearer Token: Connecting other devices requires a bearer token for authentication. Press the Create button to generate one specifically for DEVONthink's MCP server.
Install Connectors: If you are chatting or using Cowork in Claude Desktop, click Claude Desktop. If you have installed the Claude Code command-line interface (CLI), enable Claude CLI. For OpenAI's Codex access, enable Codex Desktop & CLI. To use the Hermes Agent, including via Ollama, click Install for Hermes Agent. If you're using LM Studio, enable LM Studio. Each option will only be enabled if the respective application is installed.
Copy JSON: Click this button to copy the underlying -stdio connection information as JSON. This can be added to the configuration files of compatible MCP applications.
Log: To access information about what's happening with the MCP server, click the Log button.
The Search settings control where you want external AI to search. For example, when researching medical issues, you may want to let it search PubMed but not Wikipedia. The options are: in your Databases, on the arXiv, PubMed or Wikipedia websites, or on the Web, in general.
Regarding web and Wikipedia searching, you can also choose a specific engine for deep research: DEVONthink, Perplexity, Exa AI, or ollama.com for web searches. The commercial services provide quick results but utilize indexed content. DEVONthink's results are returned more slowly but are up to date. Also, be aware the commercial services require an account and an API key.
Summarization
Model: Choose a specific AI provider and model for summarization or Default to use the Chat model.
Language: Choose a specific language to use for the summary or leave it as Automatic to use the system's primary language.
Style: Determine what summary format you'd like in response to asking chat to summarize a document. The choices are:
Text: Gives you a brief synopsis in a few paragraphs.
Bullet Points: Returns a list of the main points.
Key Points: Provide a distilled response of the main topics.
Table: Create a table of columns and rows, often used for correlating pages or links to text.
Custom: Provide a summary defined by a template you define in this settings pane.
Custom Prompt: Create your own prompt defining what kind of response you'd like, including how you'd like the summary to be structured. Use the special placeholder %@ to refer to the information being summarized.
Image Generation
Choose and set up a text-to-image AI model. These controls are dynamic and their options change depending on the model you choose.
Aspect Ratio: Choose a predefined aspect ratio for the images. Supported sizes are: 1:1, 16:9 (9:16), and 4:3 (3:4), depending on the chosen model.
Quality: Decide whether to generate Standard or HD images, if available.
API Key: Enter the API key you received from the image generation provider, e.g., Replicate.com for the Flux generator.
Transcription
AI speech-to-text processes incoming media files and processes them per these settings. For example, an .mp3 file could be transcribed into a separate annotation file for future use.
For each type of text recognition, images or audio/video, choose where to store the recognized text via the Destination dropdown. The available options are:
Searchable Text: This is similar to Apple's Live Text feature in that a text layer isn't added to the document, but instead is stored in the database's index and associated with the file.
Annotation: Create an annotation file with the transcribed text.
Comment: Add the transcribed text as a Finder comment on the file.
Images: Decide what live OCR engine you want to process images added to your database:
Fast Apple Vision text recognition: Quickly detect text in images using Apple's Vision framework. Often sufficient for many use cases.
Accurate Apple Vision text recognition: Detect text in images with an emphasis on accuracy over speed.
Text recognition via chat: Uses your chosen Chat model to detect text in images, provided the model supports image analysis. When enabled, you can choose a vision-capable AI model, if desired.
Audio & Video: Choose the transcription engine you want to process media files added to your database:
Local Apple SpeechAnalyzer transcription: Transcribe audio on your Mac with Apple's SpeechAnalyzer framework.
for reasoning, 
for vision, 
for tooling, 
for deep search, 
for coding support, and 
for cost, with the icon's boldness indicating higher or lower costs.
button has commands for creating new functions of each type: Prompt, Role, or Skill. There are also these commands: Duplicate, Rename, Import, Export, and Delete. If you delete some of the built-in functions, you can restore them via Add Default Items.
