Ask CAST Imaging
Overview
Ask CAST Imaging is currently provided in beta and is flagged as such in the interface. Review its answers carefully before relying on them.
Ask CAST Imaging provides AI-powered agents that help you explore and understand an application in plain language: how it is structured and what depends on what, which quality and security issues it has, and what a change would affect.
It differs from the CAST Imaging Assistant chatbot in the way it reaches an answer. Rather than generating a reply directly, the agent works out which tools it needs to answer your question, runs them, and uses their results to build its response. That activity is shown in the conversation, so you can see how your question is being handled.
Not every question needs a tool: for a simple one, the agent may answer directly.
In 3.6.8-funcrel, this feature was called CAST Imaging Agents.
Before you start
The following are required:
- You hold the
AI Features Accessrole - see Default profiles and roles. - An AI provider is configured - see AI Settings. For the models CAST has tested and how well each one works with the agents, see AI model compatibility.
- The Assistant chatbot option is enabled in the list of AI features in AI Settings. The same option enables the CAST Imaging Assistant chatbot.
- The CAST Imaging MCP Server is installed and running - the agent uses it to run its tools. See Install the CAST Imaging MCP Server.
- Updated in ≥ 3.6.9-funcrelThe CAST Imaging MCP Server is release 3.1.0-funcrel or later. Of the MCP servers it provides, see Architecture:
- The structural MCP server is mandatory.
- The semantic search and Text2Cypher MCP servers are optional. Without them, the agent works with the tools that remain available, and does not call the tools of a server that is unavailable.
Semantic search, which draws on the GraphRAG knowledge graph, needs a license key that includes GraphRAG - see Manage license keys. Without it, the agent still works using its other tools, but some answers may be less detailed. A warning is displayed above the input box:

Click Details to see the status of each MCP server the agent uses. imaging_semantic_mcp is shown as Not licensed:

Close the warning to hide it.
MCP server checks
When the chat window opens, Ask CAST Imaging checks the MCP servers it uses and displays a message when one of the requirements listed in Before you start is not met:
| Situation | Effect |
|---|---|
| The MCP Server release is older than 3.1.0-beta4 | A message asks you to update the MCP Server and links to the documentation. You cannot type in the input box until the MCP Server is updated. |
| No MCP server is available | You cannot type in the input box. |
| The structural MCP server is unavailable | You cannot type in the input box. |
| The semantic search or Text2Cypher MCP server is unavailable | A warning is displayed above the input box. You can still ask questions: the agent works without that server’s tools, and some answers may be less detailed. |
Run /status to see the status of each MCP server - see Specialist agents and commands.
Open the chatbot
- Open the application you want to ask about.
- Select Ask CAST Imaging to open the chat window.

The chat window opens full screen.

In 3.6.8-funcrel, this button is labelled CAST Imaging Agents. It replaces the CAST Imaging Assistant button that occupied the footer in releases before 3.6.8-funcrel. The assistant is still available and is now opened from the Action toolbar in the graph view - see CAST Imaging Assistant chatbot.
Ask a question
Type your question into the input box in plain language.
While the agent works, it reports what it is doing in the conversation, including the tools it calls. These sections are expanded by default; collapse one to hide the detail, and expand it again to review it.

Stop a response
Select Stop while a response is being generated to halt it.
To pick up where the conversation stopped, send continue as your next message. The agent resumes from that point.
Chat window options
| Option | Description |
|---|---|
| Export as Markdown | Exports the entire session to an .md file. |
| Dock | Docks the chat window to the right of the screen, so you can consult the application and the conversation side by side. |
| Minimize | Minimizes the chat window. The conversation remains available. |
| Close | Closes the chat window and ends the session - see Sessions. |
Specialist agents and commands
Type / in the input box, or select the / button there, to choose a specialist agent or a command. A specialist agent focuses on one kind of question; a command acts on the chat itself.
| Agent or command | Use it to |
|---|---|
/status |
Check the health of the agent services and confirm that the services it needs are available. |
/quality-advisor |
Get focused information about the quality and security of the application, with recommendations and a plan for addressing what it finds. |
/impact-analysis |
Find out what could be affected, or break, when an object is changed. |
/structural-discovery |
Understand how the application is put together and how its components are connected. |
AI model compatibility
Ask CAST Imaging uses the AI provider and model configured in AI Settings. The agents rely on tool calling, so the model you choose has a direct effect on the quality of the answers. CAST tested the models below on the same task, a full impact analysis, and rated how well each one works with the agents.
| Rating | Meaning |
|---|---|
| Recommended | Reliable, detailed and accurate answers |
| Works well | Good answers; some detail may occasionally be missing |
| Works with limits | Usable, but answers can be short, slow or incomplete |
| Not recommended | Unreliable or inaccurate results |
| Not yet verified | Not fully tested |
In summary:
- Best overall: Claude Opus 5.5
- Best balance of speed and quality: Claude Sonnet 5.5 and Gemini 3.8 Flash
- Best on OpenAI: GPT-6 Astra or GPT-6.1 Sol
- Best on Azure OpenAI: GPT-5.4
- On AWS Bedrock: Amazon Nova Pro or Amazon Nova 2 Lite - check long answers for completeness
- Avoid Gemini 3.5 Flash Lite and Amazon Nova Lite, and the slowest “Pro” models for everyday work
Smaller “Lite”, “Mini” and “Flash Lite” models are faster and cheaper, but give shorter and less reliable answers for a detailed impact analysis.
| Model | Rating | Comments |
|---|---|---|
| Claude Opus 5.5 | Recommended | Best default choice: very thorough, and clear about what it could not check |
| Claude Fable 5.1 | Recommended | The most precise answers. Slower; suited to complex analyses |
| Claude Sonnet 5.5 | Recommended | Best balance of speed and quality |
| Claude Haiku 5.5 | Not yet verified | Fast, but long analyses do not complete |
| Claude Opus 4.6 | Works well | Good results, but parts of the report are not shown in the chat window. Use a newer model |
| Claude Sonnet 4.6 | Works with limits | Slow, and some figures can be wrong. Use a newer model |
| Model | Rating | Comments |
|---|---|---|
| GPT-6 Astra | Recommended | The most complete OpenAI answers, with clear, practical guidance |
| GPT-6.1 Sol | Recommended | Close to GPT-6 Astra in quality, at lower cost |
| GPT-5.2 | Recommended | Clear, well-organized answers with a practical testing plan |
| GPT-5.6 Luna | Works well | Accurate, with a useful list of what to test first |
| GPT-5.6 Sol | Works well | Good guidance on how to make changes safely |
| GPT-5.6 Terra | Works well | Good overview, with code quality findings |
| GPT-5.4 | Works well | Good, practical answers. Slightly slower |
| GPT-5.2 Pro | Works well | Good answers. Slightly slower |
| GPT-5.5 | Works with limits | Good overall, but some sections of the report can be empty |
| GPT-6 Sol | Works with limits | Correct, but brief |
| GPT-5.4 Mini | Works with limits | The fastest option. Good for quick checks, less accurate for a full analysis |
| GPT-5.5 Pro | Works with limits | Very thorough, but too slow for everyday use (about 30 minutes) |
| GPT-5.4 Pro | Not yet verified | Expected to work |
| GPT-6 Luna | Not yet verified | - |
| Model | Rating | Comments |
|---|---|---|
| Gemini 3.8 Flash | Recommended | Best Gemini choice: complete, well-organized answers |
| Gemini 3.7 Flash | Recommended | Same as Gemini 3.8 Flash |
| Gemini 4 Argon | Works well | The most thorough Gemini answers, but slow |
| Gemini 3.6 Flash | Works with limits | Correct facts, but some sections can be empty |
| Gemini 3.5 Flash | Works with limits | Same as Gemini 3.6 Flash |
| Gemini 3.1 Pro (Preview) | Works with limits | Good, but some sections can be empty. Preview model |
| Gemini 3 Flash (Preview) | Works with limits | Fine for a quick summary, too brief for a full analysis. Preview model |
| Gemini 3.5 Flash Lite | Not recommended | Fast, but parts of the answer can be inaccurate |
| Gemini 3.1 Flash Lite | Not yet verified | Reported to work in earlier tests |
| Model | Rating | Comments |
|---|---|---|
| Amazon Nova Pro | Works with limits | Usually correct, but long answers can be cut off and results vary between runs |
| Amazon Nova 2 Lite | Works with limits | Finds the right target, but sometimes ends without a report or repeats example figures |
| Amazon Nova Lite | Not recommended | Unreliable: sometimes fails, or cannot find objects that exist |
OpenAI models can also be used through Azure OpenAI.
| Model | Rating | Comments |
|---|---|---|
| GPT-5.4 | Works well | Accurate, consistent results with clear advice. Best choice on Azure |
| GPT-5.4 Mini | Works with limits | Fast, but results can differ between runs |
| GPT-5 Nano | Works with limits | Finds the right target, but answers are less precise. Being retired |
Token usage and tool calls
The end of each response reports the number of tokens consumed and the number of tool calls made during the conversation.

Sessions
Closing the chat window ends the current session. CAST recommends starting a new session rather than continuing a long conversation, since responses are better in a fresh one.