Arc and Slice AI Architecture
From Arc and Slice 6.0.0, the Arc Assistant and the Slice Assistant share the same AI layer. This page explains how the pieces fit together: where users talk to the assistant, what Arc does in the middle, which tools the assistant can use, and which controls decide what it can see and change.
There is no AI built into Arc or Slice. The assistant only works once an administrator connects your own LLM provider in settings.yml. The quality of the answers depends on the LLM you choose.
Architecture diagram
CLIENTS
Arc Assistant
Chat panel in Arc (browser)
Slice Assistant
Chat panel in Excel
SOFTWARE
Arc
Runs the tools, applies the controls and sends the conversation to your LLM
Your LLM
OpenAI-compatible provider, Ollama, Hugging Face or WatsonX
AI TOOLS
TM1 AI
TM1 objects and cube data, through the TM1 REST API
Slice AI
Workbooks, Active Forms and Slice features in Excel
Pulse AI and other MCP servers
Pulse data and external tools, through MCP
Components
Clients
Arc Assistant: the chat panel in Arc. It can see the page you have open (currently the process editor, cube viewer and rule editor) so you do not have to explain what you are working on.
Slice Assistant: the same assistant, opened from a panel in Excel. It gets tools for Excel and Slice as well as the Arc tools. See Enabling the Arc Assistant in Slice.
Both assistants are windows onto the same interface: they go to the same LLM and the same controls apply.
Arc and your LLM
Arc sits between the user and the LLM. For each message it adds the relevant context (open page, chat history, tool results), sends it to the LLM configured in settings.yml, runs the tools the LLM asks for and returns the answer, rendered as Markdown in the chat panel. See Enable the Arc Assistant.
AI tools
Group | What it covers | Examples |
|---|---|---|
TM1 AI | Discovering TM1 objects, reading cube views and, in Edit mode, making changes. Every call goes through the TM1 REST API with the user's own login, so TM1 security always applies. |
|
Slice AI | Reading and writing Excel ranges, building and changing Active Forms, running Slice refresh and audit features. Any tool that changes an Active Form asks the user to confirm first. |
|
Pulse AI and other MCP servers | Arc acts as an MCP client. Each server declared in |
|
The full Slice tool list is on Slice Assistant Tools Overview and the Pulse tools on Pulse MCP Tools.
How a request flows
The user asks a question in the Arc or Slice Assistant, for example "Where can I enter salary data for this employee?".
Arc adds the context allowed by the current mode and sends it to your LLM.
The LLM asks Arc to call the tools it needs (TM1, Slice or MCP). Arc checks the controls, runs the allowed tools and returns the results.
The LLM writes the answer, or in Edit mode proposes and applies a change.
Controls
User controls (per chat)
Mode | What the LLM can do |
|---|---|
Chat only | Nothing in Arc. The LLM sees only what you type, like a normal chat. |
Read only | Uses read-only tools to look at what you are working on. It cannot change anything. |
Edit | Can also use tools that change TM1 objects or data on your behalf. |
Choose whether to include only the metadata of a view or also its data.
Review the exact data packet that will be sent to the LLM before you send a message.
The chat, read and edit modes apply to the Arc tools. They do not apply to MCP servers such as Pulse, which have their own permissions (Pulse tools are read-only).
Administrator controls (settings.yml)
Setting | Purpose |
|---|---|
| No tool calls at all: conversational use only. |
| Turn off or limit TM1 object discovery, and strip properties such as TI code or rules. |
| Turn off or cap reading of cube view data. |
| Redact any element in any dimension (regular expressions supported) so its values never reach the LLM, or block writes to it. |
| Block every external MCP server. |
| Limit how many tool calls the assistant can make in one turn. |
Set up redaction early. Unlike Pulse, which only shares metadata, the Arc Assistant can send TM1 cell values to your LLM. Blacklist sensitive data (for example salaries) as soon as an LLM is configured, and check your LLM provider's data retention policy.
Arc and external MCP clients
Arc is not currently exposed as an MCP server to external clients such as Claude Desktop, because the chat, read and edit controls could not be enforced from outside Arc. Use the Arc or Slice Assistant for Arc tasks, and connect external clients to Pulse instead (see Pulse AI Architecture).