Chat panel

The optional browser chat panel can use the same operations as an MCP client. Open it beside the 3D model when you want one browser workspace for questions, tool calls, and visual results.

The panel is off by default. It is the only ifc-console feature that may contact an external LLM provider.

Open the panel

Install the viewer bundle first:

pip install "ifc-console[viewer]"

Then use:

/chat          open chat beside the 3D viewer
/chat solo     open chat without the 3D canvas
/chat off      close chat and forget any in-memory key

In split view, press ++c++ to show or hide the panel and drag the divider to resize it.

Choose a provider

provider	credential	notes
OpenAI	OPENAI_API_KEY	choose a model available to the account
Anthropic	ANTHROPIC_API_KEY	native Claude tool use
OpenRouter	OPENROUTER_API_KEY	one endpoint for many model providers
Local	none by default	OpenAI-compatible vLLM, LM Studio, Ollama, or similar

Choose the provider and model from panel settings. Use Custom ID when the model does not appear in the fetched list. /chat anthropic can set the provider from the terminal.

For a local provider, set its base URL, for example http://localhost:8000/v1. Enable chat.local_only=true to reject any provider URL outside the local machine.

Credentials and model data

If the normal environment variable is set, the console reads the key from its own environment. Otherwise, you can enter a key in the panel. The browser sends it directly to the running console and does not store a copy. The console keeps it only in memory until /chat off or exit.

The browser never contacts the provider directly. ifc-console sends the provider:

• your messages;
• the chat system instructions;
• tool results used to answer, which may contain IFC model data.

Use a local provider or keep the chat panel disabled for confidential models that must not leave the machine.

Tools and safety

With tools enabled, the assistant can query elements, inspect properties, validate models, calculate quantities, search the knowledge index, and use any other operation permitted by the current session.

The same controls apply as with MCP:

• ask mode blocks model changes;
• only the terminal user can switch to edit;
• each tool call and result appears below the answer;
• every call is audited with provider and model metadata, never the key.

Turn chat.tools off when you want ordinary conversation without model tools.

Panel controls

• Enter sends; Shift+Enter adds a line.
• Escape closes settings or stops a running answer.
• Stop cancels the current provider response.
• New chat clears the conversation in that browser tab.
• Reloading keeps the tab's conversation; closing the tab drops it.
• Answers show tool calls, token counts, latency, and copy controls.

Provider, model, and mode are always visible above the conversation.

Python

The SDK exposes the same provider-neutral loop:

from ifc_console import Workbench

with Workbench.open("tower.ifc") as wb:
    answer = wb.ask(
        "Which walls are missing a fire rating?",
        provider="anthropic",
        model="YOUR_MODEL_ID",
    )
    print(answer["text"])
    print(answer["tool_calls"])

Pass on_event=print to stream events. Without api_key=, the provider key comes from the environment. See Python SDK.

Settings

key	default	purpose
chat.enabled_default	false	open chat at session start
chat.provider	openai	initial provider
chat.model	empty	initial model ID
chat.base_url	empty	provider URL override
chat.tools	true	expose permitted ifc-console tools
chat.max_tool_rounds	8	maximum tool rounds per answer
chat.local_only	false	allow only local provider URLs
chat.timeout_s	300	provider response timeout

Use ifc-console --chat to enable the panel at startup.