ParaCrux

Available Deeper · 4 min

Run the Agent with a local model

Point the Agent at a model server on your own computer or network, so design context never leaves your site.

Video in production

Until it is ready, this is the running order. Each step is also written out below, and you can try it in the simulator.

  1. Server

    A model server running on the same computer.

  2. Connect

    Agent settings; Local model · offline; Endpoint and Model; Save connection.

  3. Ask

    A question in a design conversation; the label beside Send; the wait message.

You need

  • The desktop app
  • A model server with an OpenAI-compatible Chat Completions API on this computer or your network, such as Ollama, llama.cpp or LM Studio

You'll come away with

  • Connect the Agent to your own model
  • Know what is sent, and where
  • Know where it stops

Step by step

  1. Start your model server

    Run it on this computer or another on your own network. The app accepts only local addresses (localhost, 127.x, 10.x, 172.16 to 172.31, 192.168.x), over http or https, with no user name or password in the address.

  2. Connect it

    Open Agent settings. Under Connect using, choose Local model · offline. Enter the Endpoint, such as http://127.0.0.1:11434/v1, and the Model exactly as your server lists it, then choose Save connection. There is no API key.

  3. Check the label

    Beside Send, the connection shows Local model and the model's name.

  4. Know what is sent

    Each step sends the conversation and the Agent's tool list to your endpoint. Nothing goes to CueCrux or OpenAI, and no AI credit is used.

  5. Seats still apply

    The Agent uses the same SOLIDWORKS tools as with any connection. Reading the open model needs no seat; changes need a valid seat, as always. On a computer that is never online, that means an offline licence.

  6. Expect it to be slower

    A step can take minutes: the app waits up to 5 minutes for each and stops after 8. The hosted design matrix is not offered. Local models vary, so check the answers.

Try it here.

The labels and messages are the app's own. The endpoint and model name are the app's placeholders.

Desktop app · Agent settings simulated · step 1 of 4

  1. Agent settings

    Connect using
    Local model · offline
    Endpoint
    http://127.0.0.1:11434/v1
    Model
    qwen2.5:14b-instruct

    Runs the product's own code

What it won't do

  • The app does not install or run the model server for you.
  • Only addresses on this computer or a private network are accepted.
  • At most 8 steps per job and one tool call per step, with no automatic retry.
  • The hosted design matrix is not offered with a local model.
  • Answer quality depends on the model you run. Treat CAD results as only what the tools record.

If you see a message

Covers: Local model connection. See the full index