Skip to content

Connecting a model

heft21 edited this page Nov 16, 2026 · 1 revision

The assistant needs a model with tool calling, served on an OpenAI-compatible endpoint. Blueprints and guides work without one.

The setting

In Drydock, open Settings and fill in:

Field What to put
Model server address the endpoint's URL, ending in /v1 (default http://localhost:1234/v1, LM Studio's)
Model pick one from the list; Drydock asks the server which models it has
API key only for hosted APIs; local servers need none
Step limit how many tool calls one answer may take (default 30)

The model shows at the top right of every page when it is set.

LM Studio

  1. Download a model with tool calling. Qwen3.5 9B is a good start on a GPU with 12 GB or more; see Eval results.
  2. Load it with a context length of 16,384 tokens or more. Drydock's tool list alone is about 7,000 tokens, so a smaller window fills before the work starts. Check the context length in the model's load settings rather than trusting the default.
  3. Start the server from the Developer tab. Its address is http://localhost:1234/v1, which is already Drydock's default.

llama.cpp

Start llama-server with the chat template turned on, which is what makes tool calls work:

llama-server -m <model.gguf> -c 32768 -ngl 99 --jinja --port 1234

A hosted API

Any OpenAI-compatible endpoint with tool calling works: put its /v1 URL in the address field and your key in the API key field.

Which models work

Drydock is built to work with small local models, but not every model calls tools the same way:

  • Qwen3.5 9B makes real tool calls and passed 7 of 18 measured tasks.
  • Qwen2.5-Coder 14B writes its tool calls as JSON text in its answer instead of making them, so nothing runs: 0 of 18.

The numbers and what failed are on Eval results.

Clone this wiki locally