Strix setup: an OpenAI-compatible base URL and a model id
Strix takes a custom model with STRIX_LLM set to openai/<model-id> and LLM_API_BASE set to that API's base URL.
Set STRIX_API_TYPE=chat_completions when the endpoint serves that API. The same variables run Kimi K3.1 Pro or DeepSeek V4.1 Flash.
When Strix is the right agent
Strix tests a web app, an API or a source tree in a Docker sandbox, and its README calls the agents "autonomous AI hackers that find and fix your app's vulnerabilities." (Strix README) Kimi K3.1 Pro is the pick. Strix's local guide says "Most local models, especially those under 70B parameters, struggle with these complex tasks." (Strix docs)
Point Strix at your model
Variable names from the configuration reference, read November 18, 2026.
Install Docker, then Strix
The README requires Docker and an LLM key. Strix README.
curl -sSL https://strix.ai/install | bashSet the model, the base URL and the API type
STRIX_LLM, LLM_API_BASE and LLM_API_KEY are the three variables, and examples end the base URL in /v1. Configuration reference. Set STRIX_API_TYPE=chat_completions because "A custom LLM_API_BASE does not change this, so set STRIX_API_TYPE=chat_completions when your gateway only serves chat completions."
export STRIX_LLM="openai/kimi-k3.1-pro-uncensored" export LLM_API_BASE="https://api.silkcompute.com/v1" export LLM_API_KEY="$SILK_API_KEY" export STRIX_API_TYPE="chat_completions" export STRIX_REASONING_EFFORT="high" export STRIX_LLM_INPUT_COST_PER_MTOK="3.00" export STRIX_LLM_OUTPUT_COST_PER_MTOK="8.00" export STRIX_LLM_CACHED_INPUT_COST_PER_MTOK="0.3" strix --target ./app-directory
Model id kimi-k3.1-pro-uncensored. Get a key.
Tell Strix what the tokens cost
An unset price counts as 0, so set both if --max-budget should mean anything. The numbers are Silk Compute list prices per 1M tokens, read November 18, 2026.
Set reasoning effort
STRIX_REASONING_EFFORT accepts none, minimal, low, medium, high, xhigh and max, and defaults to high. Use max for attack-path analysis. Configuration reference.
Name one in-scope target
Pass a directory or a URL you are allowed to test.
strix --target ./app-directory --instruction "Authorized test of this repository only. Stay inside the hosts listed in SCOPE.md."
Only on systems you own, a signed scope, a CTF or a lab. A custom model does not widen the scope.
A model on your own machine
Ollama uses a different prefix and no /v1. From the local models guide.
export STRIX_LLM="ollama/<your-model-tag>" export LLM_API_BASE="http://localhost:11434" strix --target ./app-directory
Raise Ollama's num_ctx to "at least 16k–32k", or the tool definitions get cut off. Weights that fit one GPU are on local models.
A prompt that stays inside the scope
Authorized test. Read SCOPE.md, list the in-scope hosts, the exclusions and the rules, then propose the order of review. Do not write exploit code. Stop if a check is not named in the scope.