Provider Configuration

Q: How do I configure Azure OpenAI with Humanbound?

Set `HB_PROVIDER=azureopenai`, provide your Azure API key via `HB_API_KEY`, and set `HB_ENDPOINT` to the full deployment URL including `?api-version=`, for example `https://your-resource.openai.azure.com/openai/deployments/your-deployment/chat/completions?api-version=2025-01-01-preview`.

Q: Can I use OpenRouter or another OpenAI-compatible API?

Yes. Set `HB_PROVIDER=openai`, put that service's key in `HB_API_KEY`, and set `HB_ENDPOINT` to its base URL, for example `https://openrouter.ai/api/v1`. `HB_MODEL` takes the service's own model ID, such as `anthropic/claude-haiku-4.5` on OpenRouter. The same works for LiteLLM, vLLM and other servers that expose `/chat/completions`.

Q: Can I run Humanbound with no external API calls at all?

Yes — use Ollama. Set `HB_PROVIDER=ollama` and `HB_MODEL=llama3.1:8b`, start `ollama serve`, and tests will only call your bot and the local Ollama instance. Note that local models produce lower-quality attacks than cloud providers.

The local engine needs an LLM provider for attack generation and response evaluation, and you bring your own API key. Provider settings resolve in order — CLI flags override environment variables, which override the config file at ~/.humanbound/config.yaml. Six providers are supported (OpenAI, Anthropic Claude, Google Gemini, Azure OpenAI, Grok, and Ollama for fully-local isolation); Azure requires a full endpoint URL with ?api-version=.

Configuration Methods

Provider is resolved in this order (first match wins):

CLI flags (one-off override)
Environment variables (CI/CD)
Config file (~/.humanbound/config.yaml)

Environment Variables

export HB_PROVIDER=openai
export HB_API_KEY=sk-proj-...
export HB_MODEL=gpt-4.1        # optional for openai and ollama only (see below)

HB_MODEL (or hb config set model …) is optional only for openai (default gpt-4.1) and ollama (default llama3.1:8b). Every other provider needs an explicit model; for Azure OpenAI that is your deployment name.

Config File

hb config set provider openai
hb config set api-key sk-proj-...
hb config set model gpt-4.1

# View current config
hb config

Config is stored at ~/.humanbound/config.yaml. Never sent to Humanbound.

Environment variables and the config file don't mix

When HB_PROVIDER is set, the config file is not read at all, so an endpoint or model saved with hb config set is ignored. Set everything through environment variables, or everything through the config file.

Supported Providers

Provider	`HB_PROVIDER`	Key prefix	Notes
OpenAI	`openai`	`sk-`	Default model `gpt-4.1`. Optional `HB_ENDPOINT` for OpenAI-compatible APIs
Anthropic	`anthropic` or `claude`	`sk-ant-`	`HB_MODEL` required
Google	`gemini`		`HB_MODEL` required
Azure OpenAI	`azureopenai`		Requires `HB_ENDPOINT` with `?api-version=`; `HB_MODEL` = deployment name
Grok (xAI)	`grok`		`HB_MODEL` required
Ollama	`ollama`	Not needed	Full local isolation; default model `llama3.1:8b`

Azure OpenAI

Azure requires the full endpoint URL including the api-version:

export HB_PROVIDER=azureopenai
export HB_API_KEY=your-azure-key
export HB_MODEL=your-deployment   # required: the Azure deployment name
export HB_ENDPOINT="https://your-resource.openai.azure.com/openai/deployments/your-deployment/chat/completions?api-version=2025-01-01-preview"

OpenRouter and other OpenAI-compatible APIs

The openai provider accepts a base URL in HB_ENDPOINT, so any service that speaks the OpenAI chat completions API works, including OpenRouter, LiteLLM and vLLM:

export HB_PROVIDER=openai
export HB_API_KEY=sk-or-...                      # the service's key, not an OpenAI key
export HB_ENDPOINT=https://openrouter.ai/api/v1  # base URL; /chat/completions is added
export HB_MODEL=anthropic/claude-haiku-4.5       # the service's model ID

With no HB_ENDPOINT, requests go to api.openai.com as before. If you switch from Ollama to OpenAI with hb config, clear the old endpoint with hb config set endpoint "", or requests will go to the Ollama URL. An HB_ENDPOINT left in your shell does the same thing, so unset HB_ENDPOINT too.

Ollama (Full Isolation)

For zero external network calls — everything runs locally:

# Install
pip install "humanbound[engine]"

# Start ollama
ollama serve
ollama pull llama3.1:8b

# Configure
export HB_PROVIDER=ollama
export HB_MODEL=llama3.1:8b   # optional, this is the default

# Run test (only calls: your bot + local ollama)
hb test --endpoint ./config.json --scope ./scope.json --wait

Ollama quality

Local models produce lower-quality attacks and evaluations than GPT-4 or Claude. For best results, use a cloud provider. Use ollama when isolation is more important than accuracy.

When logged in, every Humanbound account includes an LLM provider — no external API key required. Tests run on the platform automatically use it:

hb login
hb connect --endpoint ./config.json
hb test --wait
# Uses Humanbound's LLM provider — no HB_PROVIDER or HB_API_KEY needed

You can still use your own provider on the platform by adding it via hb providers add.

Frequently asked questions

Which LLM providers does Humanbound support?

Humanbound supports OpenAI, Anthropic (Claude), Google (Gemini), Azure OpenAI, Grok (xAI), and Ollama. Ollama requires no API key and runs fully locally.

How do I configure my LLM provider?

Provider configuration is resolved in order — CLI flags first, then environment variables (e.g., HB_PROVIDER, HB_API_KEY), then the config file at ~/.humanbound/config.yaml. The config file is set via hb config set provider and hb config set api-key.

How do I configure Azure OpenAI with Humanbound?

Set HB_PROVIDER=azureopenai, provide your Azure API key via HB_API_KEY, and set HB_ENDPOINT to the full deployment URL including ?api-version=, for example https://your-resource.openai.azure.com/openai/deployments/your-deployment/chat/completions?api-version=2025-01-01-preview.

Can I use OpenRouter or another OpenAI-compatible API?

Yes. Set HB_PROVIDER=openai, put that service's key in HB_API_KEY, and set HB_ENDPOINT to its base URL, for example https://openrouter.ai/api/v1. HB_MODEL takes the service's own model ID, such as anthropic/claude-haiku-4.5 on OpenRouter. The same works for LiteLLM, vLLM and other servers that expose /chat/completions.

Can I run Humanbound with no external API calls at all?

Yes — use Ollama. Set HB_PROVIDER=ollama and HB_MODEL=llama3.1:8b, start ollama serve, and tests will only call your bot and the local Ollama instance. Note that local models produce lower-quality attacks than cloud providers.