03 Providers

LLM Providers

nitpik is bring-your-own-model. You choose the LLM provider and supply your own API key — nitpik never proxies, stores, or meters your API calls.

---

Supported Providers

Provider	NITPIK_PROVIDER value	Provider-specific env var
Anthropic	anthropic	ANTHROPIC_API_KEY
Azure OpenAI	azure	AZURE_OPENAI_API_KEY
Cohere	cohere	COHERE_API_KEY
DeepSeek	deepseek	DEEPSEEK_API_KEY
Galadriel	galadriel	GALADRIEL_API_KEY
GitHub Models	github	GITHUB_TOKEN
Google Gemini	gemini	GEMINI_API_KEY
Groq	groq	GROQ_API_KEY
HuggingFace	huggingface	HUGGINGFACE_API_KEY
Hyperbolic	hyperbolic	HYPERBOLIC_API_KEY
Mira	mira	MIRA_API_KEY
Mistral	mistral	MISTRAL_API_KEY
Moonshot	moonshot	MOONSHOT_API_KEY
Ollama	ollama	(none — runs locally)
OpenAI	openai	OPENAI_API_KEY
OpenRouter	openrouter	OPENROUTER_API_KEY
Perplexity	perplexity	PERPLEXITY_API_KEY
Together	together	TOGETHER_API_KEY
xAI (Grok)	xai	XAI_API_KEY
OpenAI-compatible	openai-compatible	OPENAI_API_KEY

Third-party dependency notice: Provider integrations are powered by a third-party open-source library. This means provider support may change, break, or be removed due to upstream updates outside of nitpik's control. If you are considering a commercial license, we recommend verifying that your provider and model work correctly using the free unlicensed version of nitpik before purchasing. No license key is needed — just install and test with your own API key.

Basic Setup

Set two environment variables:

export NITPIK_PROVIDER=anthropic
export ANTHROPIC_API_KEY=sk-ant-...

nitpik looks for the provider-specific key first (e.g. ANTHROPIC_API_KEY), then falls back to NITPIK_API_KEY as a universal alternative:

export NITPIK_PROVIDER=anthropic
export NITPIK_API_KEY=sk-ant-...       # works for any provider

Choosing a Model

By default, nitpik picks a sensible model for each provider — you only need to set NITPIK_PROVIDER and an API key to get started. Override the default with NITPIK_MODEL:

export NITPIK_MODEL=claude-sonnet-4-5-20250929

Or in your .nitpik.toml:

[provider]
name = "anthropic"
model = "claude-sonnet-4-5-20250929"

Default Models

Provider	Default model
Anthropic	claude-sonnet-4-5-20250929
Azure OpenAI	gpt-4o
Cohere	command-r-plus
DeepSeek	deepseek-chat
Galadriel	llama3.1-70b
GitHub Models	openai/gpt-4.1
Google Gemini	gemini-2.5-flash
Groq	llama-3.3-70b-versatile
HuggingFace	meta-llama/Llama-3.1-70B-Instruct
Hyperbolic	meta-llama/Llama-3.1-70B-Instruct
Mira	llama3.1-70b
Mistral	mistral-large-latest
Moonshot	moonshot-v1-32k
Ollama	llama3
OpenAI	gpt-4o
OpenRouter	anthropic/claude-sonnet-4.5
Perplexity	sonar-pro
Together	meta-llama/Llama-3.3-70B-Instruct-Turbo
xAI (Grok)	grok-3
OpenAI-compatible	gpt-4o

Tip: Code review benefits from strong reasoning capabilities. Models like Claude Sonnet, GPT-4o, and Gemini 2.5 Flash tend to produce the most precise findings. Smaller or faster models work fine for quick feedback but may miss subtle issues.

Ollama (Local Models)

Ollama runs locally and does not require an API key:

export NITPIK_PROVIDER=ollama
export NITPIK_MODEL=llama3

By default nitpik connects to http://localhost:11434. To use a different host, set NITPIK_BASE_URL:

export NITPIK_BASE_URL=http://192.168.1.100:11434

Azure OpenAI

Azure requires your deployment endpoint as NITPIK_BASE_URL and the model is your deployment name:

export NITPIK_PROVIDER=azure
export NITPIK_BASE_URL=https://your-resource.openai.azure.com
export AZURE_OPENAI_API_KEY=your-key
export NITPIK_MODEL=your-deployment-name

OpenAI-Compatible Endpoints

Use any API that speaks the OpenAI chat completions protocol — self-hosted models, corporate proxies, or alternative providers:

export NITPIK_PROVIDER=openai-compatible
export NITPIK_BASE_URL=https://your-endpoint.example.com/v1
export OPENAI_API_KEY=your-key
export NITPIK_MODEL=your-model-name

This works with LM Studio, vLLM, and similar services.

Per-Profile Model Overrides

Individual reviewer profiles can specify their own model, overriding the global setting. This lets you use a cheaper model for simple checks and a more capable one for security analysis:

---
name: security
description: Deep security analysis
model: claude-sonnet-4-5-20250929
---

See Custom Profiles for the full profile format.

Per-Task Model Overrides

A review run makes several kinds of LLM call, and the cheaper, non-review ones don't need a top-tier model. You can point them at a smaller/faster model — on the same provider and API key — while the per-file review keeps the strong one:

[provider]
name = "anthropic"
model = "claude-opus-4-8"               # per-file review

[provider.models]
triage  = "claude-haiku-4-5-20251001"   # auto profile selection + threat triage
summary = "claude-haiku-4-5-20251001"   # rolling PR summary (--pr-summary)

Or via environment variables:

export NITPIK_TRIAGE_MODEL=claude-haiku-4-5-20251001
export NITPIK_SUMMARY_MODEL=claude-haiku-4-5-20251001

Each override falls back to [provider] model (or NITPIK_MODEL) when unset, so omitting them keeps the previous single-model behavior. Only triage and summary are configurable: the per-file review and the critic/verify pass always use the primary model — the critic is judgment-heavy and intentionally stays on it. Token usage is attributed per model, so the run summary shows what each model cost.

See Configuration for the full reference.

Config File Setup

Instead of environment variables, configure the provider in .nitpik.toml:

[provider]
name = "anthropic"
model = "claude-sonnet-4-5-20250929"
# api_key = "..."   # possible but not recommended — use env vars for secrets

Or in your global config at ~/.config/nitpik/config.toml to set a default for all repositories.

See Configuration for the full layering order.

Prompt Caching

For multi-file reviews, nitpik structures the prompt so that the system prefix (agent system prompt + project documentation + commit history) is byte-identical across every file in the run. Providers with prompt caching can reuse that prefix on the second task and beyond, cutting the input-token bill substantially on big diffs.

Provider	Caching	How it activates
Anthropic	yes (ephemeral, 5 min)	rig-core inserts cache_control on large system blocks
OpenAI	yes (automatic)	provider caches prompts ≥ 1024 input tokens
Gemini	yes (implicit)	context cache kicks in above provider thresholds
Azure OpenAI	yes (automatic)	inherits OpenAI's behavior
GitHub Models	unknown	upstream caching behavior not documented
Cohere, DeepSeek, Groq, Mistral, others	none / provider-controlled	no client opt-in

When cache hits occur, the run summary surfaces them:

  ▸ Tokens: 42.1K↑ in, 1.2K↓ out (28.0K cached, 67% hit)

The hit ratio (cached_input / input) tells you what fraction of input tokens were served from cache. Providers without caching simply omit the cache section from the summary.

Related Pages

• GitHub Models — free GitHub-hosted inference with GITHUB_TOKEN, ideal for CI
• Quick Start — run your first review
• Configuration — all config options
• Custom Profiles — per-profile model overrides