AI Providers

Sentri supports five built-in providers plus OpenAI-compatible slots for any vendor that speaks the OpenAI Chat Completions wire format (DeepSeek, Groq, Mistral, xAI, Azure OpenAI, Together, Fireworks, vLLM, LM Studio, LiteLLM, …). Switch at any time from the Settings page — no restart needed.

Provider Comparison

Provider	Model	Cost	Best For
Anthropic Claude	claude-sonnet-4-20250514	Pay-as-you-go ($5 min)	Best quality output
OpenAI	gpt-4o-mini	Pay-as-you-go	Fast, high-volume crawls
Google Gemini	gemini-2.5-flash	Free tier (20 req/day)	Testing / evaluation
OpenRouter	openrouter/auto (configurable)	Pay-as-you-go	One key, 200+ models (Claude, GPT, Llama, Mixtral, …)
OpenAI-compatible (compat:<id>)	Any (you set baseUrl + model)	Vendor-dependent	DeepSeek, Groq, Mistral, xAI, vLLM, LiteLLM, …
Ollama	mistral:7b (configurable)	Free, local	Privacy, no API key needed

Auto-detection order: Anthropic → OpenAI → Google → OpenRouter → any configured compat:<id> slot → Ollama

Configuration

Via Settings Page (Runtime)

Navigate to Settings in the app. Enter your API key for any cloud provider, or click "Activate Ollama" for local inference.

Via Environment Variables (Persistent)

Add to backend/.env:

# Cloud providers (pick one or more)
ANTHROPIC_API_KEY=sk-ant-api03-...
OPENAI_API_KEY=sk-proj-...
GOOGLE_API_KEY=AIza...
OPENROUTER_API_KEY=sk-or-v1-...
OPENROUTER_MODEL=anthropic/claude-3.5-sonnet   # any OpenRouter slug

# Local / Ollama (no key needed)
AI_PROVIDER=local
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=mistral:7b

Using OpenRouter (One Key, Many Models)

OpenRouter is a unified gateway to 200+ models behind a single OpenAI-compatible API. Useful when you want to A/B-test models or get fallback across vendors without juggling keys.

1. Get a key from openrouter.ai/keys
2. In Sentri Settings, paste the key into the OpenRouter card — or set OPENROUTER_API_KEY in .env
3. (Optional) Pick a specific model via OPENROUTER_MODEL:
   • openrouter/auto (default — OpenRouter picks)
   • anthropic/claude-3.5-sonnet
   • openai/gpt-4o-mini
   • meta-llama/llama-3.1-70b-instruct
   • …see openrouter.ai/models

Using OpenAI-Compatible Providers (DeepSeek, Groq, Mistral, vLLM, …)

Many vendors expose /v1/chat/completions with the OpenAI request/response schema. Sentri lets you register any of them as a compat slot keyed compat:<id> — no code change, no SDK install. Each slot stores its own baseUrl, model, apiKey, and displayName, gets its own circuit breaker, and participates in the FEA-003 fallback chain.

Add a slot via the Settings page

Navigate to Settings → OpenAI-compatible providers and fill in:

Field	Example	Notes
Slot id	deepseek	Lowercase, [a-z0-9_-]+. Becomes compat:deepseek internally.
Display name	DeepSeek	Shown in the provider badge / dropdown. Optional.
Base URL	https://api.deepseek.com/v1	Must be HTTPS public; SSRF-validated server-side.
Model	deepseek-chat	Vendor-specific model id.
API key	sk-…	≥10 characters; stored AES-encrypted in the api_keys table.

Click Save compat provider. The slot is activated immediately and joins the fallback chain on rate limits / 5xx errors.

Common base URLs

Vendor	Base URL	Sample model
DeepSeek	https://api.deepseek.com/v1	deepseek-chat, deepseek-reasoner
Groq	https://api.groq.com/openai/v1	llama-3.3-70b-versatile
Mistral	https://api.mistral.ai/v1	mistral-large-latest
xAI Grok	https://api.x.ai/v1	grok-2-latest
Together AI	https://api.together.xyz/v1	meta-llama/Llama-3.3-70B-Instruct-Turbo
Fireworks	https://api.fireworks.ai/inference/v1	accounts/fireworks/models/llama-v3p3-70b-instruct
Azure OpenAI	https://<resource>.openai.azure.com/openai/deployments/<deployment>	Your deployment name
LiteLLM proxy	http://litellm.internal:4000/v1	Any model the proxy fronts
vLLM / LM Studio / LocalAI	http://<host>:<port>/v1	Whatever the server hosts

Add a slot via the API (Docker / K8s bootstrap)

Compat slots are not configurable through env vars (see the warning below). For headless deployments, bootstrap the slot once via the admin API after the server starts:

curl -X POST https://your-sentri/api/v1/settings \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "provider":    "compat:deepseek",
    "displayName": "DeepSeek",
    "baseUrl":     "https://api.deepseek.com/v1",
    "model":       "deepseek-chat",
    "apiKey":      "sk-..."
  }'

The slot is persisted in the encrypted api_keys table and survives restarts. To delete:

curl -X DELETE https://your-sentri/api/v1/settings/compat:deepseek \
  -H "Authorization: Bearer $ADMIN_TOKEN"

No env-var configuration

Compat slots are DB-only — there is intentionally no COMPAT_<ID>_API_KEY env equivalent. If you need pure env-driven config, use OpenRouter (OPENROUTER_API_KEY + OPENROUTER_MODEL=deepseek/deepseek-chat) — OpenRouter fronts most of the same vendors and is fully env-configured.

Self-hosted / on-prem endpoints (vLLM, LiteLLM, LocalAI)

The SSRF guard blocks loopback (127.0.0.1), RFC1918 (10.x, 192.168.x), link-local (169.254.x), and localhost base URLs at save time and re-validates on every call (DNS-rebinding mitigation). To run against an internal LLM server, set:

ALLOW_PRIVATE_URLS=true

in backend/.env. This bypass is scoped exclusively to compat-slot saves and the per-call guarded fetch — trigger callbacks, preview URLs (Playwright-navigated), and notification webhooks remain SSRF-protected. Do not enable in multi-tenant deployments where untrusted users can configure providers.

Performance tuning

The decrypted compat config is cached in-memory with a 60 s TTL to avoid hitting SQLite + AES decryption on every AI call. Override the TTL via:

COMPAT_CONFIG_CACHE_TTL_MS=30000   # default 60000

When REDIS_URL is set, cache invalidations are broadcast to other Sentri instances over the sentri:compat-config:invalidate channel so credential updates are visible cluster-wide within one round-trip.

Using Ollama (Free Local AI)

1. Install from ollama.com
2. Pull a model:

   ollama pull mistral:7b          # ~2 GB, good quality
   ollama pull qwen2.5-coder:7b  # great for code generation
   ollama pull mistral           # lighter alternative

3. Start the server: ollama serve
4. In Sentri Settings, select Ollama — or set AI_PROVIDER=local in .env

Recommended Models

For best results use a model with strong JSON output and code generation: mistral:7b, qwen2.5-coder:7b, or mistral. Small models (≤3B) may struggle to produce valid Playwright code.