.env file. Setup never
prints secret values and supports interactive, scripted, and stdin-based flows.
--key for CI, --key-stdin to avoid shell history, and remove <name> --yes to revoke a local configuration entry. Common aliases include gh
for GitHub, hf for Hugging Face, gemini/gcp for Google, and ctx7 for
Context7.
The registry covers the following categories:
- Search, model providers, collaboration, and app services.
- Databases —
postgres,redis,mongodb. - Package registries — Docker Hub, npm, PyPI.
- Observability — SigNoz, Sentry, Datadog, New Relic.
- Infrastructure — Cloudflare, DigitalOcean, Hetzner.
- Hosting — Railway, Vercel, Netlify, Render, Fly.io.
- Email — Resend, SendGrid.
- Authentication — Clerk, Auth0.
- Analytics — PostHog, Amplitude.
- Feature flags — LaunchDarkly, Statsig.
LLM observability includes Langfuse.
AI development integrations also include LangSmith and Weights & Biases.
Remote compute integrations include Modal and E2B.
Local model runtimes include Ollama, vLLM, and LM Studio; use
--urlfor a custom local endpoint. Hosted model integrations include Groq and Replicate. Messaging providers include Telegram and WhatsApp.
set up Serper, show available integrations,
or check my Sentry integration.
Local and hosted models together
Usearka hybrid status to see reachable local runtimes (Ollama, vLLM, or
LM Studio) alongside configured hosted providers. arka hybrid run "..."
uses local-first routing by default and falls back to hosted models. Select
--policy hosted-first, local-only, or hosted-only; --policy parallel
runs one local and one hosted model and returns both answers for comparison.
Persist the default with arka hybrid config local-first (or another policy).
For privacy-sensitive work, force a hard local-only boundary:
local-only filters the candidate chain before inference, so hosted providers
are not used as fallbacks. If no reachable Ollama, vLLM, or LM Studio model is
configured, Arka returns an error instead of sending the prompt to the cloud.