selfhost
Other platforms
Same pattern everywhere — credentials → machine → vllm serve → /connect —
these just have console-driven flows we won't reproduce click-by-click:
- Replicate (API tokens,
REPLICATE_API_TOKEN) — strongest for their hosted model catalog; custom serving means packaging with cog. - Baseten (API keys,
BASETEN_API_KEY) — deploy via their model library or Truss; dedicated deployments expose OpenAI-compatible endpoints. - Beam (dashboard token) — Modal-style Python function deploys; adapt the Modal recipe.
- Paperspace (console API keys) — rent a machine, follow the Lambda recipe verbatim.
- AWS / GCP / Azure — a GPU VM + vLLM behind your VPN; the engineering is identical to the Lambda recipe, the IAM is yours.
- Agent sandboxes — Daytona
(
DAYTONA_API_KEY), E2B (E2B_API_KEY): for running agent code safely, not serving weights; pair with any endpoint above.
Local runtimes (Ollama, llama.cpp, LM Studio) intentionally have no pages
here — mantis setup automates them end-to-end.