Skip to main content
Last Updated: February 13, 2026 OpenClaw is a local-first AI agent framework. It can chat, run tools, automate tasks, and connect to channels such as Telegram, Discord, and other messaging platforms. This page explains:
  • what OpenClaw is,
  • why model costs can grow quickly,
  • why SaladCloud is a practical way to host models for OpenClaw.

​
What OpenClaw is

OpenClaw is the orchestration layer around your model, tools, memory, and channels. The model is the “brain,” while OpenClaw handles:
  • conversation and session state,
  • tool execution and workflow logic,
  • channel integrations,
  • scheduling and background activity.
Because OpenClaw can run frequent background calls and tool-heavy workflows, your model choice and hosting strategy have a direct impact on cost and performance.

​
Why costs increase with API-hosted models

OpenClaw itself is free, but model inference is not. With traditional API providers, billing is typically token-based. Costs tend to grow because of:
  • accumulated conversation and tool context,
  • repeated background/heartbeat interactions,
  • sub-agent and multi-step task execution,
  • always-on usage patterns.
In practice, this can make costs hard to predict for active agent workflows.

​
Why host OpenClaw models on SaladCloud

SaladCloud changes the cost model from token billing to compute-time billing for your deployment. For many OpenClaw workloads, this has three main advantages:
  • Predictability: pay for runtime, not each generated token.
  • Control: scale replicas up/down, including scheduled scale-to-zero windows.
  • Privacy posture: keep more prompt/context flow within infrastructure you control.
For users running long sessions, heavy tool usage, or frequent background checks, this can materially reduce total cost.

​
Inference options on SaladCloud

You can run OpenClaw-compatible model endpoints with:
  • Ollama: easiest path for most users.
  • vLLM: higher throughput and advanced serving optimizations.
  • TGI: strong text-generation serving option for many Hugging Face models.
All can be used with OpenClaw when configured with compatible endpoints and model metadata. A common production pattern is:
  1. run OpenClaw locally (or in your controlled environment),
  2. host one or more model endpoints on SaladCloud,
  3. set model fallbacks in OpenClaw,
  4. apply autoscaling/scheduling to align runtime with active hours.
This balances capability, reliability, and cost.

​
Next step

Follow the full setup guides to deploy OpenClaw with SaladCloud-hosted models:

​
References