Documentation / LLM setup

LLM & API key setup

Use this after your server is running and you can SSH in. Keys stay on your VPS—not in the AgenticHosting dashboard.

← Documentation overview

After your VPS is running, connect cloud LLM providers from the server over SSH. AgenticHosting does not collect API keys in the dashboard—configure keys only on your VM.

OpenClaw and NemoClaw differ in how you reach the Control UI (NemoClaw needs SSH port forwarding—see NemoClaw first login). For which LLM providers you can use, they align: the NemoClaw installer usually asks for NVIDIA first; if you skip that, you can configure the same kinds of cloud providers as OpenClaw (OpenAI, Anthropic, Google, and others the installer lists). Jump to your plan below. For OpenClaw SSH and the browser UI, start with OpenClaw first login. For NemoClaw tunnels and the NVIDIA-first / skip flow, follow NemoClaw first login.

OpenClaw

Replace YOUR_SERVER_IP with your public IPv4. SSH user is typically agentic (same as in your “server ready” email).

ssh agentic@YOUR_SERVER_IP

Provider API keys

Create keys in the provider console (you will paste or enter them during OpenClaw onboarding on the server—not here):

Onboard OpenClaw on the server

Run this in your SSH session before you use the SSH tunnel / Control UI in the browser (see the first-login guide). From /opt/openclaw:

cd /opt/openclaw
sudo docker compose run --rm --no-deps --entrypoint node openclaw-gateway \
  dist/index.js onboard --mode local --no-install-daemon

Follow the prompts to add providers and complete setup. For Docker details, see the OpenClaw Docker install docs.

How you know it worked

Onboard finishes without errors in the terminal.

Gateway health returns 200:

curl -sS -o /dev/null -w "%{http_code}\n" http://127.0.0.1:18789/healthz

Use an SSH tunnel and open the Control UI as described in the first-login guide.

Local models (Ollama)

Warning: Default plans use CPU-only instances (see product docs) with limited RAM and no GPU. Running local models is fine for experimentation only—expect slow inference and memory pressure.

For production workloads, prefer hosted APIs (OpenAI, Anthropic, Google) unless you move to a GPU tier in the future. See Ollama for installation and usage on Linux.

NemoClaw

NemoClaw is installed under /opt/nemoclaw. After SSH (ssh agentic@YOUR_SERVER_IP), check install progress if needed:

/opt/nemoclaw/health-check.sh — quick status
/var/log/nemoclaw-install.log — install log
/opt/nemoclaw/nemoclaw-install-monitor.sh — tail the log

When the stack is installed (/opt/nemoclaw/.installed on success), run interactive setup:

/opt/nemoclaw/complete-setup.sh

That script continues NemoClaw's installer. It typically offers an NVIDIA API key step first; you can skip and continue with other providers the same way you would on OpenClaw. Create keys in the Provider API keys consoles listed in the OpenClaw section above (OpenAI, Anthropic, Google, …). Full flow and tunnels: NemoClaw first login.

Keep keys safe

Never commit API keys to Git or paste them into public chats.
Prefer configuring keys only on the VPS over SSH or via the app's own secure prompts.
Rotate keys if they leak; monitor usage in each provider's billing dashboard.
Do not send secret keys to AgenticHosting support—use support for access and billing, not for sharing credentials.

Cost management

Set spending limits and alerts in OpenAI, Anthropic, and Google Cloud / AI Studio.
Start with smaller, cheaper models for tests; scale up when you need quality.
Review provider pricing pages regularly—token prices change.

Model comparison (rough guide)

Indicative only—check each provider for current pricing and context limits. Update this table in code when marketing needs fresher numbers.

For a fuller comparison, use-case matrix, provider links, and an interactive Claude configuration helper, see LLM provider reference.

Provider	Example model	Typical use
OpenAI	gpt-4o-mini	Low cost, good for drafts and tools
OpenAI	gpt-4o	Higher quality, higher cost
Anthropic	Claude 3.5 Sonnet	Strong reasoning and long context
Google	Gemini 1.5 Flash	Fast, cost-conscious
Local (Ollama)	Varies by hardware	CPU tier: testing only; GPU tier for heavier local use

Troubleshooting

Invalid API key: copy again from the provider; keys are often trimmed wrong or revoked.
Rate limited: wait and retry; check quotas in the provider dashboard.
Model not found:confirm the model name matches the provider's current catalog.
Connection timeout / firewall: for OpenClaw UI, confirm SSH tunnel and your cloud firewall rules (see first-login guide).