HomeProduct › Self-Hosted
Deployment

Self-hosted AI automation, unmetered.

Run AGNT on the desktop, on a home server, or in a container on infrastructure you already pay for. There is no execution ceiling, no per-task counter, and no tier that unlocks the ability to run your own automation more often.

Download free Self-hosting docs
Three Ways to Run

Same runtime, different footprint

Workflows and agents are portable across all three. Build on the desktop, move to a server when the work needs to run around the clock.

01 / DESKTOP

Windows, macOS, Linux

The full application with the visual builder, agent workspace, and marketplace. The fastest way to start — install and run, no account required.

02 / DOCKER

Containerized

Run headless in a container on a NAS, a VPS, or your own cluster. Ideal for scheduled and webhook-triggered work that should not depend on a laptop being awake.

03 / MANAGED

AGNT-hosted cloud

If you would rather not operate it, paid tiers run the same runtime for you with encrypted vault sync and 24/7 execution.

# Pull and run headless — persist state to a local volume
docker run -d --name agnt \
  -p 3000:3000 \
  -v ./agnt-data:/app/data \
  --restart unless-stopped \
  agnt/agnt:latest

# With a local model — no data leaves the host
docker compose up -d   # agnt + ollama, see the docs for the compose file

Current image tags, compose files, and environment variables are documented in the setup guide.

The Economics

Why unmetered changes what you automate

Per-task pricing shapes behaviour

When each run costs money, you ration automation. The work that gets automated is the work expensive enough to justify the meter — which excludes most of the small, frequent, genuinely annoying tasks.

  • High-frequency polling becomes uneconomic
  • Retries and error handling cost extra
  • Testing a workflow consumes production budget
  • Success is penalised — more volume, higher bill

Owning the runtime removes the question

Local execution has no marginal platform cost, so the only question is whether the automation is useful — not whether it clears a per-run price.

  • Poll every minute if that is what the job needs
  • Retry aggressively; failures are free
  • Test as much as you like before shipping
  • Scale volume without scaling platform spend

What you do still pay for

Model inference. If you route to a hosted provider, that provider bills you directly for tokens — AGNT does not mark it up or sit in the middle. Route to a local model through Ollama or LM Studio and that cost goes to zero as well, at the price of running the hardware yourself.