Documentation

Running in 60 seconds. Calling it in 60 more.

AGNT is a local runtime for AI agents. Install it, and everything you can do in the app is an HTTP call on your own machine.

  • No account. Community Core runs locally, free for personal use, education, nonprofits and small businesses.
  • No cloud API to learn. Your code talks to your runtime at localhost:3333.
  • Any model. A cloud key, or Ollama and LM Studio fully offline.
Quickstart · headlessPrefer the desktop app? →
  1. Start the runtime

    docker run -d \
      --name agnt-full \
      -p 3333:3333 \
      -v agnt-data:/app/data \
      --restart unless-stopped \
      ghcr.io/agnt-gg/agnt:0.6.6
  2. Check it is up

    curl http://localhost:3333/api/agents/health

    Returns {"status":"OK"} — no token needed.

  3. Make your first authenticated call

    const res = await fetch('http://localhost:3333/api/agents/', {
      headers: { Authorization: 'Bearer ' + token },
    });
    const agents = await res.json();
Base URL localhost:3333/apiAuth JWT bearerThen connect a model in Settings → Providers
Getting startedInstall, connect a model, give an agent its first job. The ten-minute path.Open guide → API quickstartBase URL, auth in ten lines, first call, then every endpoint family.Open reference → Run it headlessOne image for servers and VPSes: volumes, compose and updating.Docker guide →
Reference

Authentication

Requests carry a JWT bearer token. Where it comes from depends on where your code runs. Many local endpoints also answer unauthenticated on localhost.

Your code runs inThe token comes from
A widgetThe agnt SDK handles it
The frontendThe signed-in session
An orchestrator toolprocess.env.AGNT_AUTH_TOKEN, injected for you
Authorization: Bearer <your-jwt-token>

Copy-paste patterns for all three contexts →

First calls

The base URL is http://localhost:3333/api/. Your code talks to your own runtime; there is no cloud API to learn.

bash · no token
curl http://localhost:3333/api/agents/health
javascript
const res = await fetch('http://localhost:3333/api/agents/', {
  headers: { Authorization: 'Bearer ' + token },
});
const agents = await res.json();

Errors, rate limits, pagination, SSE and WebSockets →

Connect a model

Settings → Providers. Mix freely: a local model for routine volume, a frontier model for the hard steps, and change any agent's provider later without rebuilding it.

RouteHowTrade-off
Cloud API keyOne key from OpenAI, Anthropic, Google Gemini, Grok, DeepSeek, Mistral and othersStrongest models; per-token cost; data leaves for the model call
Fully localPoint AGNT at Ollama (port 11434) or LM Studio (port 1234)No per-token cost and complete privacy; needs capable hardware
json · Ollama
{
  "provider": "ollama",
  "baseUrl": "http://localhost:11434",
  "defaultModel": "llama3.3:70b"
}

Models & Providers API →

Run with Docker

One image with the full feature set, including browser automation. Needs Docker 20.10+ or Docker Desktop, 2 GB of free RAM and 2 GB of free disk. Pin a version tag so an update is a decision, not a surprise.

bash
docker run -d \
  --name agnt-full \
  -p 3333:3333 \
  -v agnt-data:/app/data \
  -e NODE_ENV=production \
  -e BASE_URL=http://localhost:3333 \
  --restart unless-stopped \
  ghcr.io/agnt-gg/agnt:0.6.6

Compose, volumes, updating and troubleshooting →

All documentation

Generated from the docs pipeline, so it is always current with what is published.

Loading documentation…

Frequently asked questions

Is my data sent to AGNT servers when using Community Core?

No. Community Core runs 100% locally. Your prompts, code files, credentials, and receipts stay on your machine unless you explicitly connect a remote cloud provider.

Can I use AGNT without an internet connection?

Yes. When connected to local models via Ollama or LM Studio, AGNT functions completely offline for all filesystem, local shell, and code editing operations.

How do approval gates work for dangerous operations?

Any action that modifies system settings, executes arbitrary shell scripts, or sends external emails triggers a visual approval modal in AGNT before execution.

The AGNT briefing

Stay in the loop.

New agents, workflows and field notes. Sign up for AGNT updates.

By subscribing, you agree to receive AGNT updates. Unsubscribe any time. Privacy policy.