Quickstart

Get NimbleBrain running locally with Docker Compose. You need Docker and an Anthropic API key.

Create a config file

Create nimblebrain.json in a new directory:
nimblebrain.json
```
{
  "$schema": "https://schemas.nimblebrain.ai/v1/nimblebrain-config.schema.json",
  "version": "1"
}
```
Both $schema and version are part of the config. This minimal file starts NimbleBrain with no MCP apps installed. The agent still has built-in platform capabilities (home, conversations, files, automations, skills, usage). You install apps from the web UI in step 5 to give it domain-specific tools.

Apps are not listed in nimblebrain.json. Per-workspace app installs live in workspace.json, which the runtime manages for you. The config file is for runtime settings (model, limits, features), not app inventory.

Create a Docker Compose file

services:
  platform:
    image: ghcr.io/nimblebraininc/nimblebrain:latest
    restart: unless-stopped
    volumes:
      - workspace:/data
      - ./nimblebrain.json:/data/nimblebrain.json:ro
    environment:
      - ANTHROPIC_API_KEY=${ANTHROPIC_API_KEY}
    networks:
      - internal

  web:
    image: ghcr.io/nimblebraininc/nimblebrain-web:latest
    restart: unless-stopped
    ports:
      - "27246:8080"
    environment:
      - PLATFORM_URL=http://platform:27247
    depends_on:
      platform:
        condition: service_healthy
    networks:
      - internal

volumes:
  workspace:

networks:
  internal:

The platform image runs the agent API on port 27247 (internal). The web image serves the UI and is published on port 27246.

Set your API key and start
Terminal window
```
export ANTHROPIC_API_KEY=sk-ant-api03-...
docker compose up
```
You should see output like:
```
✔ Container quickstart-platform-1  Healthy
✔ Container quickstart-web-1       Started
```
The platform container reports healthy once http://localhost:27247/v1/health responds (checked internally). The web container then starts.
Open the web UI

Go to http://localhost:27246.

There is no login screen by default. With no instance.json present, NimbleBrain runs in dev mode — requests are unauthenticated and you act as a single built-in user. This is the right setup for trying things out locally.

Dev mode has no authentication. Before exposing NimbleBrain to a network, configure a real auth adapter (oidc or workos) in instance.json. See Installation and the configuration reference.

Try sending a message:
```
What can you help me with?
```
The agent responds using its built-in platform capabilities. To give it domain-specific tools, install an app in the next step.
Install an app

Apps come from two places: the in-app connectors catalog (remote MCP connectors with point-and-click OAuth) and the mpak registry (MCP server bundles).

Open the connectors catalog from workspace settings to browse and install connectors directly in the UI. You can also drive installs by chatting with the agent — it uses the manage_connectors tool to list the catalog and install a connector, and manage_apps to manage registry apps across the org:
```
List the available connectors and install one for me
```
Once an app is installed and running, its tools join the agent’s namespace. Ask the agent to use them in plain language.

What’s next

Installation All install methods: Docker Compose or from source, plus auth setup

Core Concepts Understand the agent loop, bundles, skills, and tools

Installing Apps Browse and install MCP bundles from the mpak registry