Your First Conversation

This guide walks you through your first interaction with Talome — from creating your account to having a fully configured app running on your server. By the end, you will understand how the AI works, what it can do, and how to get the most out of it.

Step 1: Create Your Admin Account

Open http://localhost:3000 in your browser. On first visit, Talome presents a simple form to set your admin password.

Enter a strong password and click Create Account. This is the only user account — Talome is designed for single-admin use. You will be logged in immediately.

Step 2: Configure Your AI Provider

Before the assistant can do anything, it needs an AI provider. Navigate to Settings from the sidebar, then open the AI Provider section.

Anthropic (Recommended)

Anthropic's Claude is Talome's default and best-tested provider. The assistant uses Claude Haiku for fast, routine operations and Claude Sonnet for complex reasoning.

1. Go to console.anthropic.com and create an API key
2. Paste it into the Anthropic API Key field
3. Click Save

Typical cost: $0.01-0.05 per conversation. A full media stack setup with five app installs, wiring, and verification costs roughly $0.03.

OpenAI (Alternative)

If you prefer OpenAI, set your OPENAI_API_KEY and select OpenAI as the default provider. GPT-5.3 is the recommended model. All tools work identically — the AI layer is provider-agnostic.

Ollama (Local, Free)

For fully local AI with no cloud dependency:

1. Install Ollama on your server
2. Pull a model: ollama pull llama3.1
3. In Talome Settings, set the Ollama URL (typically http://localhost:11434)
4. Select Ollama as the default provider

Step 3: Open the Assistant

The AI assistant is always available. You have two ways to open it:

• Click Assistant in the sidebar navigation
• Press / on your keyboard from any page to open the command palette in chat mode

The assistant is a full page in the dashboard with conversation history, suggestion cards, and a chat input bar.

Step 4: Ask What It Can Do

Type your first message:

What can you do?

The AI responds with a structured overview of its capabilities, organized by category — Docker management, app store, media, networking, automations, self-improvement, and more. This is not a static help page. The response is generated live, reflecting which integrations you have configured.

If you have not configured Sonarr yet, the AI will not mention arr tools. Once you configure an integration, those capabilities appear automatically.

Step 5: Install Your First App

Now for the real demonstration. Type:

Install Jellyfin

Here is what happens, step by step:

Notice what happened:

1. search_apps — The AI searched across all configured stores to find the Jellyfin app, identified the correct store ID and app ID
2. install_app — It pulled the Docker image, created the container with correct ports, volumes, environment variables, and a health check
3. check_service_health — It verified the container started correctly and the HTTP endpoint responds

Three tool calls, executed automatically. You typed five words.

Step 6: Go Deeper

The assistant remembers your conversation context. Try these follow-up messages:

Check your server

Show me all running containers

The AI calls list_containers and returns a formatted table with container names, status, CPU/memory usage, and ports.

Add a memory

Remember that my media is stored at /mnt/nas/media

The AI stores this as a fact memory. Next time you install a media app, it will automatically use this path for volume mounts without asking.

Diagnose something

Why is Jellyfin using so much memory?

The AI calls get_container_stats to read the actual resource usage, then inspect_container to check the configuration. It gives you specifics — exact memory consumption, limits (or lack thereof), and offers to set resource limits if needed.

Install a full stack

Set up Sonarr, Radarr, Prowlarr, and qBittorrent and wire them all together

This is where Talome shines. The AI chains multiple tools: five install_app calls, then wire_apps to connect download clients, then arr_sync_indexers_from_prowlarr to push indexers — all in a single turn.

What Happens Behind the Scenes

Understanding the mechanics helps you get the most out of Talome. Here is what happens every time you send a message:

1. System prompt construction

The backend builds a context-aware system prompt that includes:

• Talome's core instructions (how to use tools, response patterns, safety rules)
• Your current page context (if you are on the Media page, media tools are prioritized)
• A custom system prompt, if you have set one in Settings

2. Memory injection

The top 10 most relevant memories are fetched from SQLite and injected into the prompt. Memories are ranked by a composite score:

• Recency — more recent memories rank higher
• Access frequency — memories the AI has recalled before rank higher
• Confidence — memories with high confidence (set by the AI based on how explicit you were) rank higher

This means the AI always has your preferences and facts in context without you repeating them.

3. Dynamic tool loading

Not all 230+ tools are sent to the AI on every message. Talome uses per-message tool routing:

• Core tools (Docker, apps, system, filesystem, memories) are always available
• Domain-specific tools (arr, Jellyfin, qBittorrent, etc.) load only when that app is configured in Settings
• Within active domains, keyword matching in your message further narrows the tool set — a message about "torrents" loads qBittorrent tools but not Home Assistant tools

This keeps the tool count low for each call, which improves AI accuracy and reduces latency.

4. Streaming execution

The AI's response streams back token by token. When it decides to call a tool, the tool executes server-side against your real Docker socket, database, or app API. The result flows back into the AI's context, and it continues generating its response. You see this as tool execution cards appearing inline in the chat.

5. Security gating

Every tool call passes through the security gateway before execution:

• Permissive mode — all tools execute freely
• Cautious mode (default) — read and modify tools execute normally; destructive tools (uninstall, delete volume, prune) require the AI to explicitly confirm
• Locked mode — only read-tier tools work; all modifications are blocked

Each tool is classified into a tier (read / modify / destructive), and the gateway enforces the active security mode.

Things to Try Next

System management

• "What's using the most disk space?"
• "Show me system health"
• "Clean up unused Docker images"
• "Set up nightly backups for Jellyfin"

App store

• "Search for a self-hosted note-taking app"
• "Install Pi-hole for DNS ad blocking"
• "What apps do I have installed?"
• "Update all my apps"

Media (after installing Sonarr/Radarr)

• "Search for The Bear and add it to Sonarr"
• "What's downloading right now?"
• "Show me what's coming out this week"
• "Why is Radarr not finding releases for Dune?"

Networking

• "Set up local DNS so I can access apps at *.talome.local"
• "Add a reverse proxy route for Jellyfin at media.example.com"
• "Set up Tailscale for remote access"

Home automation (after installing Home Assistant)

• "Show me all Home Assistant entities"
• "Turn off the living room lights"
• "Create an automation to turn on the porch light at sunset"

Custom apps

• "Create a bookmark manager with tags, search, and a dark UI"
• "Build me a simple dashboard that shows my Sonarr and Radarr calendars"

Self-improvement

• "Read the source code for the container list page"
• "The dashboard loads slowly — can you profile it?"
• "Add a button to the container card that opens the logs"

Next Steps