Agent builder

Overview

• Required

  Name + personality + model

  A name, personality, and model are required. Tools are optional — attach them later from the edit page.

• Optional

  Tools · Memory · Knowledge · Schedule · Tokens

  Compose the rest from the editor's tabs. Each attachment is independent — turn one off without touching the others.

Create an agent

1. 1

   Open Agents → New

   From the agent library, click New agent. The wizard opens on a single-screen form so every required field is visible at once.

2. 2

   Fill the agent details

   The wizard header reads

   Agent details

   — set the agent's name, personality, and the local model that will answer its chats.

3. 3

   Provision a model (if needed)

   If the model dropdown is empty, installation has not run yet. Run ollama pull llama3.2 on your host, then refresh.

4. 4

   Attach tools (or skip)

   The Tools section lists every enabled plugin from the marketplace — toggle the ones you want the agent to call. If nothing is enabled yet, No plugins enabled. Install plugins from the marketplace first.

5. 5

   Save

   The agent lands in the library. Click it to open the editor and add memory, knowledge, tasks, or tokens.

Personality & instructions

The Personality tab on the agent editor is where you shape the agent's voice, the constraints it operates under, and how it cites its tools. Keep the prompt tight — the model reads it before every turn, so every line of cost is paid every reply.

• Set the model at the bottom of the Personality tab — the dropdown reads from your locally installed model list.
• The prompt supports up to 8,000 characters; trim before enlarging.
• Use the Save button to apply changes; the chat thread will pick them up on the next turn.
• Each attached plugin surfaces in chat as a tool the model can call.

Tools (plugins + slash commands)

On the Tools tab of the agent editor, toggle each enabled plugin. The agent can then invoke the plugin's actions at chat time, or — when the plugin declares a slash command — intercept prompts like /weather before they reach the model, routing to /api/plugins/[id]/slash.

• Plugins → tool calls

  Model decides when to call. Useful for tools that need a structured request — fetch a calendar event, summarise a Gmail thread, post to Discord.

• Plugins → slash commands

  User types /weather (or any other /foo the plugin declares) and the plugin answer is rendered before the model is asked. Perfect for deterministic one-shot commands.

Disabled plugins are filtered out of the slash registry — a /weather reply never fires if the Weather plugin is off. If no plugins are enabled, the editor reads No plugins enabled. Install plugins from the marketplace first.

Memory

Use the Memory page to pin a fact, preference, or summary. Atlas retrieves the top-matching rows for ordinary and agent chat, bounds their prompt size, and treats their text as untrusted reference material rather than instructions.

• Label
  Short, scannable — e.g. User timezone.
• Content
  The fact itself — e.g. Eastern Time (UTC-5/UTC-4) — confirm DST shift when scheduling.
• Up to 16,000 characters per entry; trim long-form notes before saving.
• Automatic memory synthesis is off by default. If you opt in, Atlas selectively combines and revises durable details instead of saving every response. Automatic memories count toward the same quota and remain editable and deletable.

Free plan caps memory at 50 entries; Pro lifts the cap to 1,000 entries and expands the document upload library. Pro also unlocks larger knowledge-base caps — see the next section.

Knowledge base

On the Memory page, the Knowledge bases tab accepts PDF, DOCX, .txt, .md, and .csv files. Atlas chunks and embeds them locally with Ollama and surfaces the most relevant passages in chat replies — your citations carry the filename and chunk index back into the conversation, so the agent can quote rather than guess.

• Accepted

  PDF, DOCX, .txt, .md, .csv.

• Free plan

  5 docs · 25 MiB total

• Pro plan

  100 docs · 1 GiB total

Scheduled tasks

From the Schedule tab, click New task to open the Schedule a recurring prompt dialog. Each task fires its prompt at the cron schedule. Auto-derived memories are pinned back to the agent's memory tab after each run.

0 9 * * *    # every day at 09:00 (local)
*/15 * * * *  # every 15 minutes
0 9 * * 1    # Mondays at 09:00

API tokens

The Tokens tab on the agent editor issues bearer tokens to call POST /api/v1/agents/{id}/chat. Provide a label so the token is easy to identify later (Home Assistant, GitHub CI, custom API caller).