Our GitHub account is temporarily unavailable: links to the code, releases and issues will not open.

Report a bug or talk to us

AI agents

Agents, memory and skills

Agents are the assistants you talk to in Chat and hand tasks to in Work. Each one has a name, a purpose, a connection and model, and a list of tools it may use; Memory keeps what they learn about you, and Skills give them reusable instructions.

Where it is

LLM › Agents. The second row has Agents (the page is titled Agent team), Memory and Skills.

The default team

OmniGet starts with three agents:

Agent Role What it does Tools
Omni Coordinator The assistant inside OmniGet. Answers briefly, in your language, and uses a tool only when it is clearly needed. Reads files freely; asks before editing, writing, running commands or delegating.
Builder Worker The coding agent. Reads the code before changing it and makes the smallest edit that solves the task. Same as Omni.
Scout Worker The reading and research agent. Finds files and facts, says where each came from, never edits. Read-only tools.

All three pick their model by route: a local Ollama model first, then Anthropic, then OpenAI, using whichever connection you have. Add one in LLM › Configure › Connections (Connections, models and limits).

Roles: the coordinator divides and follows the work, workers carry out assigned tasks, and advisors help evaluate decisions. You can also write a Custom role.

Create an agent

  1. Open LLM › Agents and click Create agent.
  2. Under Connection, pick one in Choose a connection. If none is ready, click Connect an account or model first.
  3. Give it a Name and say What is it for?.
  4. Tick What it may do (see the table).
  5. Optionally pick a Skill (optional).
  6. Click Create agent.
Option What it does Default
Personal memory The agent remembers things about you. On
Web search The agent can search the web. Off
Reading companion Helps with what you read; also turns on Web search. Off
Project code Access to code and the shell in a project. Off
Delegate to other bots Can hand work to other agents. Off
Memory (under Advanced details) Remember on its own, Ask before saving or Only recall, never save. Remember on its own

An agent gets no code or shell access unless you tick it.

Full editor

Configure model, fallback and permissions manually (or Advanced settings (model, tools, budget) on an existing agent) opens the full editor in three steps:

  1. Name and purpose: Name, Role and What should this agent do? (its instructions).
  2. Connection and model: Fixed uses one model; Route tries a list of candidates in order (Add candidate).
  3. Review before saving, with Tools and limits:
    • Tools: each tool is auto (runs without asking), ask (waits for you) or deny. MCP tools are set in Permissions (MCP servers you use).
    • Skills: which installed skills the agent can open.
    • Budget: USD per day, Tokens per turn and Tool calls per turn.

Delete removes the agent and its settings. Saved conversations stay readable, but you cannot continue them without an agent.

Team templates

Team templates adds a preset group of agents. Preview team shows what it adds; Apply never replaces an agent you already have with the same identifier.

Memory

Memory shows what your agents remember about you, where each item came from, and how to fix or forget it. Memory is stored on your computer.

There are three scopes:

  • Your profile: shared by your personal agents in direct conversations. Rooms only see it if you share it.
  • Private notes of an agent: only that agent reads them.
  • Room notes: shared by the members of that room only.

To add something, click Add, type What should be remembered? and choose the Kind: Stated by you, Only for now or Working notes. On each item:

  • Edit (the old version stays in History).
  • Why is this here? shows the source.
  • Forget removes it and all its versions.
  • For guesses, Yes, that’s right or Not right.

The panel about what memory learns has Learn from conversations on its own, Keep a note about each person and thing, Turn on search by meaning, Search memory and Teach from a text. You can also Export readable (Markdown), Export for import (JSON) and Import an export….

Skills

Skills follow the SKILL.md format: a folder with instructions the agent opens only when it needs them.

  1. Open LLM › Agents › Skills and click Install a skill.
  2. Choose From a folder (the folder is copied, not linked), From a zip or From a git repository.
  3. After the install, click Assign to an agent.

The Explore tab has Install from GitHub (owner/repo, owner/repo/path or owner/repo@skill) and a small showcase catalog. Import from a pack lets you pick skills, agents, rules or commands item by item from a folder.

Common problems

An agent does not answer

It has no working connection. Open the agent’s editor and check Connection and model, or add a connection in LLM › Configure › Connections.

Applying a team template fails

In this release, Duo and Research squad show a save error instead of adding agents, and Solo adds nothing when Omni is already on the team. Create the agents by hand with Create agent.