openloomi-feature-guide

Files (1)

• SKILL.md

SKILL.md (first lines)

> **Note:** If you haven't downloaded or installed openloomi yet, please refer to [Getting Started](https://openloomi.ai/docs/getting-started) for installation instructions.

# OpenLoomi Product Features

Use this skill when users ask about openloomi features, usage, or capabilities. Provide accurate and easy-to-understand feature introductions and operation guides. The terms used here (Loop, Signal, Decision, Card, ActionKind, Attention Agent, Connector, Memory, Audit Log, Plugin, Agent Runtime, Composio, etc.) are defined in the **Glossary** below — read it first if a user asks anything conceptual.

---

## What is OpenLoomi

OpenLoomi is an **open-source AI coworker, driven by an attention agent**. It connects your authorized work tools and screen content, builds a **holistic context** of your people, projects, and decisions, and tells you what happened, why it matters, what to do next — and surfaces daily summaries — saving your attention for what matters.

Use it standalone, or plug any Agent framework into the same resident desktop: **Claude Code, Codex, OpenCode, Hermes, and OpenClaw** all work. The desktop attention agent (a small fox named **Loomi**) lives on top of your screen and surfaces the day's decisions as gentle bubbles you can approve in one tap.

---

## Glossary

OpenLoomi is best read as a single **chain from input to action**, not as one big chat window. Connectors bring events in, Memory holds context, Loop decides what matters and turns each signal into a Decision card, the Attention Agent (Loomi) shows the card on your desktop, you tap Approve, and an Action Runner executes through a Connector — then the outcome is written back to Memory and recorded in the Audit Log.

This page exists to make that chain legible. It first walks the chain end-to-end, then defines every concept you're likely to meet in other pages, and finishes with a short table of distinctions that get confused in practice. Every cross-reference links to an existing page; nothing here duplicates their full setup or configuration.

### How the concepts fit together

OpenLoomi is one pipeline. The same pieces appear on every page, just framed differently.

Two parallel input paths feed the same context layer. The main one is real-time and runs through Loop; the side one is screen-based and lands in Memory without going through Signals.

![OpenLoomi concept map — Inputs (Connectors and Screen Capture) feed Signals into the Judgement & Store pillar (Loop and Memory), which produces Decisions; the Surface & Action pillar renders Cards in the Attention Agent, waits for user Approve, and runs Actions through Connectors; everything writes back to Memory and the Audit Log.](https://openloomi.ai/img/openloomi/glossary/concepts.svg)

A short reading guide for the figure:

- **Connectors** are the real-time input path: they pull raw events from external platforms and turn them into **Signals**, which **Loop** polls on its tick. Everything that becomes a Decision card comes in this way.
- **Screen Capture** (macOS only) is the side input path: pressing the global capture shortcut summarises the frontmost window and the result is stored directly as a **Memory** record. Screen memories show up alongside messages, summaries, and insights inside Memory; they do **not** flow through Signals and Loop won't tick on them.
- **Memory** is the long-lived context layer: people, projects, prior decisions, summaries, insights, screen memories, and Knowledge Base chunks. **Loop** reads the relevant slice before it judges; **Chat** reads Memory through retriever skills; the result of every approved action is written back into Memory.
- **Loop** is the only thing that produces **Decisions**. A Decision is the typed judgement ("this email needs a reply", "this PR needs a review"), and a **Card** is how that judgement looks in the UI.
- The **Attention Agent** (also called **Loomi**, the **pet**, or the **fox**) is the messenger. It surfaces cards on the desktop. It does no judging of its own.
- You always tap **Approve** before anything runs. After approval, an **Action Runner** calls a **Connector** to actually send mail / post the comment / update the ticket.
- The result lands back in **Memory** (so the next Loop tick has sharper context) and in the **Audit Log** (so you can see who did what and when).

#### Worked example — "An email needs a reply"

> 9:12 AM. Sarah writes: "Hi — I tweaked tomorrow's Q2 review agenda, can you take a look? Also, I'd like to move our Wednesday 1:1 to Thursday same time — works for you?"

Step by step through the chain:

| Step          | Where it happens                                               | What is produced                                                               |
| ------------- | -------------------------------------------------------------- | ------------------------------------------------------------------------------ |
| 1. Ingest     | Gmail **Connector**                                            | A raw **Signal** — a new email event                                           |
| 2. Enrich     | **Memory** + **Screen Capture** (if you had related tabs open) | A context slice: who Sarah is, the Q2 project note, last thread summary        |
| 3. Judge      | **Loop** tick                                                  | A typed **Decision** — `email_reply`, confidence `0.85`                        |
| 4. Surface    | **Attention Agent** bubble, main-window queue, pet card        | A **Card** with subject, sender, draft preview, the four-button tray           |
| 5. Approve    | You tap **Approve**                                            | The **Action Runner** for `email_reply` is invoked                             |
| 6. Execute    | Gmail **Connector** (send)                                     | The reply is sent                                                              |
| 7. Write back | **Memory** + **Audit Log**                                     | Sarah's profile, the thread cluster, and the day's audit entry are all updated |

### Core concepts

These are the terms that anchor every other page.

#### Connector

A **Connector** is the integration boundary between OpenLoomi and an external platform — Gmail, Outlook, Telegram, WhatsApp, iMessage, Slack, GitHub, Linear, Calendar, Notion, and so on. Connectors handle both directions: pulling raw events out of the platform so the rest of the system can see them, and pushing actions back in once you approve them. See [Connectors](https://openloomi.ai/docs/connectors).

#### Signal

A **Signal** is one raw event produced by a **Connector** — "email arrived", "calendar invite received", "@mention on Telegram", "PR review requested". Signals are unopinionated: they carry the payload from the source platform and a shape so the rest of OpenLoomi can reason about them. They are the only thing Loop consumes. (Screen Capture is a separate input — see [Memory](#memory).)

#### Memory

**Memory** is OpenLoomi's long-lived context layer — people, projects, prior decisions, summaries, insights, uploaded documents, and **Screen Capture** memories — held locally on your machine. Both **Chat** and **Loop** read from Memory to ground their answers in your history, and approved Actions write back into Memory so the next judgement has sharper context. See [Memory](https://openloomi.ai/docs/memory).

#### Loop

**Loop** is OpenLoomi's proactive judgement engine. On a regular tick it pulls **Signals** from connected sources, reads the matching slice of **Memory**, and turns the ones that need you into typed **Decisions**. Loop never executes anything itself — it only "sees, thinks, and queues a card". Default cadence is every 10 minutes; tune it via `PUT /api/loop/preferences`. See [Loop](https://openloomi.ai/docs/loop).

#### Decision

A **Decision** is the smallest unit Loop produces — one signal becomes one structured judgement. A Decision carries a typed action (for example `email_reply`, `rsvp`, `review_pr`, `im_reply`, `todo`), a confidence score, a short preview of the proposed action, and any sender / subject / memory slice that's relevant.

#### Card

A **Card** is what a **Decision** looks like in the UI. A Card has the typed action, the confidence score, a preview body, the four-button tray (**Approve / Edit Draft / Later / Skip**), and — for any outbound action — a red "⚠ OUTBOUND" warning. Cards surface in three places at once: the main OpenLoomi window, the Attention Agent bubble, and the pet card. See [Loop — Approvals and dry-run](https://openloomi.ai/docs/loop#approvals-and-dry-run).

#### Action / ActionKind

An **Action** is the actual side effect that runs after you Approve a Card. The shape of an Action is fixed by an **`ActionKind`** literal — `email_reply`, `im_reply`, `calendar_rsvp`, `github_review`, `linear_review`, `todo`, `deadline_notify`, and a small set of others. Custom decision types register a label + icon, but always route to one of the built-in `ActionKind` runners.

#### Attention Agent / Loomi / Pet

The **Attention Agent** is the desktop companion — a 168×168, always-on-top, transparent little fox that lives on your desktop and shows Card bubbles when Loop queues something for you. **Loomi** is the fox's name; **Pet** is a generic nickname. The Attention Agent is **the messenger, not the judge**. See [Attention Agent](https://openloomi.ai/docs/attention-agent).

#### Audit Log

The **Audit Log** records the consequential moments of OpenLoomi's day: every Memory read/write, every Loop judgement, every Approve / Edit / Skip, every outbound Action invocation and its result, and every Connector authorization change. See [Privacy & Security — Audit Logs](https://openloomi.ai/docs/privacy-security#audit-logs).

### Supporting concepts

These are the surfaces and side capabilities that show up across the product.

#### Chat

**Chat** is the conversational entry point in the main OpenLoomi window. It solves "I have a question, a request, or a draft to write". It does **not** watch your day or queue cards — that's Loop's job. Chat reads Memory to ground answers and, when you tap **Edit Draft** on a Decision Card, you land here with the draft already loaded so you can refine it with the AI before Approving. See [Chat](https://openloomi.ai/docs/chat).

#### Automation

An **Automation** (also called a **Task** or **Scheduled Job**) is a prompt and schedule that **you** have defined in advance — "every weekday at 9 AM, summarise my unread inbox". Loop, by contrast, **discovers** Signals on its own and **judges** whether to surface them; Automation is the executor that runs whatever you told it to run. Both share the same scheduled-jobs runtime; Loop's own brief (9 AM) and wrap (6 PM) are themselves scheduled jobs. See [Automation](https://openloomi.ai/docs/automation).

#### Library / Knowledge Base

The **Library** is the user-facing surface; the **Knowledge Base** is what's behind it — the set of uploaded documents (PDF, DOCX, TXT, Markdown, spreadsheets, slides, images) that you want OpenLoomi to reason over. Library is **explicit, user-uploaded context** and is **not** the whole of Memory. See [Library](https://openloomi.ai/docs/library).

#### Skills

**Skills** are reusable capabilities the agent can call when it needs to do something specific — code generation, PDF creation, data analysis, browser automation, search, image generation, and many more. Inside a Skill, OpenLoomi runs through its **Agent Runtime**; Skills are how those runtimes are reached from Chat, Loop, or Automation. See [Skills](https://openloomi.ai/docs/skills).

#### Agent Runtime

An **Agent Runtime** is the underlying execution environment OpenLoomi uses when it needs a model that can act — Claude Agent SDK (the default), Codex CLI, OpenCode, Hermes, or OpenClaw. Runtime selection is deployment configuration. Agent Runtime powers Skills; **Plugins** (below) are the inverse bridge that lets those runtimes call OpenLoomi. See [Agent Runtimes](https://openloomi.ai/docs/reference/agent-runtimes).

From melandlabs/openloomi, Apache-2.0.

More from melandlabs/openloomi