What is Cadence?

The short version

Cadence is a personal AI assistant that runs on your computer. Linux is first: download the desktop app, hatch an assistant, name it yourself. There is no seeded house persona. Bring your own model keys. There is no Aevum-hosted cloud hatch in 0.1.

It can have its own identity, its own credentials, and its own accounts once you connect them. It can read files, manage a calendar, send mail, browse the web, run code, and remember that Alice takes coffee black.

It is not a chatbot. It is not an autocomplete engine. It is a named assistant that works for you, learns about you, and takes real actions after you approve the sensitive ones.

How is this different from ChatGPT, Claude, OpenClaw, or Hermes?

Those are conversation tools. You type, they respond, you copy-paste the answer somewhere useful. The conversation ends and everything resets.

It has tools, not just words. Your assistant can browse the web, read workspace files, run commands, send email, and talk to services you connect. On Linux, host file and shell tools reach the rest of the machine after you grant host computer access. Sensitive actions always require approval first. It does not only describe what you could do. It does it.

It remembers you. Not just inside one chat. Across days and weeks: preferences, projects, corrections. Memory is structured knowledge, not a stuffed transcript. See Memory and context.

It has its own identity. The blank assistant you name is the product. When it sends mail on your behalf, the recipient should know they are talking to the assistant, not to you. Clear boundaries.

Your data stays on your disk. Workspace files, memories, and config live in the hatch instance directory. Plain text, exportable, deletable. There is no Aevum-hosted Cadence Cloud workspace in 0.1. See The workspace.

It is personal. Not a shared team bot. Not a Slack bot for the whole company. It is Alice's assistant or Bob's assistant, not a generic agent everyone uses.

Tools like OpenClaw and Hermes also run locally and expect you to bring keys. Cadence is the same on that point in 0.1: you configure a model provider. What you get with Cadence is a desktop app, a workspace of files, a trust and permissions model, and a skill catalog you can extend. You still own the plumbing. You do not sign up for a hosted brain.

Windows, Android, and Apple clients follow Linux. They are not the 0.1 ship target.

What can it do?

A non-exhaustive list. Many skills need a one-time credential or OAuth connect. Purchases, sends, and host access ask first. Live inventory is assistant skills list. Catalog: Skills reference.

Communication

• Gmail: inbox, drafts, archive, labels, unsubscribe (OAuth)
• Email (Agent Mail): give the assistant its own mailbox once you set a mail provider
• Messaging: read, search, and send across connected platforms
• Phone: Twilio voice and SMS after you bring Twilio credentials
• Slack: channels, threads, send (OAuth or bot token)
• Contacts: people, channels, and who may reach the assistant
• Notifications: a unified router across connected channels

Productivity

• Google Calendar: schedule, create events, availability
• Tasks: templates and a work queue
• Schedule: one-shot and recurring jobs (see Scheduling)
• Followups: track messages waiting on a reply
• Playbooks: trigger-action rules for inbound messages
• Document: long-form drafts
• Start the Day: a briefing from weather, calendar, news, and tasks you have connected

Automation and the machine

• Host tools: read, write, and run commands on the Linux host after you grant computer access
• Browser: navigate, click, type, extract via assistant browser
• Watcher: poll sources for changes
• Subagent: background work in parallel

Native click-and-type of a macOS GUI (Computer Use / Screen Watch) is a Mac client skill. It is not the Linux 0.1 path. Linux uses host tools plus the browser skill.

Development and media

• App Builder: interactive apps in the conversation
• ACP: delegate coding to an external agent over the Agent Client Protocol
• Image Studio: generate and edit images when you have an image model configured
• Media processing / Transcribe: ingest video, audio, images; Whisper for transcripts

Services

• Amazon and DoorDash: shop or order after you connect those accounts. Purchases need approval.
• Weather: current conditions and forecasts
• OAuth integrations: GitHub, Notion, Linear, and others via the credential vault
• MCP servers: extra tools you add

If it cannot do something, you can teach it. See Your first skill and Skills and tools.

Where does it run?

Linux desktop (0.1)

Download the AppImage or apt package from www.aevumcadence.ai. Hatch a local assistant. Bring your own keys. Workspace and memory stay on disk. You pay your model vendor directly. See Installation and Local hosting.

Later clients

Windows, Android, and Apple follow Linux. A browser UI can talk to a local or self-hosted assistant. There is no hosted web hatch in 0.1.

Channels

Beyond the desktop app, you can reach the assistant on Telegram, Slack, SMS, or phone once those channels are configured. See Channels.

How does it work? (30 seconds)

1. You talk. In the Linux app, by voice if you set it up, or through a channel.
2. It thinks. Your message plus relevant memory, workspace files, and skill instructions go to the model provider you configured.
3. It acts. Tools run for files, browser, mail, and so on. Host and high-risk actions wait for Allow or Deny. You can set standing trust rules.
4. It learns. Useful facts land in memory and workspace files so later turns are better.

Transparency: the assistant thinks by talking to an AI model provider. Prompts and context leave your machine for that generation step. That is the trade-off of a smart assistant. We would rather say it here than hide it.

Is it safe?

Cadence has a built-in trust system. Sending mail, spending money, and touching the host require explicit approval unless you write a trust rule. Credentials live in the credential vault, not in chat and not in plain workspace text. See Permissions and Privacy and data.