HolyAgent

HolyAgent is the agent on your machines and the only thing installed on your side. Model keys, channels and memory live on the machine; only tasks, statuses and run transcripts reach the cloud, as you configure them. No telemetry.

Installation

Where	What to install
Mac (Apple Silicon)	the app from the release feed; it updates itself, data in ~/.holyagent survives updates
Server	the Docker image is being prepared for publication; write to us if you need it earlier
iPhone	the phone shell is in testing, access on request

The agent panel opens locally on the machine where it runs. You sign in to the app with HolyCode ID, the same account as the workspace.

Connecting to the organization

The app connects itself: you sign in with HolyCode ID, and on the default profile the setup wizard takes the account from that sign-in and registers the machine under it. There is no key to issue or paste. The machine appears in HolyCode under Agents as your agent. Step by step, with the wizard and the first channel: Connect an agent.

An agent for Workforce

An agent that takes tasks on Workforce scenarios runs without a window, usually on a server, and is connected with a one-time key:

1. In HolyCode: Workforce → Nodes → “Agent enrollment key…” and the agent name (lowercase Latin letters, digits, ., _, -). The key is valid for 72 hours, is shown once and works only for that name.
2. The agent needs only environment variables: the organization address, the agent name and the enrollment key.
3. The agent enrolls, receives a permanent key and appears in Nodes with a status and version. The one-time key is spent.

Optionally you can set the countries and languages the agent works with, the set of connectors and how many tasks it takes at once (5 by default).

Models

Model providers are connected on the setup wizard’s Authentication step and later in the agent panel: an API key, OAuth or a CLI sign-in for a cloud provider (Anthropic, OpenAI, Google, Cerebras); local Ollama and LM Studio are built in. Keys are stored encrypted on the machine. There can be several providers: the agent switches the model as the role and task settings say. Tokens are paid to the provider directly.

Channels

WhatsApp, Telegram, Instagram and voice are connected in the agent panel under Communication channels and run physically on this machine: customer conversations never pass through the cloud. The connection wizard shows what each channel needs. An outbound message from an r2 role is sent only after approval in the workspace.

Skills and memory

A skill is a folder with SKILL.md: what the agent can do and how. Skills are enabled per agent in the panel and move between installations with the company package (in Russian). Memory, what the agent knows about your business, lives in the agent folder on the machine.

Settings history

Every save of the agent settings writes a version (the last 50 are kept). Agent panel → settings → More → Settings history: what changed and a button to roll back to any version; the rollback itself becomes a version.

Several machines

Every machine is a separate agent. The app on each Mac registers itself when you sign in; an agent for Workforce gets its own enrollment key and is a separate node in Nodes. Roles and channels can differ per machine: for example, WhatsApp only on one machine and voice on another. A Workforce task goes to the node that has all the needed connectors.

For a machine to also build code projects, add a runner to it: see Connect an agent.