AI Hub

The AI Hub is the conversational home of your automation workspace. It gathers every way you talk to AI in ByteChef — the general-purpose assistant, chat-enabled workflows, and deployed AI agents — into one sidebar of chats, with shared attachments, memory, and model selection.

What lives in the AI Hub

The sidebar

The left sidebar is the AI Hub's navigation. Above the chat list sit two entries:

• New Chat — returns to the home view with an empty composer.
• More — a menu linking to Connectors, Memories, and Skills.

Below them, Chats lists your own chats in the current workspace and environment — chats are per-user, so you never see another user's threads — with a search box and an Archive toggle.

Starting a chat

The composer's provider popup — the model picker in the composer, on the home view — is the single launcher for every kind of chat. Opening it shows a provider search box and then two cascades above the provider list:

• Agents — every deployed AI agent (from Build → Agents) whose enabled deployment exposes a hosted chat trigger, listed by the agent's title. When none qualifies, the submenu says so rather than disappearing.
• Workflows — every chat-enabled workflow deployed in the selected environment, listed as project — workflow. Every pick starts a fresh chat; earlier chats against the same workflow stay in the list.

Agents live in hidden backing projects, so an agent never also appears under Workflows — the two lists are disjoint.

Below the cascades, the same popup selects which provider and model answers the chat.

Working in a chat

• Model picker — choose which activated AI provider and model answers each chat; different chats can run on different models. This is the same popup that carries the Agents / Workflows cascades, and it is wired into both the home composer and the in-chat composer.
• Attachments and files — the paperclip button (and drag-and-drop) upload files into a chat, and anything the assistant produces (files, charts, data tables, knowledge bases, workflow executions, workflows, custom components, code workflows) opens in a viewer panel beside the chat.
• Reference resources — the + button opens the Resources menu: a picker for referencing an existing file, workflow, data table, knowledge base, or execution in the message, plus Connectors and Skills branches. The reference is added as context the assistant can read and opens in the right-hand panel immediately — including on the home view, before the chat itself exists.
• Connectors — the Resources menu's Connectors branch lists the component connectors available to you, each with a switch controlling whether it acts in this chat, plus a Manage connectors link to add more or configure their tools. Enabled connectors let the assistant act in those apps on your behalf. On the home view there is no chat yet, so the switches are absent and the list reads as availability only.
• Specialist agents — alongside the general assistant, the AI Hub routes to focused Ask + Build specialists. The buildCustomComponent tool lists, explains, creates, and updates custom components, and the buildCodeWorkflow tool does the same for code workflows. Referenced custom components and code workflows appear in the chat sidebar and open — editable — in the right panel.
• Asset files — the assistant lists and reads asset files directly; creating or editing a file (writing new content, replacing a file's content, downloading from a URL, storing binary content, or cloning a file into another environment) routes through the asset_file_agent specialist.

• Memories — durable facts the assistant learns are saved automatically and listed on the Memories page (sidebar → More → Memories), where you can review and delete them. The same auto-memory capability is available inside workflow AI Agents via the Agent Utils Auto Memory tool, scoped to the deployment instead of the user.
• Skills — type / at the start of the composer to open a picker over your skills (the same set as the AI → Skills page), or use the Resources menu's Skills branch. Either way the picked skill becomes a chip above the input rather than text in the message, so several skills can be armed for one turn, and the assistant consults their content to steer the run.
• Chats sidebar — every chat is listed in the collapsible sidebar so you can return to or archive past threads.

Managing connectors and MCP servers

The Manage connectors link at the bottom of the Resources menu's Connectors branch opens the AI Hub's connectors page, where you decide which tools the assistant can call. It has two lists.

Component connectors

1. Click Add connector and pick a component from the grid.
2. If the component needs a connection and none exists yet, its row shows a Connect action — click it to create or choose a connection in the Connect dialog. (The row's ellipsis menu also offers Connect to change the connection later.)
3. Toggle the connector on/off to control whether the assistant may use it at all.
4. Expand the row's tools chevron to enable or disable individual actions, and click a tool's Configure (gear) button to preset parameter values for it, so the assistant only has to supply the rest.
5. Remove a connector from the ellipsis menu's Delete.

Custom MCP servers

1. Click Add MCP server and fill in a Name, the Server URL, and an optional Bearer token (stored encrypted server-side).
2. Expand the server row to see its tools, which are discovered live from the server; toggle the whole server or individual tools on and off.
3. Remove a server from its row's delete action.

Per chat, the Resources menu's Connectors branch still lets you toggle any enabled connector on or off for just that chat — this page sets what is available in the first place. A connector disabled here is unavailable everywhere and drops out of the per-chat list entirely.

How the assistant's tools are wired

Under the hood, tools reach the AI Hub agents through three tiers. A small pinned set of interaction primitives is callable on every model turn; a much larger searchable catalog is indexed in a vector store and becomes callable only after the agent looks a tool up; and self-contained domains are delegated to one-shot specialist subagents, each a focused agent with its own prompt and tool set.

This tiering is why the assistant sometimes "finds" a capability mid-chat (a catalog lookup) or narrates that a specialist is working on a sub-task (a delegated subagent). Component connectors and custom MCP servers you enable in a chat add their tools on top of this wiring.

Enabling the AI Hub

The AI Hub is an Enterprise Edition surface, disabled by default:

BYTECHEF_AI_HUB_ENABLED=true

It uses the models you activate in AI Providers, so at least one chat provider must be configured for the environment.

The workspace's AI Guardrails policy also applies to AI Hub chat turns — the same PII/secret redaction, blocked terms, moderation, and injection detection enforced on the AI Gateway and canvas AI Agent.