Concepts
Workspaces, agents, runs, approvals, automations, integrations, models and features in one page.
Workspaces and actors
Every user has a private personal workspace and can create or join team workspaces with an
owner, admin or member role. Agents, automations, connections and model records belong to
one workspace.
An actor is the signed-in user, selected workspace, and role. The API reads identity from the session; it does not trust client-supplied user IDs. Services, tools, and model queries use the actor to enforce workspace access.
Agents
An agent is instructions plus a provider, a model and a list of enabled tools.
- Code agents are Markdown files in
server/agents. They are read-only in the dashboard and copied into each workspace on first use. - Saved agents are created and edited in the dashboard, with autosave.
Providers: OpenAI, Anthropic, Google and a scripted mock. During development, agents can also
run on your local Codex CLI or Claude Code login.
Runs
A run is one agent execution with a payload. Pro runs are durable: they continue after the tab
closes and stream status as server-sent events. Status is queued, running, waiting (for
approval), succeeded, failed, or cancelled.
Tools and approvals
Tools are what an agent can do: your defineAgentTool functions, model tools such as
customer_update, integration actions such as resend_send_email, and the built-in web_fetch.
A tool can require approval. The run pauses while a person reviews the exact arguments in the approvals inbox. An approved call runs once; replays return the stored result.
Automations
An automation runs an agent on a schedule, webhook, or event. Configure its instructions, timeout, retries, and success or failure outcome. Outcomes can keep the result, send a notification, call a webhook, or update the triggering record.
defineAutomation files are templates that a workspace installs as an inactive draft.
Integrations
Integrations connect OAuth accounts, APIs with keys, or remote MCP servers. Connections belong to
a person or workspace; owners grant agents individual actions. Credentials are encrypted with
AGENTIC_ENCRYPTION_KEY and never sent to the browser or model.
Agentic Models
defineAgenticModel(table, config) adds REST routes, typed services, agent and MCP tools, prompt
resources, automation outcomes, and a dashboard view to a Drizzle table. Access rules apply in
each place the model is used. See Agentic Models.
The prompt
The dashboard's command prompt answers questions about workspace data and proposes changes. Every write is shown for review and runs only after confirmation, against exactly the records that were reviewed.
MCP
Every app serves an MCP endpoint at /api/agentic/mcp for signed-in users. It exposes the same
tools with the same access rules and approvals. During development, agentic plugin connects
Claude Code or Codex to your running dev server.
Features and packages
The server is a set of features (identity, workspaces, integrations, mcp, prompt,
agents, automations, plus models). Each owns its tables, routes, services and tools.
Identity and workspaces are always on; the others can be switched off.
| Package | Role |
|---|---|
@nuxtagentic/nuxt | The free module (MIT): agents, tools, useAgent() |
@nuxtagentic-pro/core | Browser-safe contracts, client and prompt logic |
@nuxtagentic-pro/server | Fetch handler, features, services, models, Cloudflare port |
@nuxtagentic-pro/vue | The dashboard as Vue components |
@nuxtagentic-pro/nuxt | The Nuxt layer: dashboard pages, Nitro routes, DevTools |
@nuxtagentic-pro/void | The Void adapter |
@nuxtagentic-pro/cli | The agentic command |
All @nuxtagentic-pro/* packages share one version. Upgrade them together with
agentic upgrade.