Agents overview
Agent types, capabilities, and how agents run on tickets and in chat.
Agents overview
AI agents in ArxDeck are configured assistants — not user accounts — that launch provider runs on tickets, in project chat, or from module actions. Each agent specifies a provider, capabilities, prompts, optional MCP tools, and optional skills.
ArxDeck Helper can help project admins configure agents through chat proposals. For manual setup, see agent settings in the dashboard and ArxDeck Helper.
Provider types
| Provider | Typical use |
|---|---|
| OpenRouter | Project chat and lightweight sync automations; model picker from OpenRouter catalog |
| Cursor Cloud | Async repo-backed runs on tickets; PRs and branch metadata sync to the ticket |
| Claude Managed | Async managed sessions on Anthropic infrastructure |
| Local | Windows desktop client executing via Cursor ACP on the developer machine |
OpenRouter agents focus on chat and text automations without repo execution. Cursor Cloud, Claude Managed, and Local agents are worker paths tied to repositories or local environments.
Orchestration targets such as Cursor Cloud, Claude Managed, OpenRouter, and the local desktop agent are supported in the product; this documentation set does not include dedicated setup guides for each provider in v1. Use dashboard agent settings and provider consoles for credentials and repository access.
Capabilities
Agents declare capabilities that control where they appear:
- chat — project chat picker
- automation — workflow Run agent actions
- worker — ticket worker runs (Cursor/Claude/local)
- action — content module custom actions
Can be tagged controls @mention in ticket comments. Reuse context tries to continue the same provider session on follow-up triggers for the same ticket.
Global vs project-local
Like workflows, agents can be global (superadmin-managed, delivered via content modules) or project-local (project admin-managed).
Runs on tickets
When an automation or @mention launches an agent:
- ArxDeck tracks the run with provider status, external IDs, PR URLs, and cost when available
- Optional workflow transitions fire on start, success, or failure
- Agents with MCP can call
report_agent_outcomeorrequest_user_inputto drive further automations - Terminal comments summarize outcomes when the provider returns assistant text
Ticket lists show compact agent activity; ticket detail shows a full run panel with sync, cancel, and retry actions where permitted.
MCP
Agents can enable the ArxDeck MCP server and optional external MCP servers registered at the organization. Run-scoped tokens limit tool access to the active run. See MCP integration.