What a workspace contains
| Resource | Purpose |
|---|---|
| Connections | MCP servers, REST APIs, databases the AI can call |
| Knowledge | Documents, pages, and text for context |
| Guardrails | PII filtering, cost caps, rate limits |
| Consumers | Entry points: Slack channels, API keys, WhatsApp numbers |
| Permissions | Role-based tool access patterns |
Configuration
Each workspace has:- Name: display name in the dashboard
- Team: the team that owns this workspace
- Model: the language model tier to use (e.g.
balanced,powerful) - System prompt: instructions the AI follows for every query
- Max tool rounds: how many tool-call loops the agent can run (default: 10)
- Cold timeout: minutes of inactivity before a conversation goes cold
- Close timeout: minutes after going cold before a conversation is closed
Lifecycle
Workspaces have three statuses:| Status | Meaning |
|---|---|
active | Normal operation. Queries are processed, consumers are live. |
paused | Queries are rejected. Consumers stop listening. Data is preserved. |
archived | Read-only. Historical data accessible but no new queries. |
Isolation guarantees
- Data isolation: conversations, audit logs, and costs are scoped to a single workspace
- Tool isolation: each workspace discovers tools from its own connections only
- Consumer isolation: a Slack channel can only belong to one workspace
- Knowledge isolation: documents are indexed per workspace
- Compliance isolation: guardrails are set per workspace (org-level baseline cannot be weakened)
Stats
The workspace dashboard tracks:- Queries today, this week, this month
- Cost month-to-date
- Average response time
- Error rate
- Connection count and tool count