Multi-step, identity-aware workflows with approval gates, conditional branching, and parallel execution — security enforced at every step
Submit expense → auto-approve under threshold → manager approval if over → process payment → notify
Search flights → check budget → approval if over limit → book → charge card → send confirmation
Alert → assess severity → page on-call → escalate if not acknowledged → resolve
Create accounts → assign groups → equipment approval → schedule orientation → welcome email
Visual Designer: Drag-and-drop steps in the Workflows tab. Add tool calls, conditions, approvals, notifications — wire them together visually.
JSON/YAML: Power users can write workflow definitions as code. Steps, triggers, branches, retry policies — all declared in a structured format.
Templates: Start from a pre-built template and customize. One-click install from the template gallery.
Manual: User clicks "Run" in the dashboard or types "start expense approval" in chat.
Webhook: External systems (monitoring alerts, form submissions, API calls) trigger workflows via POST /api/workflows/:id/run.
⏰ Schedule (Built-in Cron): APScheduler fires workflows on a cron schedule with scoped identity:
→ User scope: Runs as the user who created it, using their vault credentials
→ Group scope: Runs with group-level credentials (team lead creates)
→ Org scope: Runs with org-level credentials (manager only)
Presets: hourly, daily 9am, weekly Monday, weekdays, monthly. Custom cron supported.
Schedules persist to Cosmos DB and auto-restore on pod restart.
Tool call steps are not simulated — they execute against real MCP engines via agent.call_tool():
1. Tool lookup — finds which MCP engine has the requested tool (searches all registered servers)
2. Identity injection — _user_email, _user_role, _user_groups from the workflow trigger
3. Credential resolution — vault cascade (user → group → org) per the triggering user
4. Policy check — role ≥ min_role? write allowed? rate limit OK?
5. MCP execution — calls the plugin function against the external API
6. Result stored — output saved as workflow variables for subsequent steps
Retry: Configurable per step with exponential backoff. If all retries fail, workflow follows error path.
The workflow pauses and sends a notification to the designated approvers. They can:
✅ Approve — workflow resumes to the next step
❌ Reject — workflow follows the rejection path (notify user, log reason)
⏰ Timeout — if no response within the SLA, automatically escalate to the next level (e.g., director)
Multiple approvers: Configure "require 2 of 3" or "require all" policies.
When all steps finish, the workflow engine:
1. Marks the run as completed (or failed/cancelled/timed_out)
2. Saves the full execution history — every step, every result, every approval decision
3. Persists to Cosmos DB (or DynamoDB/Firestore depending on cloud)
4. Available for audit trail — who triggered it, who approved it, what was the outcome
Executes any MCP tool — email-send, payments-charge, travel-search, github-create-issue, etc. Arguments can use {{variables}} from previous steps. Configurable retry count with exponential backoff.
Pauses the workflow. Sends notification to approvers. Supports: single approver, multiple approvers (require all or N-of-M), timeout with auto-escalation, rejection path.
Evaluates an expression against workflow variables. Routes to "true" path or "false" path. Examples: amount > 500, severity == 'critical', department == 'engineering'.
Fans out to multiple branches simultaneously. Waits for all branches to complete (or N-of-M). Example: send approval to 3 managers, continue when 2 approve.
Sends a message via email, Slack, Teams, or SMS. Uses {{variable}} templates. Sends through the user's own configured channels (their Gmail, their Slack).
Pauses the workflow for a specified duration. Use for: follow-up reminders, SLA waiting periods, rate-limiting between API calls, scheduled check-ins.
Validate → check threshold ($500) → if over, get manager approval (24h timeout → escalate to director) → process payment via Stripe → email confirmation. Under $500 auto-approves.
Search flights → check against budget → if over budget, get manager approval → book flight → charge card → email itinerary with booking reference. Uses hierarchical RAG for seat preferences and airline policies.
Assess severity → if critical, page on-call (15-min ack timeout → escalate to engineering manager) → if non-critical, notify team channel → update incident status.
Create email account → assign department groups → get manager approval for equipment → schedule orientation meeting → send welcome email to new hire → notify manager that onboarding is complete.