Task scheduling for AI agents — cron, intervals, delays via MCP
schedulermcp has well-structured schemas and clear naming conventions, but suffers from incomplete parameter descriptions and missing output schema documentation. All 9 tools are explicitly registered with input schemas and descriptions. Naming follows verb_noun pattern (schedule_task, list_schedules, cancel_schedule, etc.). However, many parameters lack descriptions (e.g., 'job_id' in cancel_schedule, pause_schedule, resume_schedule, get_job, get_job_history, delete_job have minimal or no descriptions). Output schemas are not documented, LLMs cannot predict what fields to expect from responses. Error handling is generic (catch-all exception → JSON error text) with no recovery guidance. The schedule_task tool is complex with 11 parameters but lacks clear documentation of parameter interdependencies (e.g., when to use cron vs interval vs delay vs watch).
Cancel a scheduled job by ID.
Permanently delete a scheduled job
Get details of a specific scheduled job
Get execution history for a scheduled job
List all scheduled jobs with their status, next run time, and run history.
Pause a scheduled job. It can be resumed later.
Resume a paused scheduled job.
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination). Tools like list_schedules, get_job, get_job_history return data but no schema is visible in code.
Parameter descriptions missing or minimal. job_id appears in 6 tools with only 'Job ID' as description, no guidance on format, where to obtain it, or what happens if invalid.
schedule_task has 11 parameters with complex interdependencies (cron vs interval vs delay vs watch) but no documentation of which combinations are valid or mutually exclusive.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | <=2025-11-25 | v2 |
Schedule a task. Supports: once (delay), cron (recurring), interval, watch (event-driven trigger on condition).
Get scheduler statistics: total jobs, active, paused, run counts.
Error handling is generic. All exceptions caught and returned as plain JSON text with no recovery guidance, error classification, or actionable next steps for LLM.
delete_job is destructive but has no confirmation step, dry-run option, or warning in description. Agents can permanently delete jobs without safeguards.