Skip to content

MCP Tool Reference

Currently, the MCP surface provides 24 tools. Instead of just listing names, this document explains what problem each tool solves, organized by category.

1. Task Lifecycle

  • monitor_task_start
  • monitor_task_complete
  • monitor_task_error
  • monitor_task_link
  • monitor_runtime_session_ensure
  • monitor_runtime_session_end
  • monitor_session_end

This group handles task/session creation/termination and background lineage connections. It's characterized by having both runtime-scoped and explicit session-end paths.

2. Event Logging

  • monitor_tool_used
  • monitor_terminal_command
  • monitor_save_context
  • monitor_plan
  • monitor_action
  • monitor_verify
  • monitor_rule
  • monitor_explore
  • monitor_user_message
  • monitor_assistant_response

Most of the timeline is generated from this group. In particular, monitor_user_message and monitor_assistant_response are canonical paths for conversation boundaries.

3. Semantic Flow / Coordination

  • monitor_async_task
  • monitor_agent_activity
  • monitor_question
  • monitor_todo
  • monitor_thought

This group records more structured semantics than simple logs. It preserves information like background tasks, delegation, question flow, todo state, and summarized thoughts.

4. Workflow Library

  • monitor_evaluate_task
  • monitor_find_similar_workflows

It's a feature to save good examples after work is done and find similar workflows in the next task.

Common Combinations in Practice

Manual Runtime Path

  • monitor_runtime_session_ensure
  • monitor_user_message
  • monitor_explore / monitor_terminal_command / monitor_plan
  • monitor_assistant_response
  • monitor_runtime_session_end

Claude Plugin Supplementary Path

  • Manual augmentation tools like monitor_rule, monitor_verify, monitor_async_task when auto hooks miss something

Task Evaluation Path

  • Confirm user's intention to evaluate with monitor_question
  • Save with monitor_evaluate_task
  • Search with monitor_find_similar_workflows on next task start

Internal Transmission Method

All tools except the 7 lifecycle tools are sent to the POST /ingest/v1/events batch endpoint. Each tool wraps a single event in the form events: [{ kind: "...", ...input }].

Toolkind
monitor_tool_usedtool.used
monitor_terminal_commandterminal.command
monitor_save_contextcontext.saved
monitor_planplan.logged
monitor_actionaction.logged
monitor_verifyverification.logged
monitor_rulerule.logged
monitor_exploretool.used + lane: "exploration"
monitor_async_taskaction.logged + asyncTaskId
monitor_agent_activityagent.activity.logged
monitor_user_messageuser.message
monitor_assistant_responseassistant.response
monitor_questionquestion.logged
monitor_todotodo.logged
monitor_thoughtthought.logged

monitor_session_end uses the lifecycle path (/api/session-end) directly.

Maintenance Notes

  • When adding a new tool, check the kind list in schemas.ingest.ts, dispatch in event-ingestion-service.ts, and guide documentation together.
  • Keep tool names in snake_case format, and use clear verbs for user-facing descriptions.
  • Since it's manual registration without a declarative manifest, lists and implementations can easily diverge.

Local-first documentation for Agent Tracer.