Skip to content

MCP

agentistics ships an MCP server built into the binary (agentop mcp). With it, any MCP client can query your metrics, read and move Agentask tasks, file sessions into groups and build layouts for the Custom screen.

You register nothing by hand. Every time agentop server starts, it registers the MCP with every assistant installed on the machine (Claude Code, Codex, Gemini, Copilot), pointing at the binary itself. An up-to-date registration is left alone; a stale one is replaced.

Janela do terminal
agentop server
claude mcp list # should show "agentistics"

Because the MCP lives inside the binary, agentop upgrade updates the tools too. Sessions that were already open keep the old MCP until they are restarted.

The MCP talks to the local API. To point it at another address:

Variable Default
AGENTISTICS_API http://localhost:47291
Area Tools
Metrics agentistics_summary, agentistics_harnesses, agentistics_projects, agentistics_sessions, agentistics_costs, agentistics_repos
Agentask (beta) agentistics_tasks, agentistics_task, agentistics_task_create, agentistics_task_edit, agentistics_task_status, agentistics_task_statuses, agentistics_task_status_edit, agentistics_task_comment, agentistics_task_subtask, agentistics_task_link, agentistics_task_blocked_by, agentistics_task_next, agentistics_task_claim, agentistics_task_activity, agentistics_task_session, agentistics_task_delete
Session groups agentistics_session_groups, agentistics_session_group_create, agentistics_session_group_edit
Layouts agentistics_component_catalog, agentistics_get_layouts, agentistics_build_layout, agentistics_add_component, agentistics_remove_component, agentistics_create_layout, agentistics_set_active_layout, agentistics_delete_layout
Tags (read-only) agentistics_tags, agentistics_tag_detail

The metrics tools take an optional harness parameter (claude, codex, gemini, copilot, antigravity, kimi, or all, the default).

The loop an agent follows:

  1. agentistics_task_next: what can be picked up now, and what is withheld and why.
  2. agentistics_task_claim: reserves the task. The reservation expires after 30 minutes and is renewed by claiming again.
  3. Does the work.
  4. agentistics_task_session: files the session, so the task is measured.
  5. agentistics_task_status: updates the status.

In a Claude Code session:

How much have I spent this month, according to agentistics?
Which harness do I use the most, and which one costs the most?
Build me a layout in agentistics with the cost KPIs and the activity chart.
  • It has no gate of its own. Every tool runs on the first call, including the ones that delete (agentistics_task_delete, agentistics_delete_layout). The protection is your client’s tool-approval prompt.
  • It does not control the live fleet. No tool reads a session’s screen, sends a prompt, starts or stops sessions. For that, use the web interface, the cockpit, the VS Code extension or agentop session.