Skip to content
Skip to content

The interface: session fleet, costs, Agentask

The web interface opens at http://localhost:47292 after agentop server. It is responsive, works on a phone and can be installed as a PWA. Portuguese and English, light and dark themes.

Session fleet

Section titled “Session fleet”

The sessions screen (/sessions) shows the fleet: the sessions running now and the conversations that can be reopened, from every harness.

  • On the left, the session list. Group it by project, repository, task, harness, model or status, and keep sessions in named groups.
  • In the centre, the fleet overview or the open session, with its live chat and terminal.
  • Permission dialogs: when a session stops on a question, the card shows the question and the options read off the session’s own screen, and you pick one. There is no blind “approve” button.
  • Two sessions side by side on large screens (one at a time on a phone).

Each session has a state:

State Meaning
working its screen moved since the last read
waiting alive and still, waiting for you
NEEDS APPROVAL a blocking question is on screen
exited the command finished; its last screen is still readable
lost known, but no longer running (after a reboot, for example); can be reopened
external an assistant running on the machine that agentop did not start

Telling a blocking question apart from an ordinary pause only works for Claude Code, Codex and Kimi. On the other harnesses, the question shows as waiting.

A reopened session comes back with its name, note and task.

Costs

Section titled “Costs”

The Costs screen shows spend by model and by date, budget and forecast, and cache efficiency. Cost and usage also appear per session, project, repository and harness.

  • It is an API-price estimate. If you pay a flat subscription, register your plan in Settings → Billing and see the same spend compared with what you actually pay.
  • Where prices come from. Three layers: a built-in table (the floor, which always works), a community dataset and the vendors’ own pages. A price that disagrees with the other sources is dropped. Settings → Pricing lists every model this machine has used, with where its price came from (official, community or builtin).
  • Tokens are always all four counters: input, output, cache read and cache write.
  • Dollars or reais. The conversion to BRL uses the current exchange rate.

Agentask

Section titled “Agentask”

Agentask (/tasks, in beta) is the task board. A task is measured through the sessions that worked on it: cost, rounds, tokens and harnesses all come from those sessions.

  • View it as a table or a kanban; dragging between columns changes the status.
  • File sessions directly on the task or under subtasks.
  • Statuses can be edited (new ones, renamed, with a colour).
  • blocked requires a reason or a blocking task; done requires at least one filed session.
  • A task nobody could price sorts last by cost, instead of showing up as zero.

From the terminal, use agentop task and agentop session batch --task. From your agents, use the agentistics_task_* tools of the MCP.

The other screens

Section titled “The other screens”
Screen What it is for
Home KPIs, activity over time, heatmap, hourly usage, highlights
Top usage what consumed the most
Repositories grouped by git remote, so the same repository unifies even across different folders
Tags saved groupings, such as “what did client X cost this month”
Tools tool-call ranking and tokens per tool
Compare harnesses side by side
Custom build your own layout by dragging components
Export a PDF report with the sections, period and theme you choose
Settings preferences, notifications, data and sources, backup, harnesses, pricing, billing

When a harness does not produce a metric, it shows as N/A, never as 0.