The console

The Hub serves a web console at its address (http://127.0.0.1:7070/ with agen up). agen ui prints a sign-in link; any API token works, and the console shows and allows what the token’s scopes and namespaces allow.

agen ui

The console updates live: the Hub streams changes (tasks, runs, model and tool calls, approvals, instances, deployments) to it, so lists, traces and the fleet map change as work happens. Press Ctrl+K anywhere to jump to a page or a deployment, or paste a trace or task id.

Overview

The first page answers “what is happening, and does anything need me?”:

Needs attention: counts by kind and the three most serious items, each with what to do about it. Kinds include approvals waiting for a decision, deployments with failed runs in the last 24 hours (delegated calls included), daily budgets that are nearly or fully used, failed instances or tool servers, deployments waiting for room because every Nest is full, and live warnings: runs near their turn limit or run budget, runs that have been quiet for 5 minutes, long approval waits, queues waiting or blocked behind a pending approval, bursts of tasks in one conversation, model error rates, held provider accounts, missed schedules and webhooks refusing calls. The full list, paged, is in the Inbox.
Fleet map: who calls whom in the last 24 hours.
- Where work enters (API and CLI, schedules, webhooks, A2A callers outside the fleet) is drawn on the left.
- Calls are solid lines, work started with start_agent is dashed, and results going back are dotted; an agent calling itself is a loop. A line’s colour shows the share of calls that failed; click it for counts.
- Dots travel along a line while a call is in flight (a red diamond when it failed), and an agent pulses while it waits for its model.
- Select an agent or a line for details: what an agent runs now and its backlog, or a line’s calls, errors, p95 latency and the target’s latest failures.
- Above 12 agents the cards get compact and tall columns wrap. Drag to pan, Ctrl + wheel or the buttons to zoom, and Fit to see everything. Find an agent highlights it and its direct neighbours; agents with nothing in 24 hours can be hidden.
Now: every run in progress, nested under the run waiting for it, with what it is doing (model, tool, waiting for an agent or an approval) and for how long.
Activity: tasks as they are queued, run and finish, grouped by conversation.
Timeline: runs per agent over time, to spot overlaps, bursts and long runs.
Latest traces.

Inbox

Approvals that agents wait for, with the tool, its arguments, who asked and what the agent was doing (from its transcript), and tasks that failed since your last visit. The person whose work asked for an approval cannot decide it (see Permissions). The inbox also has the full list of everything that needs attention, paged.

Runs and traces

Runs lists one row per trace by default: the run a task or call started, with the tokens, cost and number of agents of everything it delegated. Filter by deployment, status, time and label; switch to All runs to see delegated runs too.

Opening a run shows its trace:

Timeline: every span across agents as a tree with bars on a shared time axis: agent runs, model calls, tool calls, delegated calls (with the callee’s run nested under them) and approval waits. Spans that are still running show as they happen. Find a span by name, or show only failures; the arrow keys move through the tree.
Inspector: select a span to see what happened in it. A model call shows the messages it sent and the reply it got (text and tool calls), its tokens, cost and finish reason. A tool call shows its arguments and its result, the permission decision and whether a side effect was reused. A run shows its input, output or error, usage and labels.
Transcript: a run’s whole conversation, tool calls with their results, and earlier messages of the conversation it continued.

Explain asks the assistant to walk through the trace.

The same data is in the API and CLI: see Traces.

Deployments

A deployment’s page shows its last 24 hours (runs, failures, duration, tokens, cost), where it stands against its limits (spend today, queued tasks with how many are scheduled for later and how long the oldest has waited, instances), its instances with their tool servers, and tabs for runs, tasks, logs (live), the bundle’s files and triggers. A deployment whose provider account is held says provider held (see Models).

Run a task submits one, with an optional conversation key, labels, a start time (the task waits until then) and a model tier (for agents on a routing model).
A task’s details show its start time, model tier (and whether it ran near the daily budget), the task that started it and where an async result goes. Cancel task cancels one that has not finished.
The Bundle tab lists earlier versions of the bundle. View shows a version’s files; Roll back to this deploys that version’s files again, and instances restart on it.

New agent (on Deployments) writes a bundle from a short form: name, description, instructions, and the model: one OpenRouter model, a routing model (Jev Router or Auto Router) with its cost tier, highest tier and a fallback model, or scripted replies for trying things out. Edit bundle opens a deployment’s files in an editor; saving checks the bundle and rolls the new version out (a broken bundle is refused with the reason).

Schedules

Everything that starts work on its own: cron triggers with their next and last run (Run now), one-off tasks waiting for their start time (Cancel), webhooks (New secret), and missed and refused trigger events. See Triggers.

Costs

Spend per deployment over 24 hours, 7 or 30 days, with runs, failures, duration percentiles and tokens, and spend over time. Cost needs a provider that reports it (OpenRouter does).

Models: calls by the model that served them (a routing model or a fallback may serve another model than the one configured), with failures, fallbacks, latency, tokens and cost.
Provider accounts: accounts held after an account error, with the reason and the next try, and Resume (admin). See Models and providers.

Templates

Ready-made agents to start from: a scripted hello agent, a web researcher, a writer, an editorial team that delegates to both, a pull request reviewer, coding, CI, support, incident and data agents, and operations agents (fleet steward, approval triage). Use opens a form for the name, model and instructions, sets the secrets the template needs and deploys it. See Templates.

Fleet

The Nests that run instances, with their capacity and instances. For admins, Add a nest shows the agen nest run command to run on the new machine, with a fresh join token (it works once, for an hour) and, when the Hub uses TLS, the CA hash to pin; Revoke removes a Nest, and its instances move to other Nests.

Assistant

The assistant is an agent that works for the person chatting with it: it answers questions about the fleet and makes changes, through the Hub’s MCP tools, with that person’s permissions. Before a change it checks it (validateOnly where there is one), says what it will do and waits for a yes.

An admin sets it up from Templates → Console assistant. That deploys the assistant template and gives it a token created with --on-behalf: such a token only acts for the submitter of the task it works on, never has more scopes or namespaces than that person, never has admin rights and can never decide approvals. Each answer links to its trace.

While it works, the chat lists each step (the tools it calls, with their arguments); a finished answer keeps them under Steps. Drag the panel’s left edge to widen it. History lists your past chats in this browser; opening one continues that conversation, since the assistant remembers each by its conversation key.

Settings

Your token’s identity and scopes; for admins, API tokens, platform secrets and notification targets.