Claude Code Agent View Alternative, Self-Hosted: When to Graduate

What Claude Code agent view actually is

Run claude agents and you get a table. Every background Claude Code session on your machine, grouped by state: pinned, ready for review, needs input, working, completed. Each row carries a status icon, a one-line summary of what the session is doing, and how long it has been running. Press Enter and you are inside the full conversation. Press Space and you peek without attaching.

According to Anthropic's own docs, sessions run under a local supervisor process, so closing the terminal does not stop them. That process saves transcripts under ~/.claude/jobs/, and idle sessions get parked after about an hour, then woken back up when you reattach. You can dispatch new work straight from the dashboard with claude --bg "<task>", or push a foreground session to the background with /bg.

It shipped as a research preview on 11 May 2026, on Pro, Max, Team, Enterprise and API plans (claude.com/blog/agent-view-in-claude-code). "Research preview" is doing real work in that sentence. The docs themselves say the interface and shortcuts may still change.

Agent view vs a self-hosted cockpit

What each one actually does, not what the marketing implies. Checked against Anthropic's docs and the community reviews cited below, read 2026-09-23.

When agent view is enough

Be honest about this before selling anyone a bigger tool. If you are one developer on one laptop, running three or four Claude Code tasks and checking on them between meetings, agent view solves your actual problem. It is free. It ships with Claude Code. There is nothing to install, patch, or host.

CloudZero's rundown of 2026 orchestration options lines up the same way: independent tasks with no dependencies between them are exactly what agent view is for. Fix a bug in one service, review a pull request in another, watch a flaky test in a third. Dispatch all three, check back when they finish.

Don't reach for a cockpit because the word sounds more serious. Reach for one when the constraint changes: more than one person needs to see the work, or the work has to leave your laptop and land on a server, or code the agent wrote has to reach production through something more accountable than a summary you chose to believe.

Where agent view stops: the hard boundaries

Anthropic's own docs are specific about scope, and the community reviews confirm it from the outside. Three limits matter most.

It is single-machine. Agent view reads sessions from your own ~/.claude directory. It has no idea what your teammate is running on their laptop. A review from amux.io puts it plainly: agent view's scope is a single session or a single machine, full stop, and that is a hard boundary, not a bug to file.

Transcripts live on the machine that made them. They are useful right up until that laptop is asleep, off, or belongs to someone who left the company. There is no shared store, no server-side database, nothing a compliance review can point at as the record.

Nothing about production. Agent view will show you that a session finished and even that it opened a pull request. It has no concept of a dev, test, prod ladder, no migration step, no health check, and no audited override when someone wants to skip a step anyway. That is a different job, and it is not the job agent view was built for.

The migration signals: shared visibility, transcripts on a server, personas, and promotion

Four signals tend to show up together, and any one of them is reason enough to look past a local dashboard.

Shared visibility. The moment a second person needs to see what an agent is doing, a per-laptop tool stops working. amux.io's review names this directly: without a shared board, agents duplicate work or sit idle, and nobody notices until the standup.

Transcripts on a server. If a session's history needs to survive a laptop reboot, a resignation, or an audit request, it cannot live under one person's home directory. It needs a database someone administers on purpose.

Personas with boot prompts. A named subagent in agent view is a prompt someone remembers to type. A persona is a recipe: model, effort, boot prompt, skill packs, and MCP bindings, saved once and reused by anyone who is allowed to launch it. That difference matters more as the team grows past the person who wrote the original prompt.

Dev, test, prod promotion. Agent view tells you a session is done. It does not tell you whether the tree was clean when it finished, and it will not stop a broken commit from reaching production. That gate has to live outside the agent, on the server, where an agent's confidence cannot override it.

How Commander attaches to sessions agent view started

This is not a rip-and-replace story. Commander does not compete with the Claude Code daemon; it sits on top of it. When Claude Code holds a session in the background, Commander detects it through the same daemon roster agent view reads, and offers an Attach to button that opens the running conversation in a Commander terminal without losing its state. That matters because recent Claude Code releases refuse a plain resume on daemon-held sessions, so a direct attach is the reliable path.

From there, the session gets everything a laptop dashboard cannot give it: a transcript kept in your own PostgreSQL, a persona record of what it was launched with, a git panel showing what actually changed, and a place in a promotion ladder if the work is headed to production. Nothing about the underlying Claude Code process changes. Commander is the room the session runs in, not a replacement for the session itself.

Frequently asked questions

Related reading