Agentara

The deciding trait is that the agent stops being something you sit in front of. You send it work from a chat window on your phone, it runs on the machine under your desk, and the answer comes back to the same thread. That changes what you are willing to hand it, because the cost of asking drops to a message.

What you give up is polish. You install it from a clone, and it only runs on macOS or Linux. Pick it if you want a resident assistant on hardware you own. Pick a plain terminal session if you would rather not run a daemon.

The failure mode is a quiet stall. Work is queued and dispatched one item at a time within a session, which is the right ordering choice, and the documentation describes no watchdog, no timeout and no stuck-session detector to go with it. A backend that hangs waiting for input holds its queue until a person notices, and the whole point of the product is that nobody is watching.

What it does right is refuse to reimplement the agent. It drives the vendor CLIs directly, so the tool loop underneath is the one those vendors test and ship.

1. Persisting the complete transcript in a line-delimited format is the correct archival decision: it is appendable during a run, greppable afterwards, and it makes a disputed result reconstructible instead of merely remembered. 2. The system deliberately owns no reasoning of its own. It supervises and schedules; the tool loop, the context assembly and the verification all happen inside the delegated process.

2. That is an honest boundary, and it means the architecture cannot be evaluated separately from whichever backend is configured. No measurement is published, and none would be meaningful without naming the model underneath.

Start with the paperwork, because the paperwork is the asset. A badge claiming a permissive licence with no corresponding file gives an acquirer's counsel nothing to diligence and gives a corporate user nothing to rely on. That is not malice, it is a side project, and side projects are valued accordingly: five hundred stars of interest and no entity behind them.

Moat: none. Likely acquirer: none, though the pattern it demonstrates gets copied into a funded harness within a year. Position: run it, do not depend on it, and read the repository before your employer's counsel does.

The economics are simple and the governance is absent. There is nothing to buy, and the model spend lands on whichever vendor subscription each engineer already holds, so finance sees sixty unchanged invoices and I see sixty unmanaged daemons. Chat-channel access means work can be dispatched into a repository from a phone with no identity check I control.

There is no directory integration, no retention policy and nothing that plugs into a build pipeline. Onboarding is a source build per machine. Not yet, and not on any repository that matters.

The cost column is easy: it authenticates as me, through the CLI logins I already pay for, and adds nothing of its own to the bill. The bundled skill files sit in the repository as plain markdown, so changing what the assistant knows how to do is an edit and a restart rather than a feature request.

The gap is protocol. It is not an MCP client, so every server I run has to be wired into the backend CLI instead of into this. Install is make install and make dev from a clone, which at least means I can patch it.