Codeg

The trait that decides it is continuity across tools. A conversation you started in one agent can be picked up in another, which sounds minor until you have abandoned a half-finished thread because the tool you wanted it in was not the tool you started in. For anyone who has more than one agent installed, that is the daily annoyance this removes.

It is also a large product with a lot of surface. Pick it when switching is your habit. Pick Vibe Kanban if you only ever wanted the board.

The integration is unilateral. None of those tools publishes a stable storage contract, so every import path is reverse-engineered and every upstream release is a chance for one to break, quietly and one at a time. A workspace whose value is completeness degrades badly when three of fifteen sources stop parsing and nothing announces it.

What it does right is unattended containment. Each background task runs in its own checkout and waits for review before landing, so autonomous work cannot merge itself.

1. Cross-tool delegation is a genuinely novel capability and it rests on an unexamined premise: that the state one agent holds can be handed to another whose context representation, tool schemas and system instructions are different. 2. What survives that translation is the question, and nothing published describes the mapping.

2. No measurement exists. No comparison of a task completed within one agent against the same task delegated across two, which is the experiment the feature invites.

Aggregators occupy a structurally weak position: they add value by depending on products whose makers gain nothing from the dependency and can end it by changing a file format. There is no revenue here to negotiate with those vendors and no leverage to ask for a stable interface.

Moat: the breadth, which is exactly the thing that is expensive to maintain and cheap to erode. Likely path: continued volunteer maintenance chasing upstream changes. Position: useful today, and not a foundation, because the foundation belongs to fifteen other projects.

A single container image and a server reachable from a browser means one managed deployment rather than sixty installations, which is the shape I want and rarely get from projects like this. There is no licence cost at all.

The scheduling is where I hesitate. Saved configurations run on a cron inside this product, with no documented integration into our existing automation, so I acquire a second scheduler with its own failure modes and no shared alerting. Approved with conditions: one central deployment, and automations mirrored into our own monitoring.

Registration through an open agent protocol is the part that earns the score. It means the fifteen built-in integrations are a convenience rather than a boundary, and something I write myself joins on the same footing without a pull request being merged. That is extensibility done in the right direction.

Permissive licence, a one-line container start, and protocol client support so my servers attach. Token reporting per session is a nice touch for someone who checks. Very little here needs patching, which is my highest compliment.