ggcode

The deciding trait is peer discovery. Start it and other copies on the same network appear, so you can message a colleague's instance, broadcast to the room, or hand a task to somebody else's agent without a relay, an account or a signup. Nothing else in this category treats a shared office as infrastructure.

That premise is also the limit: remote teams get none of it, and what remains is an ordinary terminal agent with a nice interface. Pick it if you sit together. Pick anything else if your team is spread across three time zones.

The dealbreaker is the trust model, or its absence. Anything on the local network can announce itself, and instances then exchange messages, share files and delegate work across the boundary. On a coffee shop network, a conference floor or a flat corporate VLAN, that is an inbound path to a tool holding shell access and a provider key. The row documents the mechanism and no control over it.

What it does right is refuse a middleman. There is no relay server holding the traffic and no account required to use it, which removes a vendor from a conversation that did not need one.

1. Resumable state means an interrupted task has a defined restart point rather than a fresh transcript, and that is the difference between recovery and repetition. It is the one part of the design that shows a long-horizon assumption. 2. The stated model support is unbounded, which is not a specification; tool-calling behaviour varies enough between families that a claim of universal support requires a table.

2. No evaluation is published and none is claimed, which is at least consistent with the rest of the row.

A registered domain and a single author is the most common shape in this category and the least durable. There is nothing to acquire, nobody to hire away, and no commercial surface that could fund the maintenance a networked tool needs, which is more than a solo agent needs because protocols age badly.

Moat: none. Distribution is a star count you could gather in a week. Likely path: it stops when the author's attention moves, and the interesting idea in it gets rebuilt elsewhere. Position: fine to try, wrong to standardise on.

The cost line is trivial and the questionnaire is not. Sixty developer machines exchanging traffic with each other is a data flow my security team has to describe, approve and monitor, and there is no central console producing a record of any of it. That work costs more than the tool saves.

There is no provisioning, no policy surface and no retention statement, and nothing here runs in a pipeline, so it never becomes a delivery metric I can report upward. Onboarding is easy, which is the wrong thing to be easy first. Not yet.

Somebody cared about packaging. Six channels including source means whatever my machine already uses, there is a path, and the permissive licence means the fork is mine if the author stops. Compiled and self-contained, so no runtime tags along and no interpreter version ruins my afternoon.

Then the ceiling. No MCP client, so the servers I run are outside; no local model support, so my own hardware is idle and every token leaves the building. For a tool whose entire personality is keeping traffic on the local network, sending inference to a cloud is a strange asymmetry.