Shippie

The deciding trait is that there is no service. It ships as a prebuilt workflow you drop into the automation you already run, so nothing new gets access to your repository and nobody new appears on the invoice. For a small team that has already lost an argument about third-party bots, that changes the conversation entirely.

You are the operator, which means upgrades and failures are yours. Pick it when control matters more than convenience. Pick PR-Agent when you would rather someone else ran it.

The exposure is the combination. A key sits in the repository's secrets, an agent with shell access runs in the same job, and the review is triggerable from a comment. On any project taking outside contributions, that chain lets an untrusted change influence a privileged execution, and nothing documented restricts which events or which authors may start one.

What it does right is specificity. The finding categories are named up front, exposed secrets, inefficient code, unhandled edge cases, so the output has a shape a team can calibrate against.

1. Context is gathered rather than supplied. The agent uses tools to explore the surrounding codebase and delegates to subagents, so a finding can rest on a definition three files away instead of on whatever fitted in a prompt. 2. That is the architecturally correct answer to the truncation problem every diff-only reviewer has.

2. It also costs more per review, and no measurement of that trade is published: no precision figure, no token accounting, no comparison against the single-pass approach it replaces.

Three years in the same category with a rename in the middle is persistence rather than traction. There is no company, no commercial tier and no revenue to fund the ongoing work of keeping pace with funded competitors who ship weekly. Stars measure goodwill, and goodwill does not pay for a second maintainer.

Moat: none, though the permissive licence means users are never stranded. Likely path: continued solo maintenance, or the author's attention moving to whatever comes next. Position: adopt it freely, and keep the workflow simple enough that replacing it later is a morning's work.

Cost is inference and compute minutes we already buy, so there is no seat negotiation and no new supplier in the estate, which removes a security review rather than adding one. It runs unattended in the automation both of our code hosts already use, so it becomes a required check rather than a habit.

What is missing is central visibility. There is no console, so configuration lives in each repository and drift across sixty developers' projects is invisible to me. Approved with conditions: one shared configuration, owned by the platform team, and no per-repository forks of it.

The protocol support is what lifts this above the other review bots. It consumes servers, so a reviewer can consult my documentation server or drive a browser without anybody writing an integration for it, and the tools I built for other purposes suddenly earn a second use.

Four providers are supported including a routing service, so the model choice is mine and cheap to change. Permissive licence, no service in the middle, and a single command to start. This one I would keep even if the maintainer disappeared, which is the point.