Adnify

The deciding trait is that Plan Mode will not start until you have read the plan. A request becomes a task graph you inspect, approve or send back, and only then does anything get written. If you have ever watched an agent confidently do the wrong six things in a row, that pause is the whole product.

What you pay for it is speed, and a desktop app you install from source rather than a store. Pick it when the cost of a wrong change is high. Pick Cursor when you would rather move fast and read the diff afterwards.

The risk is that the guard rails are lists. A whitelist of git subcommands protects against the commands somebody thought of; a shell tool that runs on the host protects against nothing at all. There is no container between the agent and the filesystem it was pointed at, so approvals are the only thing standing between a mistaken plan and a working tree.

What it does right is the diff preview. Edits are scoped and shown before they land, which is the correct default and one many faster tools skip.

1. Retrieval is two-sourced: a language server supplies symbol-accurate structure while a vector index over the indexed repository supplies similarity, which are different failure modes and therefore complementary. 2. The pipeline ends in a declared validation stage. Most designs stop when the model says it stopped; naming verification as a step means the loop has a defined terminal condition rather than a conversational one.

2. None of this is measured. No evaluation accompanies the architecture, so the claim is structural rather than demonstrated, and a reader should treat the task graph as a design argument.

The interesting asset is the licence itself. A custom agreement that reserves commercial use is a founder keeping optionality open: it converts every business user into a conversation, and a conversation is where pricing power starts. Two hundred and seventy-three stars is not distribution, though, and there is no entity, no tier and no revenue line to inspect.

Moat: the reserved rights, which is thin. Likely acquirer: nobody buys this; an editor vendor hires the author and the repository goes quiet. Position: use it personally, and do not let a team depend on one person's calendar.

The line item I cannot fill in is the one that matters. Personal, research and non-profit use is free; commercial use requires permission from the author, which means the price for sixty seats is whatever an email negotiation produces. That is worse than a published number, because I cannot budget it and legal cannot review it.

Operationally it is sixty desktop installations with no console, no directory integration and no audit trail I can hand an auditor. It does not run in a pipeline, so it never becomes a measurable step. Not yet.

Model routing lives in the kernel, so I point it at a local Ollama endpoint and nothing leaves the box except what I choose to send. MCP servers plug in. The source is public and I build it with pnpm from a clone, which means I can read every tool policy and budget rule before I trust one of them.

Then I hit the licence. Public source under a bespoke agreement is not open, and a fork that survives the author is exactly what that agreement is written to prevent. Readable, editable, and not mine.