SRD CodeFree

The deciding trait is the plan step. Craft and Ask are the two core agents, and a plan mode sits in front of them so the approach is agreed before anything is written, which is the difference between a session you review and a session you undo. Once a tool makes planning the default rather than a discipline, the failure rate on larger tasks changes noticeably.

The catch is that all of it assumes you are inside one company's platform. Pick it if you are. Pick something self-contained if you are not.

The extension mechanism is the exposure. A hook is arbitrary code executed at a lifecycle point, and an organisation-scoped server catalogue means the set of tools reaching an engineer's agent is curated centrally rather than chosen locally. That is a supply chain inside the company, and nothing published describes review, signing or provenance for what enters it.

What it does right is separate the approval modes. File editing and shell execution are gated independently with an explicit whitelist for what may run without asking, which is the correct granularity.

1. Taint tracking is a named method with a literature, a false-positive profile and results a second tool can reproduce, which is a different epistemic class from asking a model whether code looks unsafe. Naming the technique rather than the outcome is the part worth crediting.

2. Grounding search in a language server means symbol resolution comes from a parser rather than from text similarity. 3. No accuracy figures accompany either, so the methods are stated and their performance on real repositories is not.

Bundled tooling has a different survival curve from standalone tooling. It lives as long as the platform it defends, which for infrastructure attached to a national operator is a long time, and it is priced at whatever makes the platform stickier rather than at what the market would bear. That is why the pricing question has no answer visible from outside.

Moat: the platform underneath, which is the strongest kind and also the least portable. Likely acquirer: none. Position: durable if you are already a customer, unreachable if you are not.

The install is free and the relationship is not defined, which is the wrong way round for anything that touches source code. I cannot review terms I cannot obtain, and without them there is no retention clause, no processing location and no support commitment to hand my security reviewers, whatever the product itself does well.

Nothing runs unattended either, so it never becomes a pipeline step I can measure. Not yet, for a team outside its home market. Inside it, with the account already in place, this is one of the better-governed rows on the board.

This is the most closed row I have read this quarter. The model list is fixed, my own credentials are not accepted anywhere, and nothing about the loop can be pointed at an endpoint I run, which means every turn happens on terms I did not set and cannot inspect.

The marketplace is the exception and it is a real one: servers are published, installed and counted, so a tool I write can reach every engineer without going through the vendor. That is a genuine extension point bolted to an otherwise sealed product, and I resent how much I like it.