GigaCode

You will pick this for integration rather than intelligence. It is built into the platform where your code already sits, it is free for individuals, and it ships an agent mode alongside the completion, so the daily experience is that everything is already connected. That proximity is the deciding trait, because the assistant you actually use is the one that is already there when you open the editor.

Pick it if you are on this platform. Pick SourceCraft Code Assistant when you want a comparable regional tool without committing to this host.

Filtering on both sides of the exchange is a design that fails silently. A prompt rejected before it reaches a model and a response altered after it returns produce the same symptom, which is an assistant that is unhelpful for reasons it will not state, and there is no key of your own to route around the filter because none is offered. Debugging a refusal becomes guesswork about somebody else's policy.

What it does right: the terminal companion works against a local repository, so the working tree stays where it is rather than being uploaded to a workspace.

1. Retrieval over the codebase supplies context to the chat and the agent, described but never specified, so the selection and ranking mechanism is unknown. 2. The routing layer is the substantive claim: dozens of open models behind one interface means two identical requests can be served by different systems, and no documentation says which, when or why.

Nothing is pinned and no benchmark is published, so no result here is reproducible and no regression is attributable to a change. The observation: a routing layer this wide is an operational achievement and an evaluation problem, and only the first half is marketed.

The parent is the whole thesis. A financial institution funding a developer platform is not chasing software margins, it is buying position in a domestic ecosystem, which means this product survives commercial underperformance for years in a way a startup's would not. Owning models rather than renting them keeps the cost of goods internal, and pricing at that level for business is deliberately below what an international competitor could match.

There is no acquirer and no exit, because none is wanted. Position: durable at home, structurally unavailable elsewhere, and priced to stay that way.

One sentence on the demo: the agent mode did what it said. Everything else is a procurement obstacle. The only number I can find for the edition security would accept is quoted in a currency our purchasing system does not carry, from a vendor our compliance function will not clear, and no identity integration or retention policy is published in a language my team reads. Nothing here reaches our pipelines either.

Onboarding is an hour for the plugin and a project for the on-premises edition. Not yet.

The one thing I would want is technically present and practically out of reach. Models can run inside your own perimeter, and getting there requires an organisational licence rather than a download, so an individual has no path to it at all. That is local inference as a sales tier rather than a capability.

Everything else is shut. The source is closed, I cannot supply my own credentials, and there is no client for the tool servers I already run, so nothing I have built attaches. What I am offered is a well-integrated product on a platform that owns every layer beneath it, and I own none of them.