bbarit-oss

The deciding trait is that it is one compiled binary. No runtime to install, no version manager, no dependency tree that breaks on a machine you set up last year. Rust means startup is not something you notice, which sounds trivial until you have waited on an interpreter forty times in an afternoon.

What you give up is documentation and a stable surface, because this was pulled out of the vendor's desktop product and still reads like an internal tool. Pick it if you like fast and small. Pick Claude Code if you want the version other people have already debugged.

The risk is concurrent writers. Several processes editing one checkout at once produces the class of corruption that is hardest to unpick: two partial edits to the same file, a git index touched mid-operation, and a final state no single agent intended. The row records git operations and parallel sub-agents together, and no arbitration between them.

What it does right is refuse to hide the loop. This is a small legible control flow inherited on purpose from a project known for exactly that, and a legible loop is one you can reason about after it goes wrong.

1. Storing long-term facts in a categorised, human-readable index is the unusual choice here. Most designs persist memory as opaque vectors, where a wrong belief is undetectable until it surfaces; an inspectable index makes correction a text edit. 2. That taxonomy is borrowed from a prior project rather than invented, which the row states, and borrowing a tested scheme is the right instinct.

2. Retrieval over the repository is semantic and bundled rather than delegated to an external service. 4. No evaluation accompanies any of it, and none is claimed.

The interesting fact is the extraction. A company with a paid desktop tool carving out its engine and publishing it is buying distribution with code it already wrote, and the calculation only works while the desktop product remains the thing that earns. Free is a customer-acquisition line item here, not a position.

Moat: none in the giveaway; whatever moat exists sits in the product this advertises. Likely path: it tracks the commercial tool closely and slows the moment that tool stops selling. Position: use it, and understand you are standing in someone's marketing funnel.

Finance approves this in a sentence and security stops it in the next. Sixty local installations with sixty developer-held provider keys is sixty places our source is read and no central record of any of it. There is no administrative surface, so model policy and version drift are handled by whatever desktop tooling I already own.

It does not run unattended, so it never becomes a measurable step in code review or delivery. Onboarding is fast for anyone comfortable in a terminal. Not yet, until keys are issued centrally and there is somewhere to read what it did.

Inheriting my existing configuration is the detail that wins me over. My servers and skills are already defined for two other tools, and this one reads them rather than inventing a third file format for the same information. That is respect for the person doing the setup, and almost nobody ships it.

The licence keeps a fork viable and the provider registry is genuinely agnostic. The gap is local weights: no local model support, so my own hardware sits idle and every request goes to somebody's API. That is the one thing I would patch first.