OpenABCode

The deciding trait is that you stop choosing. Each task is classified and sent to whichever provider suits it, so the mental overhead of deciding which model handles this particular job disappears from your day. If you have ever kept three terminals open for three vendors, that is the friction being removed.

You are the wrong buyer if you have one provider and one bill, because then routing is complication without benefit, and you can switch it off but you paid for it in setup. Pick it when you are already multi-vendor. Pick a single-model agent when you are not.

The safety model is a reading assignment. There is no permission layer in the product, and the documentation responds by listing three external isolation approaches the user is expected to arrange. That is honest and it is also a default that fails open: the person who installs this quickly gets an agent with a shell and no boundary.

What it does right is admit it in writing. A project that documents the gap is easier to deploy safely than one that implies a sandbox it does not have.

1. The classifier is a model, which means each task pays for an extra inference before any work begins, and the routing decision has an error rate nobody has published. 2. That would be a fatal omission except for the second design choice: every decision is written to the session record, so a user can measure the error rate themselves.

2. Making a system auditable when you cannot make it correct is the right order of priorities. The apparatus for evaluation exists; no evaluation using it has been published.

358 stars, and a value proposition that only exists while three vendors differ in price and strength. Routing is arbitrage, and arbitrage is the least defensible business in software, because the counterparties can close the gap whenever it becomes worth their while.

Moat: none, and the sign-in path suggests an account layer whose purpose has not been explained. Likely acquirer: none. Likely path: the routing idea becomes a feature inside larger agents and this stays the reference implementation. Position: use it and keep the configuration portable.

The design multiplies my procurement work by three. Each provider it routes to is a separate agreement, a separate data-handling review and a separate invoice, and sixty engineers spread across all of them makes the monthly number harder to forecast rather than easier.

There is no SSO, no audit log and no retention policy from this component, and it does not run unattended, so it never becomes a measured step in delivery. Onboarding is short. Not yet, and the blocker is that I would be approving three vendors to save developers one decision.

Every part of the decision is mine to rewrite. Rules, keyword heuristics, file-extension mappings, project markers and the fallback provider are configuration, per project or globally, and one command turns the routing off entirely when I disagree with all of it. Configurable defaults with an off switch is the shape I want.

AGPL-3.0 is the strongest copyleft on this board, which suits me and will trouble anyone planning to embed it. The npm install even passes a flag to skip install scripts, which tells me somebody here thinks about supply chains.