Build +112% more on your weekly limit, 10× cheaper, with the same code quality.

Merging is one part of it.

Your Claude, Codex, Cursor, Gemini and Grok subscriptions, on one repo, as one agent. antiloki routes each job of an ask to the seat measured best for it, automatically, under one rule: keep at least 80 % of the best measured quality, then the cheapest, then the fastest. The frontier seat is asked only for the parts that need it, and a seat that is out of quota — or nearly — is skipped before the job, not after it fails. It controls what may be touched, records every turn — priced, timed and attributed to a seat — tests every run before it counts, and measures the whole thing in seconds and dollars. The merge is what routing produces — the rest is why it can be trusted.

Adoption is finished and control never arrived: 90 % of professional developers use coding agents weekly and 68 % daily, but only 29 % trust what comes out — down from 40 %, with 3 % who highly trust it. And the failure mode is not theoretical: an agent quoted the rule that forbade it and deleted a production database and its backups, and the deletion call took nine seconds — 93 % of organisations have had an AI-caused incident, 19 % have governance for the next one. Rules in a prompt are not enforcement. JetBrains 2026 · Stack Overflow 2026 · PocketOS · Spacelift 2026

One agent per role, on the seat measured best for the job.

You do not staff a roster. A setup wizard finds the AIs on your machine — Claude, Codex, Cursor, Gemini and Grok, each listed as ready, not signed in or not installed, with the command that fixes it — and asks for one main provider. Then you ask, and the agent takes the role the ask needs (Backender, Frontend designer, Tester, Reviewer and others), arriving with its scope, skills and specialists. The router chooses the seat for every job, from the seats you are signed in to and that still have room in their window; there is no per-job picker to set.

Each agent keeps a queue of asks you can reorder, cancel and re-prompt, with a checkpoint per response to undo to. A run can be paused, resumed or stopped, and the QA can be stopped on its own. The QA test is filmed — a live browser frame while it runs and the kept recording afterwards — and every turn closes with what it did, which seat did it, what it cost, and the QA verdict. A long conversation closes into episodes that consolidate into a summary.

The codebase judged by arithmetic, before anything ships.

X-ray is the deterministic half. Every file gets a verdict — a score and why it is not 100, axis by axis — computed from your ASTs and git history, so the same commit scores the same twice and no model is asked for the number. Around it: file reviews with comments at the lines, the rules your codebase lives by checked at the line, a map of the codebase (every file a box; hubs, orphans, cycles), the coverage gaps, and the blast radius of a change before it lands.

It reads TypeScript and, since 2026-10-01, Flow-typed JavaScript: on a copy of React the Flow files it had to skip went from 1,160 to none, and 11 of 1,177 are still unreadable. It says what it did not index — excluded, an unsupported language, a parse that failed, changed since the index. The same facts reach the agent as its context pack. Measured on a wide TypeScript rename (three agents, one run per cell), the map alone was 31 % faster on 48 % fewer input tokens at the same pass rate. On a small change the map alone passed 0 of 3, because the index had left out the file that needed the change; that is why the pack now says what it could not read. On a second repository the map made Codex use more tokens (472k to 1.15M). The report has the cells.

What sits around the work.