antiloki · a desktop app · macOS · Windows · Linux

Faster with agents. Safer with proof. Local, yours.

Work faster — parallel sub-agents, one job per thread, the tools beside it. Work safer — every change on an audit trail, every file judged by deterministic analysis, no AI in the verdict. On your machine, on the subscription you already pay for.

Download — free for 7 days
one binary · no account · no API key — your subscription is the engine
Threads

Speed — agents that work in parallel, each in its own box.

Threads: two agent panes side by side, the scope sentry refusing an out-of-scope edit, a live preview of the worktree in the same thread
  1. 01

    One job, one thread.

    The chat, the agents and the tools that job needs, side by side — drag, resize, collapse. Leave for X-ray and the agents keep working; a floating stack shows each one's line.

  2. 02

    Boxed, not trusted.

    Every agent gets a worktree and a scope. A write anywhere else is refused by the tool itself, before the file changes — and the scope sentry shows the attempt, with one click to widen or keep.

  3. 03

    Heimdall routes every ask.

    A high-tier planner reads the ask first: one job or a plan, which model, which tier — the cheapest that can do it. Sub-agents fan out in parallel, each in its own box, on the subscription you already pay.

  4. 04

    Planned, verified, judged.

    A goal splits into tasks that own disjoint files; each is verified against its acceptance; an isolated QA judge reviews the whole run; one pull request at the end — sealed if the judge failed it.

  5. 05

    Preview beside it. MIN when you want it cheap.

    ▶ preview boots the agent's worktree next to the chat, live. MIN mode turns exploration off — the agent navigates by antiloki's own intel (reviews, graph, docs, ASTs) and touches the disk only to edit.

X-ray

Safety — every file judged, every number explained, before anything ships.

X-ray: a file's review — its score, why not 100, the law-based and measured health bars, the review over the snapshots
  1. 01

    The verdict, not a vibe.

    Every file reviewed and scored — and why not 100, axis by axis: coupling, structure, cohesion, defects, tests, change. Deterministic analysis; no AI in the number.

  2. 02

    Laws, not lint.

    The rules your codebase lives by, checked at the line: what breaks a law, where, and what compliance is worth. A senior review that reads the same way every time.

  3. 03

    The codebase as a map.

    Every file a box, area its mass, colour its health; hubs, orphans, cycles, the most central and the most risky — at a glance, and per folder.

  4. 04

    What the tests never reach. What a change would touch.

    Coverage gaps by file, the blast radius before a change lands, and the review over the snapshots — the score's history, scrubbable.

  5. 05

    The same intel feeds the agents.

    Reviews, graph, docs and ASTs become the context pack every agent starts from — what MIN mode navigates by. Design too: your screens captured from the running app, the design system and its drift.

the real app · the same codebase on both screens · Threads is where the agents work, X-ray is what they know

how teams work today

Adoption is done. Trust and control never arrived.

90 %of professional developers use coding agents weekly; 68 % daily.JetBrains, 2026
29 %trust what the agent produces — down from 40 %. 3 % "highly trust" it. 7 in 10 won't merge without a manual review.Stack Overflow, 2026
+200 %code output per engineer in a year. Review throughput: flat. Review — not generation — is the bottleneck.Anthropic, internal
9 sfor an agent to delete a production database and its backups — after quoting the rule that forbade it. 93 % of orgs have had an AI-caused incident; 19 % have governance for the next one.PocketOS · Spacelift, 2026
Rules in a prompt are not enforcement.

"Don't touch production" lived in the agent's instructions. It read them, and ran the delete anyway. What an agent may do has to be enforced by the tool, not requested of the model.

More agents means more to supervise, not less.

At five or six agents on one repo the questions become: which one is blocked, which one changed the wrong file, which branch is safe to merge. The operator becomes the bottleneck.

Nobody can say what it changed or what it cost.

Review finds it late, production finds it later, the invoice finds it at month end. Trust stays at 29 % because there is no record to check.

Free for a week. Then $200 a year, or $500 once.

Trial
$07 days
  • everything — Threads and X-ray
  • no card, no account
  • your projects stay when it ends
Download
Yearly
$200/ year
  • updates + support for the year
  • one machine at a time, move it anytime
  • cancel anytime — runs to the end of the year
Buy a year
pay once
Lifetime
$500once
  • yours forever — never expires
  • 2 years of updates + support
  • one machine at a time, move it anytime
Buy lifetime

USD · tax at checkout · key by email, pasted once · works offline up to 14 days between checks · invoices with VAT for companies

Install. Open a repo. Start an agent.

macOSApple silicon · Intel · .dmgDownload ↓ Windows.exe · .msi · x64Download ↓ Linux.deb · .AppImage · x64Download ↓

needs git and at least one of claude · codex · opencode signed in

What is antiloki, in one sentence?

A desktop app that runs your coding agents in parallel on your machine with folder-level write enforcement, per-task verification and a full audit timeline — the control layer the agents don't ship with.

Does it replace Claude Code, Codex or OpenCode?

No — it runs them, twice over: as the agents in their boxes, and as the engine behind antiloki's own AI (the chat, the reviews, the ultra review). Keep your subscriptions; there is no API key to paste.

What does the X-ray do that a linter doesn't?

It judges every file against your rules and its own measurements (tests, coupling, structure, change) and tells you why a file is not 100. Then the ultra review: a $0 survey of the territory, and your engine reading the worst units to report what it can verify — with the fix. The chat sees all of it, so "why is this file 81?" gets a real answer.

Isn't this just worktrees and tmux?

Worktrees are the easy part. The write gate at the tool call, the verifier before merge and the cost-per-task record are what you can't script in an afternoon — and what the trial is for.

2× the tokens sounds expensive.

It is more. That's why the bill is on screen and the agent decides when splitting is worth it — in our run it declined a small task on its own. You pay for verification, not speed.

Can my company install it?

One binary, local, no account, no telemetry. The complete list of network calls is: your CLIs' own calls, a daily license check, an update check against a static file. Send that to security.