For Claude Code, Codex, Antigravity and Kimi

Same model.
Better mind.

Onyx wraps the coding agent you already use in a harness that remembers what worked, checks every action and seals every step. Same model, same settings, head to head: it has beaten plain Claude Code 112 games to 40.

macOS, Windows and Linux. Uses your own agent subscriptions.

Onyx vs plain Claude Code Stockfish scored
Onyx112
:
Claude Code40
13 series of ten or more games, 165 games, 13 drawn Onyx draw Claude Code
20–0our best series. Every game a win, every game more accurate
27×less context carried into new work (18k vs 479k tokens)
−77%cost per game in one week of harness work
29%cheaper all in, checkpoints included, than the same model in plain Claude Code

The product

Memory. Governance. Provenance.

An agent on its own keeps one growing session, forgets it when the session ends, and leaves no record you can check. Onyx is a desktop app that fixes all three.

Memory

Remembers what matters. Knows what it doesn't know.

  • Checkpoints keep every new task lean: 18k tokens, not 479k.
  • Topics carry across checkpoints, so a thread never loses the plot.
  • Lessons start as hypotheses and get disputed when evidence disagrees. One lucky case never becomes truth.
Governance

Every action checked before it runs.

  • Per-project access: its folder, ask first, or the full machine.
  • A policy gate on every tool call, with refusals recorded too.
  • Shield redacts sensitive output before the model sees it.
Provenance

Every step sealed. Every line traceable.

  • Each action is a sealed object of 8 hashed sections. Change one byte and it shows.
  • All of it lands in a hash-linked chain you can verify.
  • Ask any line of code which turn wrote it.
CodeFlow semantic maps of your code, kept fresh as you work Action cues that react to what the agent does Skills on demand instead of on every request Bring your own Claude Code · Codex · Antigravity · Kimi

Benchmarks

Proven head to head. Next, everything else.

Same model and effort on both sides, scored by something neither side controls, every run published. Chess comes first: long games played from memory, rated move by move by Stockfish. Onyx has won 12 of the 13 series.

Onyx stays light. Claude Code fills up.

Tokens carried into each game of the latest series.

OnyxClaude Code

So cost per move stays flat.

List cost per move, game by game. Claude Code's rose fivefold as its session grew.

Onyx, avg 5.7¢Claude Code, avg 9.4¢
Running

Chess arena

112 wins, 40 losses, 13 draws across 13 series.

Next

Coding arena

Fixed coding tasks repeated over rounds, graded by hidden tests.

Planned

Every agent

The same arenas through Codex, to show the gains carry across models.

Planned

Every platform

Windows, Linux and the packaged app, alongside macOS.

Costs use one list-price table for both sides (Opus 5.5, 1-hour cache writes). In the latest series Onyx cost $2.43 a game to Claude Code's $4.03 on play; its checkpoints add $0.74, leaving it 21% cheaper all in, and 29% across the series' three Onyx arms. Onyx is cheaper per move once a session passes about 160k tokens. Still improving: output length and the checkpoint model.

Give your agent a better mind.

Early access for macOS, Windows and Linux.