All AI guides
Comparison13 min read

Codag vs Crukx: Crukx Changed Jobs Again, So Pick by Job

As of 7 November 2026, Codag is a live tool that cuts coding-agent tool output at $0.05 per credit, with 10 free credits. Crukx has left agent testing and now sells lossless delta storage for fine-tuned models, measured at 2.44x to 3.65x smaller on fine-tune pairs, with a free alpha and no listed price.

The short version

Codag and Crukx no longer compete. Codag trims coding-agent tool output at $0.05 a credit, with 10 free credits and a $499 Team plan. Crukx now stores fine-tuned model checkpoints as exact deltas, free in alpha, with no published price. Pick by the problem you have, not the brand.

Not sure which one fits you? Describe your situation and Smart Match names three tools.Find my AI tools

In this guide15 sections

Two unrelated jobs: Codag trims coding-agent logs and test output before the agent reads it at 0.05 dollars a credit, and Crukx stores a fine-tuned model as a CRX delta 3.61x smaller Hokai diagram of the two products' jobs; the 3.61x figure is Crukx's own Qwen2.5-14B benchmark. Source: codag.ai pricing and crukx.dev, checked 7 November 2026.

Codag and Crukx no longer compete for the same budget, because Crukx changed its product for the second time in 2026.

In August this guide told readers that only one of the two shipped, since Crukx sold a regression gate behind a waitlist. That framing is gone. On 7 November, crukx.dev opens on "Know what every model is. Move only what changed," and the product is a lossless storage format for fine-tuned model checkpoints. Codag is still a metered tool that trims what coding agents read. A team choosing between them is really choosing between two unrelated problems.

Key numbers
Qwen2.5-14B fine-tune, full model against a CRX delta
3.61x
Crukx benchmark, 7 Nov 2026
One Codag credit, about 100 MB of log output
$0.05
Codag pricing, 7 Nov 2026
Free Codag credits, about 1 GB, no card
10
Codag pricing, 7 Nov 2026
Crukx on a model trained apart from its base
1.52x
Crukx benchmark, 7 Nov 2026

What changed since 24 August

This guide was last judged on 24 August 2026. Re-read against both vendors' own pages on 7 November, the picture moved in four places.

Item24 August7 November
Crukx productAgent release gateCheckpoint delta storage
Crukx accessWaitlist onlyFree alpha, scoped pilot
Codag framingLog compressionTool-output gateway
Archal paid useDescribed as billingCredits only for now

Codag's prices did not move. Its pitch did: the Y Combinator profile now calls it "the AI gateway for cheaper, faster, permissioned agents," which widens the story from log trimming to cost control and tool governance across providers. The rest of this guide is rebuilt around the two jobs the products now do, and it is the second time this pair has needed that treatment. The reasonblocks-vs-touchmark comparison hit the same problem with two other pivoting startups.

Why both pitches moved

Both companies changed what they sell within about six months, which is worth knowing before you trust either roadmap. Codag's blog keeps its launch post of 23 May 2026, "Hello Codag," as history and notes that the product has since grown from log compression into cutting coding-agent costs across all large tool results. That is a widening of one product, and the original Codag buyer is still served by it.

Crukx is a different kind of change. The August site sold a gate that blocked risky agent releases, and the November site sells storage for models. On the pages read, nothing on the new site mentions the old pitch, and we found no announcement explaining the move. A reader who bookmarked Crukx as an agent tool in the summer would now land on a page about checkpoints, which is a fair reason to distrust any listing that still describes it the old way.

Pivots are common in a young batch, and they are not a verdict on quality. They do change how much weight a buyer should put on a roadmap slide, so compare the product that exists today. For the storage side, the models directory shows which base models your own fine-tunes descend from, and the guides index holds other comparisons where one side moved.

What Crukx actually is now

Crukx stores a fine-tune as the difference from its parent model, then rebuilds the original files byte for byte. Its homepage example is Qwen2.5-14B to its Instruct version: 29.5 GB as a full model, 8.2 GB as a CRX delta, a 3.61x reduction with an identical SHA-256 on rebuild. Packing took 2 minutes 14 seconds and unpacking 2 minutes 1 second on 32 CPU threads.

The benchmarks page lists every pair Crukx measured, including the ones where it does not help. The numbers are specific enough to check against your own models:

PairReductionNote
Qwen2.5-1.5B to Instruct3.65xFine-tune
Qwen2.5-Coder-7B to Instruct2.44xFine-tune
Qwen2.5-14B to Instruct3.61xFine-tune
Qwen2.5-0.5B-Instruct to merged LoRA5.35xMerged adapter
Qwen2.5-1.5B to Math-1.5B1.52xUnrelated sibling

The last row is the honest one. Math-1.5B was trained separately, so the base gives no help and the delta matches what Crukx alone achieves. The page also says untouched modules in a merged LoRA stay byte-identical, which is why a plain deduplication tool such as Xet already finds 27.6% overlap there, and that keeping the unmerged adapter is smaller still.

Against the tools people already use, the same page shows the 1.5B pair stored at 3.09 GB raw, 2.69 GB with Xet deduplication against the base, 2.05 GB with ZipNN and 0.85 GB as a CRX delta. Crukx's pricing page adds that Xet found 0 to 0.03% overlap on full fine-tunes.

Both sets of figures are Crukx's own measurements, so treat them as a vendor benchmark rather than an audit, and rerun them on your checkpoints before committing storage plans. Hugging Face, where most public fine-tunes live, is the obvious place to compare.

Crukx is also blunt about the limits:

  • A delta needs its parent to rebuild, so an unrelated model gets no delta benefit.
  • CRX "does not make inference or training faster" today, and Crukx claims no GPU savings it has not measured.
  • Packing and unpacking run on an ordinary CPU, so no GPU is needed.
  • The CLI that pushes your own models is a private alpha, while the model explorer and benchmarks need no account.

What Codag still is

Codag attaches to Claude Code and Codex with one codag setup command, then routes model traffic through a local proxy on the developer's machine. It reduces only large, eligible tool results: logs, test and build output, search results, file listings and document reads. Source code, diffs, configuration and exact structured data pass through unchanged.

Anything it omits stays encrypted on the developer's machine for seven days, capped at 1 GiB, retrievable by line range, JSON path or group. If Codag times out, fails validation or runs out of quota, the agent gets the original bytes. That fail-open design matters more than any savings figure, because a compression layer that can break a session is a layer people switch off.

Its numbers come in two flavors. The founder's launch post says optimized results are 75% smaller on average, and that more than 90 organizations use it, including teams of up to 35 people.

The Y Combinator summary is more modest, saying it cuts token spend 20 to 30% while retaining output accuracy. Both are the vendor's own claims, and the gap between a 75% smaller tool result and a 20 to 30% smaller bill is normal, because tool output is only part of what a session pays for. The 38.4 million tokens and $291 on the homepage dashboard are a labelled example, not a customer result.

What each one costs

Codag is metered in credits at $0.05 each, which it estimates covers about 100 MB of log output. The catch is in a line a skim misses: its inference-based agent-context reduction burns credits about a thousand times faster, around 100 KB per credit, or roughly $500 per gigabyte.

PlanPriceWhat you get
Pay as you go$0, then $0.05 a credit10 free credits, about 1 GB
Credit packs$0.50 to $25About 1 GB, 10 GB or 50 GB
Team$499 a month12,500 credits, about 1.25 TB
EnterpriseCustomSSO, VPC peering, version pinning

Team overage bills at the standard $0.50 per GB, and a hit quota means passthrough of the original output, never a broken session. A 14-day pilot needs no card and never auto-charges.

Crukx publishes no amount. Its pricing page lists three doors:

  • Explore: free, no account, with the benchmarks and a model explorer.
  • Build: a private alpha of the CLI (crukx scan, push and pull), free while in alpha.
  • Pilot: an engineering engagement scoped together, with a written report of bytes, time and rebuild checks.

The page's storage estimator says 20 full-weight fine-tunes of a 14B base would take 191 to 271 GB instead of 620 GB. That is a storage-only estimate, with transfer time and egress left out.

Where Codag wins

Codag wins whenever the cost is a coding agent rereading its own output. It is live today, billed in cents, installable in minutes and reversible with codag disable. A solo developer on Claude Code or Codex can try it without a card (users of Cursor should check support first, since it is not on Codag's attach list), and nothing in the setup changes the model, the prompts or the provider account.

It also has more public surface. Beyond the product, Codag publishes open-source SDKs for Python, TypeScript and Go, a Rust log-templating library under the MIT license, and a VS Code visualizer with more than 500 stars. A buyer can read code and test claims, which counts for more than any dashboard screenshot.

Where Crukx wins

Crukx wins on a job Codag does not attempt. A lab, a hosting team or a company that keeps dozens of fine-tuned variants of the same base model stores most of the same bytes over and over. Crukx's measured 2.44x to 3.65x on fine-tune pairs, and 5.35x on a merged LoRA, is a direct storage saving, and an exact rebuild means no quality question: the files are identical.

The private alpha is also free right now, which makes the downside of testing small. If your storage bill is dominated by checkpoints rather than by agent tokens, a pilot is the only way to find out, and Crukx says it will tell you where CRX did not help.

Who should pick which

Pick Codag if developers run Claude Code or Codex all day and the pain is cost or context filling up with passing test output. Pick Crukx if the pain is terabytes of fine-tunes that share a parent. A team with both problems can use both, because they touch different layers: one sits in the agent's tool path, the other in the model store.

If you arrived here looking for what Crukx sold in August, a release gate that blocks a bad agent deployment, neither tool does that now. The closest live option is Archal, covered below. A team that wants to compare it with a rival can read the Archal vs ReasonBlocks breakdown.

A one-week test for each

Neither tool should be bought on a vendor chart. A week is enough to measure both on your own work.

For Codag:

  1. Run codag setup on one developer machine and note the agent's usual token use for a typical task first.
  2. Work normally for three days, then read codag status and codag savings for the counts.
  3. Spend the 10 free credits only on logs, so the $0.05 rate applies, and avoid the inference reducer.
  4. Compare the bill for that machine with the week before, and ask the developer whether any answer got worse.

For Crukx:

  1. Pick one base model and the fine-tunes you keep from it, and total their size today.
  2. Use the free Explore tier to check whether your parent appears in the model explorer.
  3. If the pairs fit, join the alpha and run crukx scan before you push anything.
  4. Rebuild one delta, compare the checksum with the original and keep the full copy until it matches.

Switching cost and lock-in

Neither product traps you. Codag attaches through a local service and codag disable rolls it back, so leaving costs a minute. Crukx is the stickier one by design: a delta can only be rebuilt from its parent using Crukx's tooling, so storing your only copy of a model as a delta makes the CLI part of your recovery path. Keep full copies of anything you cannot afford to lose until a pilot proves the rebuild on your own files.

Data handling and security

Codag says eligible tool output is processed transiently and not retained, that billing runs on contentless counts and sizes, and that omitted content stays on the developer's machine. That is a meaningful claim for teams worried about source code leaving a laptop, though Codag's own page states it, so verify it with your security team before enabling it on private repositories.

Crukx says weights stay on your infrastructure during a pilot and that the alpha CLI runs on your machine. Its homepage shows a sign-in and a dashboard, but we found no SOC 2 or compliance page on the pages read, so a regulated team should ask before pushing any private model.

The live alternatives for the job Crukx left

Crukx's August pitch, an agent regression gate, still has buyers, and a few live products cover parts of it. Archal sells stateful sandboxes of services such as GitHub, Linear, Slack, Supabase and Stripe, exposed as REST and MCP endpoints, so a test run never touches a real account.

Its billing details:

  • Each environment costs $0.10 per minute, billed by the second.
  • New accounts get $20 in credits: $5 on verified signup and $15 once the first sandbox is ready.
  • Archal states that paid continuation is not yet available while its safety rollout stays disabled.

So today Archal is a free-credit trial, not a purchase. Kashikoi runs no-code agent simulations and is reachable only by booking a demo. Maitai covers production monitoring and correction, and when we read it on 24 August it listed a $50 Starter and a $200 Professional plan. Its pricing page showed no plan figures when we opened it on 7 November, so confirm the current price directly.

Other evaluation and monitoring tools on Hokai include Confident AI, Openlayer and ReasonBlocks, none of which we re-read for this refresh.

For the spend half of Crukx's old observability pitch, SuperPenguin publishes four plans:

PlanPriceManaged spend
Free$0Up to $2K
Growth$30 a monthUp to $5K
Pro$200 a monthUp to $20K
EnterpriseCustomCustom

Teams comparing broader options can start from the AI agent observability directory, or the three jobs guide, which separates monitoring, testing and spend control.

The challengers for Codag's job

Codag is not alone in trimming tool output. Edgee sells a drop-in CLI wrapper that adds three techniques: tool result trimming, tool surface reduction and output brevity. It publishes two figures and tells readers not to add them together: a 15 to 20% cut in token spend across customers in production, and 50% on SWE-bench Lite with all three techniques on. It has no Hokai record yet, and a vendor benchmark on one suite is not an audit.

A search also surfaced IterateAI's compression plugin, sqz and anvanster's compressor, all open-source hooks aimed at Claude Code. We did not read those first-hand, so this guide quotes no figures for them and does not rank them. For a team that would rather route traffic than wrap a CLI, LiteLLM sits at the gateway layer, though it does not trim tool results. Browse the wider tools directory for other agent-cost options.

Sponsor spot open

What would change this verdict

Three developments would reopen this comparison. If Crukx publishes a price or a general-availability date, the cost section needs a rewrite, and if it ships a third product, so does the rest.

If Codag's inference reduction becomes the default instead of the opt-in rate, a credit stops covering 100 MB and the headline price stops predicting a bill. And if a model vendor such as Anthropic or OpenAI builds this trimming into its own agent, a separate tool for it loses much of its reason to exist.

Both companies are small. Codag is a one-person team by its Y Combinator profile, and Crukx has changed direction twice this year, so give either a short leash before standardizing on it.

The case for waiting

The sensible objection is that a product that has pivoted twice is not worth evaluating. For Crukx that is a fair read: the alpha is free, the pricing is unpublished and the page you read today may not describe the product next quarter. Waiting costs nothing but the storage you would have saved.

For Codag, waiting costs more, since the 10 free credits are enough to measure the effect on one real session. If you want a shortlist matched to your own agents, budget and setup, Smart Match takes a few questions. Developers on GitHub Copilot rather than Claude Code should note that Codag lists only Claude Code and Codex today, so check support before assuming it fits. Both vendors' own pricing pages, cited at the foot of this guide, outrank any third-party listing, including Codag's and Crukx's entries elsewhere.

Frequently asked questions

Are Codag and Crukx competitors?

Not any more. Codag reduces the tool output a coding agent reads, and Crukx stores fine-tuned model checkpoints as exact deltas from a parent model. In August 2026 Crukx sold an agent release gate, and its homepage on 7 November describes the storage product instead, so they now solve unrelated problems.

How much does Codag cost in 2026?

Codag bills in credits at $0.05 each, about $0.50 per gigabyte of log output, with 10 free credits and no card to start. Team is $499 a month for 12,500 credits, and Enterprise is custom. Its inference-based context reduction burns credits about 1,000 times faster, near $500 per gigabyte.

Does Crukx publish a price?

No. Its pricing page lists a free Explore tier, a Build tier that is a free private alpha of the CLI, and a Pilot that is an engineering engagement scoped with the customer. No amount is shown, and the alpha is free only while it is in alpha.

Does Crukx shrink every fine-tune by the same amount?

No. Its own benchmarks range from 2.44x on Qwen2.5-Coder-7B to 3.65x on Qwen2.5-1.5B, and 5.35x on a merged LoRA. A model trained separately from the base, Math-1.5B, got only 1.52x, because the base gives no help. Rebuilds are byte-identical in every case it lists.

What replaced the release gate Crukx used to sell?

Archal is the closest live option, with stateful sandboxes of services like GitHub and Slack at $0.10 per environment per minute. It currently offers $20 in free credits and says paid continuation is not yet available. Kashikoi runs no-code agent simulations through a booked demo.

Covered in this guide

  • Codag: Cuts AI coding-agent costs by trimming oversized tool results, from test runs to grep output, down to the lines that matter, with the rest kept on your machine.
  • Crukx: Stores related model checkpoints as exact deltas from their parent model, so fine-tunes take less space to keep and move. Free Explorer, CLI in private alpha.
  • Anthropic: Anthropic is a Public Benefit Corporation started by 7 ex-OpenAI researchers. It builds the Claude model family and counts Amazon and Google among its largest backers.
  • Archal: Stateful API sandboxes for testing AI agents and integrations: 22 environments such as GitHub, Slack and Stripe, billed per environment-minute. YC S26.
  • Confident AI: LLM evaluation and tracing platform from the DeepEval team, with a free plan, CI regression tests and SOC 2 Type II.
  • Cursor: Cursor is an AI code editor built on VS Code, widely adopted across large enterprise engineering teams, with Agent Mode, Tab completion, and Cloud Agents.
  • GitHub Copilot: GitHub's AI coding assistant with agent mode, Copilot Fusion model routing and MCP support across 6 IDEs, plus a free plan with no credit card required.
  • Hugging Face: The AI community building the future. Platform for discovering, sharing and collaborating on machine learning models, datasets and applications.
  • Kashikoi: Kashikoi is a no-code simulation platform for evaluating AI agents, built by the ex-Moveworks lead behind 250+ shipped enterprise agents.
  • LiteLLM: Open-source AI gateway and Python SDK that puts 100+ LLM providers behind one OpenAI-compatible API, with budgets and fallbacks. Free to self-host.
  • Maitai: Maitai is a $50 to $200/mo control plane that indexes production LLM traffic and auto-corrects bad output in real time with Sentinels.
  • OpenAI: OpenAI builds ChatGPT, the GPT-6 family (Astra, Sol, Terra, Luna), Codex and the OpenAI API. It reports 1.2B weekly ChatGPT users and raised new funding in March 2026.
  • Openlayer: AI governance platform with an AI system inventory, 175+ evaluation tests and runtime guardrails that produce audit-ready evidence for regulators.
  • ReasonBlocks: YC-backed startup that audits a team's AI evals and tests judge scores against people, with candidate model training on the side. No public pricing.
  • SuperPenguin: AI spend platform with 2 native SDKs (Python, TypeScript) that attributes provider costs to a customer, feature, or pull request instead of one monthly total.

Sources

Still deciding?

This guide covers a handful of options. Smart Match checks every listing in the directory against how you actually work and what you can spend, then hands you the shortlist and the reason behind each pick.

Find your tools

Keep reading

Find your toolsAll AI guides