This is a desk review of official vendor pages and dated launch sources, fetched 2026-10-20. No tool in this post was tested hands-on. Vendor figures that matter to a buying decision are labeled as vendor claims. Where a source was incomplete, we say so: ABrush’s plan prices were not extracted, Designeer’s primary page was blocked, and Modeinspect’s own prices disagree.
What’s New This Week (and What We’re Not Repeating)
This launch watch covers the week of 2026-10-20. The 2026-10-13 edition covered Kombai 2.0 basics, the Figma agent, Modeinspect, and Resurf. This one focuses on crew-style and bring-your-own-key tools: PixelCrew and ABrush, plus Kombai’s changelog, which that edition did not cover, and a pricing discrepancy on Modeinspect’s own site. The comparison table from last week is not repeated.
The Perception Question: What Does an Agent Actually See?
An agent can only design as well as it can perceive. The lens for every tool here is its perception surface: what the agent is allowed to read. A free-text brief (PixelCrew). An open Photoshop document (ABrush). A codebase with its components and tokens (Modeinspect). The second half is the landing surface: where output goes. Files, pull requests, PSD pixels, or chat-linked canvases each imply different integration work for your team.
PixelCrew: A Crew of Agents, Your Own Keys
PixelCrew’s tagline is “Your design brief. Agents get to work.” The official homepage describes a named crew pipeline rather than a single model call. Elena (Researcher · Strategy) runs first and produces audience personas, a competitive audit, and an information-architecture recommendation. Marcus (Director · Art Direction) runs second and produces visual direction and a creative brief. Mira (Designer · UX Architecture) runs third and produces wireframes, user flows, and a section-by-section spec. A “build crew” then turns the blueprint into production HTML.
Pricing on the homepage has three parts. The bring-your-own-key tier is $0 from PixelCrew during alpha: you connect an OpenRouter, Anthropic, or Google Gemini key, and model costs bill to that provider. The Pro tier is marked “Coming soon” and is described with hosted keys and one bill, which would end pure BYOK. Enterprise is custom-priced. The homepage says “a typical brief costs a few dollars in model usage” and “average delivery: 25–45 minutes”. Both are vendor claims, unverified.
Perception surface: PixelCrew reads the brief and its own intermediate files. The page describes no codebase or Figma input. The homepage says the crew “assembles the design system: tokens, type scale, color, components” and lists design-system.json among its outputs. That system is generated by the crew, not read from your library. The Enterprise tier is where the page mentions “your design system baked in”. The question to ask a vendor is therefore whether your system or the crew’s generated one drives the output.
The handoffs are file-based, with markdown and HTML artifacts passed between stages (research-brief-output.md, CREATIVE_BRIEF.md, wireframes.html). That gives a pipeline you could audit and diff, which matters more than the delivery-time claim.
ABrush: AI Generation Inside Photoshop
ABrush is an AI image generation and editing plugin inside Adobe Photoshop, pitched as “AI studio directly in Photoshop for artists.” The primary page lists 23+ models, 5 AI model mods, 3 UI mods, ControlNet support, IP-Adapter and references, generation history, presets with sharing, and LoRA styles. Named models include Stable Diffusion, Flux, and Gwen Image. The page also says “No training on your data. Full privacy for commercial projects” (vendor claim). Its FAQ covers commercial use, but the answer text was not extracted, so check it before commercial work.
Perception surface: the open PSD and the artist’s own presets and LoRAs. No tokens, no code, no design system. Output is pixels in a PSD. The “Simple plans” section exists, but its figures were not extracted, so no price is reported. According to hunted.space (secondary), ABrush launched on Product Hunt on 2026-09-12 and had 147 upvotes per hunted.space. The maker, AB Games (the studio behind Hidden City), says its CEO’s launch post claims about three years of internal use on its own game production (vendor claim). ABrush is a real tool for a clear audience, but it is an artist’s tool, not a design-system tool.
Kombai’s Changelog, 2.0.113 to 2.0.149
Kombai’s changelog records these releases:
- 2.0.113 (2026-09-12): “Local Agents” lets Kombai use your existing Claude Code or Codex subscription through the installed CLI. The same release added the Kombai MCP Server, so you can use Kombai inside Claude Code, Codex, or Cursor.
- 2.0.137 (2026-10-02): a quick-edit toolbar on the design canvas for fonts, spacing, borders, corners, backgrounds, links, and alignment, with a Save/Discard/Undo strip.
- 2.0.144 (2026-10-10): an embedded chat view, where a coding agent such as Claude Code, Cursor, or Codex can return a link to the Kombai conversation and design canvas.
- 2.0.149 (2026-10-14): Gemini 4 Argon added to supported models, and plans can be copied or saved as files.
The 2026-10-13 edition did not cover the changelog, so all four entries are new to this blog. The homepage lists availability as a desktop app and as an extension for VS Code, Cursor, Antigravity, Trae, and Kiro. The vendor’s benchmark figures, originality up to +13 percentage points and creative range up to +23 points, remain vendor-stated and unverified; we did not review the benchmark page. Pricing was not verified on primary pages this run. A LaunchArchive post (secondary) dates the Kombai 2.0 launch to 2026-06-02.
Modeinspect: A Pricing Discrepancy Worth Flagging
Modeinspect sells a design canvas with your codebase and agents built in. Its homepage says “Design in code. Skip the handoff,” and “Vibe coding gets you to 80%. Modeinspect is the last 20%.” It also says “No MCP servers, no localhost,” which contrasts with Kombai’s MCP and CLI integrations. The homepage states that “every change opens as a normal pull request on a branch, with a description and screenshots.” Its comparison blog post says the canvas runs on a Next.js React codebase with Tailwind CSS.
Here is the discrepancy. The homepage pricing section shows $0/mo, $24/mo, and $48/mo tiers. The comparison blog says “Hobby is free and Pro is $20 a month per person.” We report both figures and state neither as fact. The vendor needs to reconcile its own pages before you commit budget.
On perception, Modeinspect is the strongest story in this set. It targets your real components and tokens. Its homepage shows three outcome statistics: “22 days saved on delivery cycle”, “0 engineering handoffs”, and “95% of UI from existing components”. A separate Product Hunt pilot claim, carried from the 2026-10-13 watch, says “design-to-merged-PR in 13 days” for one customer, Prelude. These are all vendor or single-customer claims, not independent measurements. The homepage’s 95% figure is not tied to a named customer, and the Prelude pilot’s roughly 95% may be the same claim. The customer logos on the homepage (Kiwi.com, moss, Apify, E2B, Prelude, NCCER, Deepnote) show who uses it, but they carry no verified outcomes.
Shotcandy and Designeer: Scope Notes
Two entries get scope notes rather than coverage. Shotcandy, per hunted.space (secondary only, because its primary site could not be fetched), is a free browser-based screenshot beautifier. It launched on 2026-09-28 with 132 upvotes at snapshot. It has no agent-perception surface and is not a design-agent tool. It appears here only because it was in a recent launch set. Designeer is excluded entirely: its primary site returned an antibot block, so we do not describe it.
Comparison Table: Perception Surface, Output, Key Model, Lock-In Risk
This table covers this week’s four tools only. Vendor claims are flagged inline. The lock-in column is editorial judgment, not a measurement.
| Tool | Perception surface | Output landing | BYOK support | Verified pricing status | Lock-in risk (editorial judgment) |
|---|---|---|---|---|---|
| PixelCrew | Free-text brief plus its own intermediate files; generates its own design-system.json; Enterprise adds yours | Production HTML plus markdown artifacts | Yes: OpenRouter, Anthropic, or Gemini key during alpha; Pro “coming soon” with hosted keys | BYOK $0 from PixelCrew (alpha); Pro coming soon; Enterprise custom | Low today; rises if Pro’s hosted keys become the default |
| ABrush | Open PSD plus the artist’s presets and LoRAs; no tokens, no code | Image pixels in a Photoshop document | Not described | Plan figures not extracted | Medium: pixels only, no token or code path back into your system |
| Kombai | Own design canvas; MCP and CLI links into coding agents; no codebase-read claim verified on primary pages | Canvas plus code links into Claude Code, Cursor, or Codex; plans exportable as files | Partial: Local Agents use your existing Claude Code or Codex subscription | Not verified on primary pages this run | Medium and falling: MCP and CLI integration reduce platform coupling |
| Modeinspect | Your real codebase, components, and tokens (Next.js React with Tailwind) | Pull request with description and screenshots | Not described | Discrepancy: homepage $0/$24/$48 vs blog “$20 Pro”; neither stated as fact | Low on output (PRs merge into your repo); the pricing ambiguity is its own risk |
The Third-Party Ranking Context
A third-party tracker, KanonAgent, lists these upvote counts in a snapshot dated 2026-10-17: a Claude-branded design entry at 873, Stitch by Google at 642, Inspector at 622, “Figma for Agents” at 578, Resurf at 244, PixelCrew at 138, and ABrush at 120. These are aggregator numbers, not Product-Hunt-verified. ABrush’s 147 (hunted.space) and 120 (KanonAgent, 2026-10-17) come from different sources and dates, and we do not reconcile them. Treat the ranking as directional context, not a leaderboard. A LaunchAIJam aggregator counts 22 AI tools launched in October 2026. We did not verify each one individually.
FAQ
Is PixelCrew actually BYOK? During alpha, yes. The official homepage says “PixelCrew charges nothing while in alpha. Connect your own API key,” and model costs are billed by your provider. The Pro tier is “coming soon” and is described with hosted keys. Verify current terms on the official site before relying on it.
Can ABrush output design tokens? No. Its perception surface is the open PSD and the artist’s presets, and its output is pixels in Photoshop. No token or code output is described on its primary page.
Is Modeinspect $20 or $24 per month? We don’t know, and that is the point. Its homepage shows $0/$24/$48 tiers, while its own blog says Pro is $20 a month per person. Ask the vendor before purchasing.
Did Kombai add MCP support? Yes. The changelog records the Kombai MCP Server in 2.0.113 (2026-09-12), which lets you use Kombai inside Claude Code, Codex, or Cursor. Version 2.0.144 added an embedded chat view where coding agents return canvas links.
What changed since the 2026-10-13 watch? Two tools new to this blog (PixelCrew and ABrush), Kombai’s changelog (not covered on 10-13), and a newly flagged Modeinspect pricing discrepancy. The Kombai 2.0 basics, the Figma agent, and Resurf are not repeated.
Why is Designeer not covered? Its primary site returned an antibot block. We excluded it rather than describe it from incomplete sources.
The Bottom Line
Per tool, one verdict. PixelCrew is the most interesting BYOK experiment this week: markdown handoffs, your own keys during alpha, and a spend cap you control. It perceives only its own brief, though. Per the page, your own system enters through the Enterprise tier, so ask for a demo that uses your system before you judge its output. ABrush is for artists, not design systems: strong inside a PSD workflow, irrelevant to token pipelines. Kombai’s changelog is dated and moving fast, with Local Agents, MCP, embedded chat, and Gemini 4 Argon all shipping since September. Its benchmark numbers remain vendor claims. Modeinspect has the strongest perception surface for the question “how do agents learn to design better”: it reads your codebase and lands output as pull requests. Its unresolved pricing discrepancy is a diligence blocker, though. If you are evaluating this week, start with Modeinspect’s pricing question and PixelCrew’s Enterprise tier. Ask both vendors before you buy.
How This Guide Was Built
This is a desk review of primary vendor pages and aggregator-verified launch dates, fetched 2026-10-20 via an in-session research lane; no tool was tested hands-on. Official pages were prioritized: kombai.com, kombai.com/changelog, pixelcrew.ai, abrush.ai, modeinspect.com, and the Modeinspect comparison blog. Secondary aggregators, including Product Hunt’s ABrush page, hunted.space, LaunchArchive, KanonAgent, and LaunchAIJam, are used only for launch dates, upvote context, and ranking snapshots, and are marked secondary. Every vendor number is labeled as a vendor claim. Where extraction failed (ABrush plan prices, its commercial-use FAQ answer) we disclosed the gap instead of filling it. The Modeinspect pricing discrepancy is disclosed, not resolved. Designeer was excluded after an antibot block.