The useful dividing line in AI design tooling is not whether a product has an agent button. It is whether that agent can inspect a meaningful input, produce a meaningful artifact, and leave behind evidence that another person or program can check.
This week’s launch watch covers changes and updates available from October 27 through November 3, 2026, plus one clearly labeled boundary item from October 26. It does not claim that anything shipped on November 3.
That leads to three surfaces:
- Protocol: an agent can communicate through a documented interface such as MCP, a CLI, or an API.
- Files: an agent can read or write inspectable artifacts, even without a formal protocol.
- Nothing: no agent-facing read or write path is documented on the public page.
The interesting design problem lives in the middle of this spectrum. A file-based interface can be practical, extensible, and surprisingly agent-friendly; it can also expose more ambiguity than a nominal API. Protocol-based products may be easier to compose, while file-based products may be easier to inspect. This watch therefore scores what the public documentation makes visible, not what a product might be capable of doing internally.
The Scoring Rubric
Every tool receives four scores from 0 to 2, based only on its public pages:
- Read surface: 2 for a documented protocol or programmatic read path such as MCP, CLI, or API; 1 for named, inspectable artifacts or observable workspace and file reads; 0 when nothing agent-facing is documented.
- Write surface: 2 for a documented protocol or CLI write path, or content an agent can author; 1 for files, pull requests, or diffs that appear as reviewable artifacts; 0 when no agent write path is documented.
- Verification surface: 2 for public, version-numbered, or third-party-hosted artifacts; 1 for structured artifacts visible only inside the tool’s environment; 0 when there is nothing for a third party to check.
- Lock-in exposure: reported as a named dependency rather than a numerical score—for example, “host: Photoshop,” “protocol: none, files only,” or “workflow: VS Code + Codex/Claude Desktop named.”
A zero is therefore a legitimate scoring result. It means the public documentation does not establish that surface, not that the team failed to cover the product.
Kombai Extends Its Workspace, Review, and Chat Boundaries
Kombai’s October 27–November 2 changelog adds several concrete agent-facing surfaces. The source is its public Kombai changelog, which is dated and versioned.
Version 2.0.160, dated October 27, expanded VS Code workspace support to multi-root workspaces. The agent can read and write files across every open folder rather than only the first. Tabs and the active-file indicator identify the relevant folder, while the Changes panel and file history label edits with their folder name.
That release also made edits made by terminal commands visible in the Changes panel with a real diff and an undo action. Instead of terminal-driven changes disappearing from the review trail, sed and mv operations now become reviewable artifacts. The same release added GPT 6.1 Terra and MiniMax M3.1 Flash to the model dropdown.
In 2.0.161, dated October 28, Kombai added live-page inspection when asked to build something inspired by an existing website. The changelog says the agent inspects the page, scrolls through its sections, and captures colors, fonts, spacing, and animations. It does not explain how this inspection is implemented, so no DOM, browser, or screenshot mechanism can be inferred.
The release also introduced quarterly billing for personal plans, alongside monthly and yearly billing. 2.0.163 then moved purchasing into the Codex app, where the plan card offers either a plan or a one-time credit pack. These are billing cadences, not prices: the changelog publishes no plan price, and it does not establish that quarterly billing costs less than monthly billing.
In the same October 30 release, when Kombai credits run out, a plan card appears in chat, with payment handled through Stripe in the browser. A design brief produced by “Kombai Taste” can appear as a card in Codex desktop and Claude Desktop.
2.0.167, dated November 1, made Kombai’s latest skills available automatically at startup in Codex, Claude Code, and Cursor. The same release simplified design settings by removing Passes.
Finally, 2.0.169, dated November 2, improved the Codex canvas. “Open in browser” now opens a design’s actual preview page in the system browser instead of the Codex in-app browser. Selected designs also render fully, with Google Fonts, stock photos, icon libraries such as Font Awesome, Phosphor, and Lucide, module imports, and placeholder images loading in the canvas. Clarifying questions now always appear as interactive question cards and wait for an answer rather than sometimes appearing as plain text while the model continued.
Kombai’s read and write surfaces are therefore files, workspace context, diffs, and cards — plus a protocol path that its changelog documents but this week’s entries do not mention: version 2.0.113, dated September 12, 2026, introduced Kombai’s MCP server, which lets an agent “start a Kombai chat in a named workspace folder, follow a running chat and read its result, and read the designs on a canvas” per its MCP server documentation. Its public changelog makes the change trail independently checkable.
Modeinspect Connects Code and PRs, but Leaves a Limit Specification Hole
Modeinspect’s pricing page is new material for this watch. It is an undated page, fetched 2026-11-03, so the date of its underlying offer cannot be established from that page alone.
The page lists a Hobby tier at “$0/ month,” including “The full canvas, connected to your codebase,” “Unlimited connected repositories,” “Monthly agent limits for personal projects,” and “No credit card required.” A Pro tier is listed at “$20/ month” per person, with “Extended agent limits for daily work,” “Merge-ready PRs, straight from canvas,” “Design QA your live product,” and “Priority support.” It is “Billed monthly. Cancel any time.” Teams and enterprises are listed as “Custom,” with scale-based pricing, SSO/SAML, SCIM and SOC 2 Type II, centralized billing, and a dedicated contact.
The pricing FAQ explicitly says: “Will you bill me by tokens or usage? No. On Pro, $20 is $20: flat, per person, per month.” It also says: “What happens if I hit my Hobby limits? Nothing breaks and nothing gets charged. Your work stays where it is, and your limits reset at the start of the next month.”
The important issue is not simply that two tiers have different prices. It is a specification hole. Neither “monthly agent limits” nor “extended agent limits” is quantified. An agent—or a team deciding what can be delegated—cannot determine how many actions are available from the public page.
The code connection supports a file-level read interpretation, while “merge-ready PRs” establishes a reviewable write surface. But the evidence does not document an MCP server, API, or CLI. Verification also depends on artifacts visible in Modeinspect’s own environment rather than a documented third-party hosting or versioning path.
Shotcandy Has a Broad Tool Surface—and No Protocol
Shotcandy’s tools page enumerates a substantial browser-only surface:
- A screenshot beautifier with backgrounds, shadows, frames, exact social sizes, 42 one-click styles, wallpapers, and 3D tilt.
- A batch editor for up to 100 screenshots, with ZIP export, individual adjustments, retained filenames, and folder input.
- A multi-screen mockup tool for two to six screenshots in six named layouts: Side by side, Overlap, Hero, Cascade, Fan, and Grid.
- An App Store screenshot maker for three to 10 slides at exact sizes, including headlines and PNGs without transparency.
- A code screenshot tool that outputs PNG, MP4, or GIF, with eight themes, highlighted lines, line numbers, and a window title.
- An Open Graph image maker at 1200 × 630, plus X, LinkedIn, and Facebook sizes, plus a file-size check.
- A macOS window-frame tool, including a browser frame with a custom URL.
- A redaction tool offering blur, pixelation, or solid-box coverage before sharing.
The product is described as free, open source, and browser-only: nothing is uploaded and no account is required. Its source is available in the public Shotcandy GitHub repository, under the MIT license.
That public repository gives Shotcandy a strong third-party verification surface. Its functional breadth, however, does not create an agent interface. The reviewed pages document no MCP server, API, or CLI. Under this rubric, its read and write scores are zero even though people can use the tools interactively. The MIT repository is also the clearest example of the distinction between verifiable software and an agent-callable surface.
PixelCrew’s Real Integration Surface Is Its Artifact Chain
PixelCrew presents a sequential agent crew: Elena handles Research and Strategy, Marcus handles Direction and Art Direction, Mira handles Design and UX Architecture, and a build crew follows. The page describes average delivery as 25–45 minutes.
Its named artifact chain is more important for agent perception than that time estimate. The page names:
research-brief.md, CREATIVE_BRIEF.md, wireframes.html, ux-flow.json, copy.md, tokens, components, qa-report.md, landing-page.html, design-system.json, and docs.
These are machine-readable files in the sense used here: named artifacts that can carry structured content between stages. The evidence does not establish that design-system.json or ux-flow.json conform to an external schema, so no stronger claim should be attached to them.
PixelCrew also documents a bring-your-own-key alpha arrangement. It charges nothing while in alpha, agents run using the user’s own key, and model costs are paid directly to the provider. OpenRouter is recommended; Anthropic and Google Gemini keys also work. The page says there is no subscription, markup, or card on file. A Pro tier without a published price lists hosted keys, one bill for model costs, priority processing, brief history and versioning, and team seats.
There is no documented public API or MCP endpoint. Here, the integration point is files, not protocol. PixelCrew therefore offers a visible artifact chain and reviewable outputs, but not a documented programmatic composition interface. The homepage is an undated page, fetched 2026-11-03.
Resurf Documents Protocol Access Read-Only — and Contradicts Its Own Price
Resurf is a personal context library that Claude, Cursor, and Codex can search and read over MCP. Its homepage also lists a CLI, supports keys from OpenAI, Anthropic, Gemini, or OpenRouter, and offers local models through Ollama. The page advertises a seven-day free trial, macOS 14.3+ support, and free iPhone and iPad apps.
Its comparison page adds a permission detail the homepage does not: “Through MCP or the command line, Claude, Codex and Cursor can search and read the library, and you decide whether an agent may also change it.” So the protocol surface is read by default with an opt-in change permission the user grants — not a separate documented write API. Verification remains limited by the review evidence; there is no public, version-numbered artifact path shown here.
The price information requires careful handling. Resurf’s Resurf 2 launch post says: “The licence: $49 for two years of updates, or $99 for lifetime updates. One-time, no subscription.”
Its comparison page, which says product and pricing details were checked against official sources on October 19, 2026, instead says Resurf is “$39 one time with two years of updates, or $79 for lifetime updates.” The homepage purchase CTA also reads “Buy a license · $39.”
Those are two price pairs from the same vendor. Neither figure is declared correct here. The more recently checked comparison page is not automatically authoritative, because its recency establishes when its editor checked the information, not whether that information supersedes the dated launch post. The Resurf homepage is an undated page, fetched 2026-11-03.
ABrush Is a Photoshop Boundary Item
ABrush’s public page describes AI generation directly inside Photoshop, with Stable Diffusion, Flux, Gwen Image, and other models; “23+ Top models inside”; ControlNet; IP-Adapter/References; generation history; presets and preset sharing; and LoRA styles. It also states that users own generated images, that data is not used for training, and that commercial projects receive full privacy.
This is an undated page, fetched 2026-11-03. A dated company LinkedIn announcement says, “ABrush is now live on Adobe Exchange,” dated October 26, 2026. It is therefore a boundary item, falling one day before the watched October 27–November 3 window.
The Adobe Exchange listing provides a third-party verification point. But no MCP server, API, or CLI is documented, and no price is published on the public page. The principal dependency is explicit: host: Photoshop.
Designeer Is a Human Index, Not an Agent Surface
Designeer calls itself “A curated collection of interface craft, component libraries, AI tools, and resources for builders.” Its pages are reviewed by hand.
The visible taxonomy covers All, Interface, Motion, 3D, Typography, Branding, Logo, Illustration, Landing page, Mobile app, Dashboard, Icons, Prototype, Generative, and Game UI, among others. A color filter and “Recent” sort help humans navigate the collection.
Designeer is an undated page, fetched 2026-11-03. It has no public API, machine-readable feed, or other documented programmatic surface. Its curation is useful to people, but a person would need to copy the resulting information elsewhere before an agent could use it.
Accordingly, Designeer receives zeros for both agent read and write surfaces. Its curated pages are visible inside the tool’s environment, yielding a verification score of 1. Those zeros are legitimate results: the absence of documented access is exactly what the scorecard is designed to expose.
Comparison Table
| Tool | Read surface | Write surface | Verification surface | Lock-in exposure | Agent interface documented? | Score (R/W/V/L) |
|---|---|---|---|---|---|---|
| Kombai | Multi-root VS Code workspace files; MCP server (2.0.113) | Files, terminal-command diffs, undo, chat cards; agent-started chats over MCP | Public versioned changelog and public MCP docs | Workflow: VS Code + Codex/Claude Desktop named | Yes: MCP server; plus files and cards | 2/2/2/VS Code + named apps |
| Modeinspect | Connected codebase; agent limit amount unspecified | Merge-ready PRs | Structured artifacts in product environment | Host: vendor product; exact agent limits unspecified | No MCP/API/CLI documented | 1/1/1/vendor environment |
| Shotcandy | No agent read path documented | No agent write path documented | Public MIT GitHub repository | Protocol: none; browser and local files | No | 0/0/2/browser + MIT files |
| PixelCrew | Named artifact chain and machine-readable files | Named artifacts produced by crew | Structured artifacts in product environment | Provider keys and vendor workflow; no protocol | Files only | 1/1/1/files + model provider |
| Resurf | MCP and CLI | Read plus user-granted change permission (comparison page); no dedicated write API | No public versioned artifact path documented | Protocol: MCP/CLI; client choices remain optional | Yes: MCP and CLI | 2/1/0/protocol: MCP/CLI |
| ABrush | No agent read path documented | No agent write path documented | Adobe Exchange listing | Host: Photoshop | No | 0/0/2/host: Photoshop |
| Designeer | No agent read path documented | No agent write path documented | Hand-curated pages in product environment | Workflow: human browsing and copying | No | 0/0/1/human-curated index |
The lock-in column intentionally contains no numbers. A protocol such as MCP is not the same dependency as a desktop host such as Photoshop, and neither is equivalent to a file-only workflow. Naming the dependency is more useful than compressing unlike risks into one score.
FAQ
Which tool has the strongest documented agent interface?
Kombai documents the broadest agent interface in this group. Its changelog announces an MCP server in 2.0.113 (September 12, 2026) that lets an agent start a chat in a named workspace folder, follow it, read its result and read canvas designs, and this week’s releases extend the file-level surface underneath it. Resurf documents MCP and CLI as well, but as read access to a context library, with a change permission the user grants rather than a dedicated write API.
Which tool is best for inspecting an agent’s work?
That depends on what “inspecting” means. Kombai has a public versioned changelog and exposes terminal-command changes as diffs. Shotcandy and ABrush have public third-party hosting points—the MIT repository and Adobe Exchange listing. PixelCrew, Modeinspect, and Designeer expose structured material primarily inside their own product environments.
Does a zero mean the product cannot be used by agents?
No. A zero means no agent-facing surface is documented on the reviewed public pages. It is not a claim about hidden capabilities, internal implementation, or hands-on behavior.
How many agent actions does Modeinspect permit?
The public page does not quantify them. “Monthly agent limits” and “extended agent limits” are unquantified, so the number cannot be responsibly stated.
Which Kombai pricing cadence is cheapest?
The evidence does not answer that. The changelog documents monthly, quarterly, yearly, and one-time credit-pack options but publishes no prices. It therefore cannot support a comparison between cadences.
Why are both Resurf prices shown?
Because the same vendor’s dated August 8 launch post and an October 19-checked comparison page disagree. Presenting only one would conceal the contradiction. The newer source is not automatically authoritative, so this guide does not resolve which pair is current.
Was Product Hunt ranking used?
No. No Product Hunt data was fetched for this week. There is consequently no leaderboard, rank, or upvote count in this assessment.
Did the author test these tools?
No. This is a desk review of public documentation fetched on November 3, 2026, with no hands-on testing.
The Bottom Line
This week’s strongest development is Kombai’s expansion of observable workspace behavior: multi-root reads and writes, terminal-command diffs, interactive question cards, automatic skill delivery, and more complete canvas rendering. Its public, dated changelog also makes the change trail unusually easy to inspect.
Resurf documents MCP and CLI read access for other agents’ context, with a change permission the user grants instead of a dedicated write API, and its public pricing evidence conflicts. Modeinspect provides a clear code-to-PR workflow while leaving agent limits unquantified—a material spec hole for delegation. PixelCrew demonstrates the value of a named artifact chain, but its documented integration point remains files rather than protocol.
Shotcandy, ABrush, and Designeer score poorly for agent access because their public surfaces are human-oriented, not because their software is necessarily weak. ABrush has the clearest host dependency, Shotcandy the clearest third-party code verification, and Designeer the clearest example of human curation without a machine-readable layer.
The central lesson is straightforward: design-agent quality depends on the quality of the surfaces agents can perceive, not on the number of buttons or models attached to the product. The most promising tools expose work across at least two of the three surfaces—read, write, and verification—and make the boundary between them legible to another agent or reviewer.
How This Guide Was Built
This is a desk review only, conducted from public vendor pages and the dated public announcements listed inline. Sources were fetched on November 3, 2026. No hands-on testing was performed, so the guide does not claim that any tool behaves in practice as its documentation describes.
Dated sources are identified by their published dates. Where a page itself was undated, it is labeled “undated page, fetched 2026-11-03.” The ABrush Adobe Exchange announcement is explicitly treated as an October 26 boundary item. No Product Hunt data was fetched for this week, so no leaderboard, ranking, or upvote information appears.
Scores use only the public evidence summarized in the rubric. A zero records missing documentation; it is a legitimate finding rather than an expression of uncertainty about undocumented behavior. Where vendor sources conflict, as with Resurf pricing, both are reported without selecting an authoritative figure. No schema, implementation, benchmark, price, or capability has been inferred beyond the cited source material.