Build1 distinct publisher3 min readPublished
A low-code vendor is quietly positioning itself as the layer a coding agent writes on top of, and the evidence for that sits in its documentation and its package manifest rather than in any announcement.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Follow any of these and your For You feed starts watching them — no settings page required.
build
AgentCheck reviews what the agent actually changed, and does it without another model1 distinct publisher
build
Resend is selling to the agent and handing the humans free templates1 distinct publisher
product
Four leaderboards, four denominators: what you buy when you standardize on a coding agent1 distinct publisher
build
Copilot is now the minority tool, and your 2025 standardization decision knows it1 distinct publisher
The number that carries the argument is two. `@nocobase/preset-nocobase` is the package that bundles the built-in plugins, and its dependency list holds 104 entries on stable 2.2.5, the same 104 on 2.3.0-beta.6, and 106 on 3.0.0-alpha.13, with nothing removed [11]. On composition evidence, then, the 3.0 line is the 2.3 beta plus two plugins, a delta of under 2 percent of the bundle [22]. The post counts those two packages without naming them [25]. The scope is bounded; the contents are not.
The release notes prove the weaker instrument here. Across alpha.4 through alpha.13 the component tally comes out flat at the top: Portal manager 9, AI employees 9, client-v2 9 [13]. Split by date the shape returns, because every portal change landed in the first ten days and the later alphas absorb general work shared with the 2.3 line, including AI employees, client-v2, Workflow and Multi-space [14]. alpha.10 on its own runs to roughly 11,000 characters of notes covering the whole product [15]. That is a release note doing a newsletter's job, covering the whole product instead of just what changed.
The version plumbing is worth reading before you script against it. Package composition here came from npm and the notes from the GitHub Releases API [9], and alpha.1 through alpha.3 exist only on npm, with GitHub Releases starting at alpha.4 [12]. So an enumeration from the Releases API undercounts the line by three. The documented floor of 3.0.0-alpha.6 sits above that gap and well below the tip: npm's alpha dist-tag pointed at 3.0.0-alpha.13 [10], seven releases past the stated minimum [23]. Thirteen alphas between July 30 and August 31 works out to a new one about every 2.7 days [24].
The Quick Start states its design problem in its own words: AI vibe coding can produce a good-looking page, but it has a hard time connecting to a real business system, or it reimplements authentication, permissions and collection design from scratch [17]. The proposed division hands NocoBase the kernel work, meaning authentication over OIDC, SAML, CAS, LDAP, SMS and others, collections and external data sources, CRUD REST APIs, field- and record-level ACL, workflows and file storage, and hands the agent the React pages and business logic served on top [20]. AI Portal is that entry, its source lives locally, and it is reserved for the AI Agent to write [19]. One portal change from those first ten days closes the loop: copy a prompt from an AI portal's card to connect a coding agent [16]. NocoBase's own framing is that you treat the platform as the foundation of the system kernel so the agent can focus on business logic [18].
For the trade to pay in your context, the work you have left has to genuinely sit above that substrate. Your entities need to model cleanly as collections, your authorization needs to be expressible as field- and record-level ACL, and your identity provider needs to be one of the ones already supported [20]. Miss on any of those and the agent writes around the kernel instead of on top of it, which is the reimplementation the Quick Start is arguing against [17]. Credit where it is due on the reporting: the runtime checks ran on the official 3.0.0-alpha.9 Docker image with PostgreSQL 16 on a fresh install, and the portal seed data, path assignment and template version were re-verified against the alpha.13 packages [8].
Ranked by verification strength, evidence, and original report placement.
NocoBase has been publishing 3.0 alphas since July 30th, starting with 3.0.0-alpha.1.
As of the post's publication the 3.0 line stands at 3.0.0-alpha.13.
There is still no release announcement for 3.0: nothing on the NocoBase blog and no "Announcing NocoBase 3.0" post.
NocoBase 2.2 shipped as a stable release on August 23rd and received its own announcement, and the beta line has already moved on to 2.3.
docs.nocobase.com carries an AI Builder section spanning 18 pages.
The AI Builder section's centrepiece, the AI Portal Quick Start, states a prerequisite of NocoBase >= 3.0.0-alpha.6, and the docs include a link to request a demo environment.
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 31, 2026
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Primary artifacts, one set of eyes
Nearly every load here can be re-run by a reader: npm dist-tags, the GitHub Releases API, an 18-page docs section, a numbered forum thread, and a fresh Docker install of alpha.9 re-checked against alpha.13. That is better hygiene than most release commentary. It is also still one dev.to post — no second publisher has touched the dependency counts, NocoBase has said four words on the record, and the two packages that carry the entire 3.0 thesis are counted but never named.
Vendor shipping; no users in view
Thirteen alphas in a month and a documentation section with a request-a-demo link tell us NocoBase is building hard. They say nothing about anyone using it. The only running instance in this reporting is the author's own fresh install; there are no download figures, no customers, no third-party deployments — and the documented path demands Node 22, pnpm, an alpha CLI and a coding agent of your own before a single page exists.
Restraint in the text, reach in the thesis
The framing — a low-code vendor recasting itself as the kernel a coding agent writes on top of — leans on two extra packages in a 106-package bundle and a docs section nobody has announced. That is real weight on a small delta. What keeps the number near zero is the author's own bookkeeping: the alpha caveat sits above the findings, the release notes are explicitly disclaimed as a poor instrument because they also carry 2.3 work, and NocoBase itself is overselling nothing, having published no announcement to oversell.
The vendor's documentation does the arguing
The persuasive sentences in this story are not the author's; they are lifted from NocoBase's own AI Builder pages, which cast vibe coding as the problem and NocoBase as the kernel, and which sit beside a link to request a demo. The author is an independent dev.to writer with no stated relationship to the vendor, volunteers the method and flags the alpha risk — but every strategic claim still routes through marketing-adjacent documentation, and naming Claude Code, Codex and Cursor in your own prerequisites is positioning as much as it is instruction.
Checkable today, provisional by design
Two things hold this down, neither of them sloppiness. The subject is prerelease software whose composition can shift between builds — the author already had to re-verify seed data and template version against alpha.13 — and the whole account rests on one author, one publisher and one install. What is verifiable is verifiable right now; what it implies about shipped 3.0 remains a bet, and the version floor in the docs already trails the shipping alpha by seven releases.