NocoBase recorded 37 release events in 30 days, but the interesting part is not raw speed. The project is shipping stable 2.1 patches while pushing multiple 2.2 alpha and beta builds, and the same fixes often move through more than one channel. That makes NocoBase a useful test of a harder idea: can an open-source no-code platform add AI agents without weakening the permissions, workflows, and auditability that business systems require?

The current evidence says the team is treating that problem seriously. It also says buyers should evaluate the release train with care. High frequency here reflects both product momentum and the operational cost of changing a large platform in public.

NocoBase is positioning AI as an operator, not a code generator

The official site describes NocoBase as an AI plus no-code platform for building business systems. Its core pitch is specific. Coding agents get a CLI and skills, human builders get a WYSIWYG configuration interface, and both work against the same data models, pages, workflows, and permission rules.

That is a more credible direction than asking a model to generate an entire internal application as disposable code. Business software rarely fails because nobody can draw a form. It fails when permissions drift, workflow state becomes opaque, data models cannot evolve, or nobody can explain why an automated action happened.

The GitHub repository frames the technical foundation as a microkernel with composable plugins. It also presents standard relational data, workflows, access controls, and audit logs as reusable infrastructure. ToolVitals classifies the repository as OSI-approved open source under Apache-2.0. That matters for teams that need to inspect, self-host, extend, or fork the system, although licensing alone says nothing about the effort required to operate it.

The April NocoBase 2.1 beta announcement made the AI strategy concrete. The nb CLI became the official entry point for agents. The release introduced eight NocoBase Skills covering environment management, data modeling, UI building, workflows, permissions, solution generation, plugin management, and publishing. It also expanded AI employees with MCP support, a workflow node, web search, and reporting tools.

The distinction between building and operating is central. NocoBase uses coding agents to configure or extend an application, while AI employees work inside the running application. According to the official site, those employees receive roles with field-level read and write permissions, and their data changes and workflow triggers can be audited. That is the right security model on paper. It puts an agent inside an existing authorization boundary instead of giving it a magical side door.

The metrics show an unusually busy release train

ToolVitals gives NocoBase a 96 overall score, a 93 health score, and a 98 shipping score. Its hot score is 229.4. The repository has 23,345 stars, and ToolVitals observed 30 GitHub releases over 90 days plus 37 release events in the latest 30-day window. Data confidence is 100.

Those numbers support a clear conclusion: NocoBase is actively maintained and publishing at a high rate. They do not prove that every release is equally significant.

The recent event stream includes stable v2.1.24 and v2.1.25 releases alongside v2.2 alpha and beta builds. Several changes appear in multiple channels. The Redis cache URL fix, for example, is visible in v2.1.24 and v2.2 prereleases. Client V2 fixes and workflow corrections also move across the release branches. A release-event counter correctly records every publication, but readers should not interpret 37 events as 37 unrelated feature packages.

That overlap is not a flaw in ToolVitals data. It is useful evidence about NocoBase’s delivery model. The team appears to be maintaining a stable line while testing a broad client and AI transition in prerelease channels. For operators, the implication is simple: release count measures maintenance pressure and delivery cadence, not upgrade safety.

The missing metrics matter too. The payload has no 30-day commit count and no active-contributor count. ToolVitals therefore cannot use this snapshot to tell whether the work is broadly distributed or concentrated among a small group. The 23,345-star total shows substantial GitHub attention, but stars are not deployments, retained users, or successful upgrades.

Recent releases are mostly about boundaries and failure modes

The v2.1.25 release added DingTalk Stream synchronization and LDAP user synchronization, including optional Active Directory department sync. Those are enterprise integration features, not demo polish. The same release reduced main-process memory use while parsing knowledge-base documents and changed file URLs so permissions can be checked before redirecting to storage.

Its fixes are even more revealing. The release corrected AI tool-call approval checks so decisions are authorized against the conversation that owns the message. It also fixed attachment access through an external NocoBase data source. These are narrow changes, but they sit on important trust boundaries: identity, conversation ownership, file access, and external data.

The v2.2.0-beta.14 notes include another attachment hardening change that avoids trusting client-provided file paths or URLs. They also include permission filtering for approval-file associations and iframe sanitization for Markdown content. The v2.1.24 notes show the attachment fix reaching the stable line, plus Redis worker isolation and workflow corrections.

This pattern supports the article’s thesis. NocoBase is not merely adding chat boxes. It is encountering and fixing the security and reliability problems that appear when agents touch real application state. That is healthy engineering behavior. It is not proof that every relevant path is safe.

The client transition adds another layer of risk. The v2.2.0-alpha.8 release contains fixes across public forms, workflow V2 editors, the flow engine, relationship validation, and external data-source variables. The 2.1 announcement says client V2 introduces a new package, route, and per-plugin source structure, with the kernel and plugins migrating gradually.

NocoBase’s roadmap connects these changes to a larger plan: complete plugin migration to client V2, AI worktrees, automated validation and debugging, sub-agents, sandboxes, custom skills, and parallel multi-agent support. That is ambitious. It also means the project is changing the application builder, plugin model, and agent layer at the same time.

The background-research example shows the product’s strongest idea

A July NocoBase article on automating company background research provides a better product demonstration than generic AI claims. The example separates the current company profile from a table of background-check tasks. Workflows prevent duplicate unfinished jobs, scheduled checks fill missing research, and each generated report retains a history record.

That data model is the important part. The AI output does not silently overwrite the only copy of a business record. Current state and processing history remain separate. A user can inspect the report history, and the workflow can check for existing tasks before creating another one.

This is where NocoBase’s no-code heritage helps its AI story. Models are probabilistic, but database constraints, workflow states, permissions, and audit records are not. NocoBase can be valuable if it makes those deterministic controls easier to build around uncertain agent behavior.

The example is still first-party material. It demonstrates an intended architecture, not an independent production benchmark. A skeptical engineering lead should reproduce the pattern with realistic permissions, failure injection, concurrent workflow runs, model timeouts, malformed attachments, and rollback procedures before approving a deployment.

Shipping faster than larger developer tools, with important caveats

NocoBase’s 37 release events in 30 days exceed LangChain’s 17 in the supplied comparison set, even though LangChain has 141,817 GitHub stars and a 240.0 hot score. NocoBase also exceeds Kilo Code’s 27 release events, while Kilo Code has 26,226 stars and a 226.9 hot score. All three are classified by ToolVitals as OSI-approved open source.

This is not a quality ranking. LangChain and Kilo Code solve different problems, use different versioning practices, and may package work differently. The comparison shows that NocoBase’s publication cadence is unusually high even beside prominent developer tools. It does not show that NocoBase produced more code, more user value, or fewer regressions.

Composio offers another useful calibration point. ToolVitals records 59 release events in 30 days, 29,243 stars, and a 232.5 hot score. NocoBase is not the fastest publisher in the group. Its differentiator is the combination of high cadence with a full business-application substrate: relational data, forms, workflows, permissions, plugins, external data sources, and now agent interfaces.

What ToolVitals cannot infer

ToolVitals can observe stars, releases, site health signals, and other public activity. It cannot verify code quality, test depth, upgrade success rates, support response time, user satisfaction, revenue, or production adoption. It also cannot confirm the site’s claim that the infrastructure is production-proven.

The 93 health score and 98 shipping score are strong public-maintenance signals. They are not a security review. Recent fixes involving attachment paths, tool-call approvals, approval-file ownership, iframe sanitization, and permission-aware file URLs show active hardening, but they also identify areas that deserve focused testing.

“No hosted pricing tracked” does not mean NocoBase costs nothing. Self-hosting has infrastructure and operational costs. Some release-note labels also identify Pro components such as S3 Pro, so buyers should inspect the current feature and commercial boundaries directly before planning an implementation.

For maintainers, the release volume is both an asset and a communication problem. Stable, alpha, and beta entries propagate overlapping fixes, which can make 37 monthly events look noisier than the actual change set. A concise release-channel map, explicit backport labels, and a security-impact classification would help operators decide what to deploy without reading several near-duplicate changelogs.

Recommendation

Evaluate NocoBase if your team needs self-hosted internal software where agents and non-developers must work on the same governed data and workflow model. The strongest reason is not automatic app generation. It is the attempt to put AI building and AI operations behind ordinary permissions, traceable workflows, and relational records.

Run the evaluation on the stable 2.1 line first. Build one workflow that touches sensitive records, then test field-level access, approval ownership, attachment handling, retries, duplicate prevention, audit logs, and rollback. Treat 2.2 as an active migration track until its client and plugin changes settle.

If your requirement is only a conversational prototype, NocoBase may be more platform than you need. If the application must survive changing business rules and explain what an agent did, its architecture deserves a serious trial. The 37 release events show a team moving quickly. The permission and workflow fixes show what that speed is being spent on.

Sources