↑ ↓ to move · Enter to open · Esc to closeTap a result

Architecture becomes data. The agent obeys.

An agent harness for Spring Boot on Claude Code.

guides · sensors · blueprints

32 skills·7 agentsin the harness, as of release v0.44.0 →
§ in 70 seconds

Claude Code brings the model. Nerviz brings the rest.

It is not a coding assistant: it brings no model and no editor. It is not an agent framework: those build agents, this one constrains one. It is not a template: it generates through Spring Initializr at run time, and its job continues on every feature.

nerviz-pitch · 69 s · no audio
01

Architecture as data

Nine blueprints in YAML, plus your own. One declaration feeds the model, the hook and the compiler.

Read more →
02

Spec before code

One use case per run, designed layer by layer by an owner skill, approved by you, then frozen.

Read more →
03

Blocked at write time

A forbidden import, a write outside a skill's territory, an edit to an approved spec, a force push: a hook refuses it. The model is not asked to remember.

Read more →
04

Cost audit

A hook records each run: tokens per piece, the chain of skills, files touched, questions asked. It costs zero tokens, and /audit-usage aggregates the runs.

Read more →
05

Self-contained projects

A generated project carries the whole harness and its own CI. /arch-adopt updates it later; below Spring Boot 4.0, /platform-upgrade gets it there first.

Read more →
06

Linux, macOS and Windows

Every hook is one ArchHook.java, with no shell. CI injects each violation on Ubuntu, macOS and Windows and requires the block.

Read more →
§ the problem

Writing a skill takes minutes. Governing them is the hard part.

Teams adopt coding agents fast, each in its own way. Months later the same problems show up in every repository.

  1. 01

    No standard across teams

    Each team writes its own prompts, skills and conventions. The company uses AI, but there is no one way to use it, and no way to compare results.

    −7.2%in delivery stability, an estimated drop as AI adoption grew.source: DORA 2024 ↗

    One harness exported into every project, from a blueprint in YAML.

  2. 02

    Easy to create, hard to maintain

    A skill takes minutes to write. Keeping it scoped, named and current never ends, and a misspelled frontmatter field is ignored without a word.

    25% → <10%In 2021, 25% of changed lines were moved code, the trace of refactoring; in 2024, under 10%. For the first time, copy/paste outran moved code.source: GitClear 2025 ↗

    A schema checked in CI; /arch-adopt updates without losing local edits.

  3. 03

    No governance

    Which agent may write which folder? Who approved the design? Without a rule in code, the agent decides, and the answer changes between runs.

    1 production databasedeleted by an agent during a code freeze.source: AI Incident Database, incident 1152 ↗

    Write territory per class, enforced by a hook. Deny by default.

  4. 04

    Cost and traceability you cannot see

    Tokens go out per piece, per model, per run. Nobody can say which skill spent them, which files it touched, or what failed.

    4 monthsfor Uber to exhaust its whole 2026 AI budget, as Claude Code adoption spread faster than planned.source: Forbes 2026 ↗

    /audit-usage: tokens, model, files and failures per piece, at zero token cost.

  5. 05

    Same prompt, different code

    The model reads a skill's description and decides whether it applies. Loaded in one run, skipped in the next, the same request yields different code.

    66%of developers name AI solutions that are “almost right, but not quite” as their top frustration. 46% distrust AI accuracy; 33% trust it.source: Stack Overflow Developer Survey 2025 ↗

    Commands are typed, norms load by path, hooks run on every event.

  6. 06

    The architecture erodes quietly

    The build is green and a domain class now imports a framework. It surfaces in code review weeks later, if at all.

    45%of the time, given a secure and an insecure way to write the code, the models chose the insecure one. In Java, the failure rate tops 70%.source: Veracode 2025 ↗

    forbidden_imports declared once, refused at write time.

§ the life of a use case

From a one-line request to a green build

  1. 01/arch-doctor

    Is enforcement active on this machine?

    ✅ operational
  2. 02/init-project

    Initializr, blueprint, harness exported into the project.

    20 min
  3. 03/new-feature

    A short interview, one skill per layer, a frozen spec.

    10 min
  4. 04/new-feature-implement

    A restricted executor in a clean session.

    310 tests · 0 fail
  5. 05/audit-usage

    Time, tokens and failures of every piece.

    0 tokens to produce
§ where to start

Three starting points, one harness.

JDK 21+, git and curl. No Python, no Node, no shell scripts.

I have no project

A new project

/init-project

A Spring Boot project from the Initializr, with a blueprint, executable boundaries and a green build. No business code. It carries its own .claude/.

Guide · /init-project →
I have a project on Spring Boot 4.0 or later

An existing project

/arch-adopt

Installs the harness into a project born elsewhere, or updates one that is behind. It refuses a dirty worktree and stamps a hash for every file it writes.

Guide · /arch-adopt →
new · v0.41
I have a project below Spring Boot 4.0

Raise the line first

/platform-upgrade/arch-adopt

/arch-adopt refuses that project. /platform-upgrade takes the build to the line the Initializr serves today:

  • the version, the BOMs, and renamed starters and artifacts
  • renamed types and configuration properties
  • tests before and after: test-diff tells what the upgrade broke from what was already broken
  • it never commits and never loosens a failing test
Guide · /platform-upgrade →
§ public proof · nerviz-ai/bank-app
Measured by others, live
  • GitHub Actions
  • SonarQube Cloud
  • GitHub

Nerviz does not draw these: GitHub and SonarQube Cloud render them on every visit.

Don't trust it, run it
git clone https://github.com/nerviz-ai/bank-appcd bank-app./mvnw verify

JDK 21 and Docker. Without Docker, the integration tests are skipped. The latest main build printed:

[INFO] Tests run: 287, Failures: 0, Errors: 0, Skipped: 0[INFO] Tests run: 80, Failures: 0, Errors: 0, Skipped: 0[INFO] All coverage checks have been met.[INFO] BUILD SUCCESS