Documentation

From zero to an agent that knows your React Native project. The full command reference lives in the repo — this is the fast path.

$ man vectalon

v0.22.4

name

vectalon — the AI engineering control plane for React Native teams: give it a repository and it continuously understands, reviews, diagnoses, upgrades, and validates the application

synopsis

$ npx vectalon <command> [options]
$ npx vectalon feature "login screen with auth API"

quickstart

[01]npm install --save-dev @vectalon-dev/rn && npx vectalon init

Install & init

Scans the repo, seeds the knowledge base, and enables the ecosystem items your project needs. No config files to write — Vectalon owns its own knowledge.

[02]npx vectalon serve

Serve the agent

Boots the MCP server. The package registers 64 project-aware tools and 44 deterministic commands, but availability follows the released lifecycle catalog. Beta onboarding and policy surfaces are enabled by default; experimental analysis, knowledge, model, and integration commands require opt-in and may need configured models, credentials, or network services.

[03]npx vectalon feature "login screen with auth API"

Use it

PRD → stories → acceptance criteria → implementation → tests → review. Every fix is compile-checked before it lands, and the terminal explains itself — live phase progress, a command feed, and parsed failure cards on failure.

commands · free — genuinely useful

fix "issue"

THE workflow — tell it what's broken (or pass --log) and get root cause → evidence → impact → recommended fix → applied → verification → confidence in one structured verdict, applied in a sandbox by default

score

The Vectalon Engineering Health Score — one 0-100 number from eight dimensions, the delta vs your last run, and P0/P1/P2 actions

init

The 15-minute proof of value — scan the project, build the knowledge base, and end with the scan summary + Health Score + Top 5 problems. No LLM config asked

mode

Where your source runs — Cloud (hosted models) / Private (company LLM) / Air-gapped (local model, nothing leaves the machine); enforced, not labeled

demo

The flagship demonstration — the feature workflow, live: Requirement → … → PR + the self-healing loop, from a real prior run when present, zero model calls

brain

The productized Team Brain — ask "Why Zustand?" and get the decision card (ADR, reason, approver, related, reviewed); ask "Who owns auth?" and get the expertise tree (owner, experts, ADRs, services, changes)

plan

Commercial-use plans and their qualified capability IDs; beta access is labelled separately and experimental capabilities require opt-in

outcomes

Engineering outcomes, not feature counts — the sales material: issues detected, automatically fixed or prevented, PR issues caught, build failures resolved, RN upgrades completed, tests generated, perf regressions detected, and the estimated developer-hours + dollars saved from your committed reports (--rate to set the blended rate)

serve

Run the MCP server — agents connect from your editor

feature "…"

Generate components, write tests, run workflows

doctor

Ecosystem + native toolchain + leaderboard readiness, with numbered fix steps

refresh

Re-fetch web intel + re-seed knowledge from the repo

status

One read-only health screen — daemon, MCP server, model, license

ecosystem

Browse the tooling catalog — MCP servers, skills, hooks; grouped, with --info cards

models

List the local model tiers — fast (1.5B) / balanced (3B) / quality (7B) — with the one auto-selected for your RAM; init picks it for you

pull [tier]

Download the local GGUF model — a usage tier (fast|balanced|quality) or a model id; defaults to your machine’s auto-selected tier

selftest

Test every harness feature in isolated sandboxes — live pass/fail stream

impact

Cross-package blast radius — affected screens, navigation stacks, and the Maestro E2E flows that must run (with accessibility variants for covered screens)

coverage

Per-screen E2E + accessibility gap dashboard with links to the open follow-up tasks

perf

Static performance scan — render-phase setState, memo-defeating props, heavy startup imports, legacy bridge traffic, with ranked fixes

bench / leaderboard

Run the RN benchmark suite against any model — --preset fast|balanced|quality runs the local tiers

commands · pro — teams & hard problems

upgrade

React Native / Expo upgrade copilot — rn-diff-purge diffs, AST impact analysis, codemods

ci

Self-healing CI generation

visual-ci

PR-mode visual regression — capture affected screens, diff vs committed baselines, post the report on the PR, exit with a gating code

ci-incident

Self-healing CI gate — file a triaged incident (severity, cause, rollback suggestion) for a failed CI gate into the team brain

visual-baseline

Manage the committed visual baselines — list, capture, update, prune, quarantine

bundle

Bundle budget guardrails in code review

profile

Hermes runtime analysis — JS-thread blocks, retained objects, leak candidates

sandbox

Run commands with deny-by-default env, no network, hard time/memory limits

render

Compile + headless-render generated code before the diff — Metro transform, sandboxed

sync

Experimental cross-project Team knowledge; explicit opt-in required

team-policy

Org-wide guardrail policy — publish/pull the team policy + shared bundle budgets through the sync remote (Team)

agents — deterministic, free, report-driven

phase 8 — review

review · arch · sec · build-fix · test-repair · refactor · deps · a11y · release-ready · bug-fix · score

phase 9 — release eng

crash · arch-score · cicd · app-store · soc2 · tokens · team-stats · perms · dashboard

phase 10 — enterprise

figma · sentry · observability · governance · audit · repos · release-predict · play-store · dataset · lora

phase 11 — platform & github

gh-pr · gh-issue · gh-ci · gh-sec · monitor · evals · search · incident · train · cost · dx

One card per agent — verdict, triggers, and the report it produces — on the agents catalog →

reports — see your own

Reports never leave your project unless you share them. Three ways to read them:

[01]vectalon <agent>

In the terminal

Every run prints the verdict, severity-ranked findings, and the fix plan to stdout — `vectalon dashboard` prints the aggregate across every agent.

[02]docs/vectalon/<cmd>/report.md

In your repo

Each agent writes report.md + report.json into your project — plain markdown/JSON that renders on GitHub/GitLab and in editors, and is machine-readable for your own dashboards or CI gates. Gitignored by default: reports stay local unless you commit or share them.

[03]vectalon dashboard

One HTML file, in a browser

Aggregates every agent report into docs/vectalon/dashboard/report.html — a self-contained page with per-agent drill-down, search, and severity filters. No server, works offline, portable: attach it to a PR or host it anywhere.

The documents on this site are generated from a demo project — see the real output →