JARVIS

Your AI engineering team, one command away.

JARVIS is a model-agnostic multi-agent orchestrator that coordinates specialized AI agents in configurable pipelines — with quality gates, cost controls, and full audit trails.

What JARVIS does

You say: "add dark mode support"
JARVIS runs: PLANNER → BUILDER → FIXER → TESTER → REVIEWER + DESIGNER
You get: working code, tests, quality audit, PR created

Each agent is a specialist:

• PLANNER decomposes the task into sub-steps
• BUILDER implements the code
• FIXER resolves type/lint/build errors
• TESTER writes tests
• REVIEWER audits quality (/100 score)
• DESIGNER audits UI/UX (/100 score)
• DEVOPS handles infrastructure
• RESEARCHER analyzes and compares options
• SECURITY scans for vulnerabilities

Install

git clone https://github.com/bzbhh/JARVIS-IA.git
cd JARVIS-IA
npm install
npm link    # makes 'jarvis' available globally

Quick start

# Initialize JARVIS in your project
cd your-project
jarvis init

# Run a pipeline
jarvis run feature "add user authentication"
jarvis run fix "resolve TypeScript errors"
jarvis run research "compare Supabase vs Firebase"

# Run a single agent
jarvis simple REVIEWER "audit the latest changes"

Providers

JARVIS works with multiple AI providers. You need at least one:

| Provider | Setup | Tool use | Cost | |----------|-------|----------|------| | Claude CLI | npm i -g @anthropic-ai/claude-code | Full (native) | Via Claude subscription | | Anthropic API | export ANTHROPIC_API_KEY=... | Full (ReAct loop) | Pay per token | | OpenAI API | export OPENAI_API_KEY=... | Full (ReAct loop) | Pay per token | | Gemini API | export GEMINI_API_KEY=... | Full (ReAct loop) | Pay per token | | Ollama | Install Ollama | Text-only | Free (local) |

If a provider is unavailable, JARVIS automatically falls back to the next one in the chain.

Pipelines

| Pipeline | Agents | Use when | |----------|--------|----------| | feature | PLANNER → BUILDER → FIXER → TESTER → REVIEWER + DESIGNER | New features | | fix | FIXER → TESTER | Bug fixes | | infra | DEVOPS → FIXER | DB migrations, CI/CD | | design | DESIGNER → FIXER | UI/UX audit | | research | RESEARCHER | Analysis, comparison | | pentest | SECURITY → REVIEWER | Security scan | | fullstack | PLANNER → DEVOPS → BUILDER → FIXER → TESTER → REVIEWER + DESIGNER | Full stack changes |

Quality gates

Audit agents (REVIEWER, DESIGNER, SECURITY) score code /100:

• >= 90: PASS — pipeline continues
• 70-89: CONDITIONAL — FIXER runs, then re-audit
• < 70: FAIL — pipeline stops

Cost management

• Smart routing: simple tasks use cheap models, complex use premium
• Budget caps: per-agent ($5 default) and per-pipeline ($50 default)
• Cost tracking: every agent's cost is logged
• jarvis estimate feature — forecast cost before running

Commands

jarvis init                    # Setup JARVIS in your project
jarvis run <pipeline> <task>   # Run a pipeline
jarvis run "natural language"  # Auto-detect pipeline from description
jarvis simple <AGENT> <task>   # Run a single agent
jarvis status                  # Show running pipeline + queue
jarvis runs                    # List recent pipeline runs
jarvis cost                    # Cost report
jarvis estimate <pipeline>     # Cost forecast
jarvis resume <trace_id>       # Resume a crashed pipeline
jarvis postmortem <trace_id>   # Analyze a failure
jarvis login [provider]        # Set API keys
jarvis providers               # List available providers
jarvis config                  # Show resolved configuration
jarvis doctor                  # System health check
jarvis create-agent            # Build a custom agent
jarvis dashboard               # Web UI at localhost:3333
jarvis watch                   # Auto-process GitHub issues
jarvis serve [port]            # REST API server
jarvis run <pipeline> --dry-run # Preview without executing
jarvis config --validate       # Validate configuration

Configuration

jarvis init creates jarvis.config.json with auto-detected settings:

{
  "project": "my-app",
  "rulesFile": "CLAUDE.md",
  "commands": {
    "build": "npm run build",
    "test": "npx vitest run"
  },
  "qualityGate": {
    "threshold": 90,
    "conditionalMin": 70
  },
  "limits": {
    "maxBudgetPerAgent": 5,
    "maxPipelineBudget": 50
  }
}

See CLAUDE.md for full configuration reference.

SDK

import { Jarvis } from "@jarvis-ai/cli";

const jarvis = new Jarvis({ projectDir: "." });

// Listen to events
jarvis.on("pipeline:done", (r) => console.log(`Done: $${r.cost}`));
jarvis.on("pipeline:fail", (r) => console.log(`Failed: ${r.error}`));

// Run a pipeline
const result = await jarvis.run("feature", "add dark mode");
console.log(result.success, result.cost, result.traceId);

// Estimate cost before running
const estimate = await jarvis.estimate("feature");
console.log(`Estimated: $${estimate.total}`);

// Route natural language
const route = jarvis.route("fix the login bug");
console.log(route.pipeline, route.complexity); // "fix", "simple"

Full TypeScript support included (.d.ts).

REST API

jarvis serve --port 3001                    # Start API server
jarvis serve --port 3001 --key=my-secret    # With authentication

# Endpoints
POST /api/pipelines          # Start a pipeline
GET  /api/pipelines/:id      # Pipeline status
GET  /api/pipelines/:id/events # SSE event stream
GET  /api/runs               # Recent runs
GET  /api/cost               # Cost report
GET  /api/providers          # Available providers
POST /api/route              # Route text → pipeline
GET  /api/health             # Health check

GitHub Action

- uses: bzbhh/JARVIS-IA/action@main
  with:
    pipeline: feature
    task: "add user authentication"
    budget: "10"
  env:
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

Skills

Each agent loads domain-specific skills that enhance its capabilities. JARVIS ships with 24 skills covering coding standards, testing patterns, security review, deployment, and more.

Custom skills can be added to templates/skills/<name>/SKILL.md.

Known limitations

• Ollama provider is text-only (no tool execution) — useful for analysis, not code changes
• Dashboard file watchers only activate after first pipeline run
• npm package not yet published (install from source)
• Orchestrator test coverage is in progress

Documentation

• CLAUDE.md — Architecture, config, conventions
• AGENTS.md — Agent coordination guide
• CONTRIBUTING.md — How to contribute
• ROADMAP.md — Development plan
• docs/PRODUCT-PLAN.md — Product strategy

License

MIT