Aura Workshop

AI Agent Orchestration
That Self-Organizes

It's the difference between having one AI assistant and having an autonomous AI-powered company that thinks, acts and performs on your behalf

Download for macOS Download for Windows
7+
LLM Providers
6
Classification Types
11
Validated Scenarios
3
Platforms

Describe what you want. The system classifies your intent, assembles the right team, and delivers working results. No workflow configuration. No agent selection.

Examples

What can Aura Workshop figure out?

Just describe what you want. Aura analyzes your intent, decides the best approach, and executes. Here are real examples.

01 Simple Code Question
"Write a Python function that reverses a string"

Classified as SINGLE. One agent responds directly with the code. No team creation overhead.

02 Full-Stack Application
"Build a REST API for a todo app with database schema, CRUD endpoints, and unit tests"

Classified as NEW. Creates a 5-agent Software Dev Team (PM, Architect, Developer, QA, DevOps). PM asks clarifying questions before development begins.

03 Vague Request
"Help me with my project"

Classified as CLARIFY. The agent asks what kind of project, what language, what you need help with. After your response, it re-classifies and routes appropriately.

04 Existing Team Routing
"Add a dark mode toggle to the settings page"

Classified as TEAM. Routes to the existing Software Dev Team. Workflow resumes with the existing roles and context.

05 Data Analysis Pipeline
"Analyze our Q4 sales data, generate charts, and write an executive summary"

Classified as NEW. Creates a Data Analysis Team with Data Engineer, Analyst, and Report Writer roles. Fan-out step for parallel chart generation.

06 Documentation Generation
"Generate API documentation for this Express.js router file"

Classified as SINGLE. One agent reads the file, generates OpenAPI-style documentation, and writes it to disk.

07 Multi-Service Architecture
"Design a microservices architecture for an e-commerce platform with auth, catalog, cart, and payment services"

Classified as NEW. Creates a team with Architect, multiple Developers (fan-out), and DevOps. Each service developed in parallel, then merged.

08 Code Review
"Review this pull request for security vulnerabilities and performance issues"

Classified as SINGLE. One agent with security expertise reviews the diff, flags issues, and suggests fixes.

09 Workflow with Fan-Out
"Run security audit, performance benchmark, and accessibility check on our web app simultaneously"

Classified as WORKFLOW. Three parallel fan-out tasks execute simultaneously, results merge into a unified report.

10 Research and Summarize
"Research the top 5 Rust web frameworks, compare their features, and recommend one for our project"

Classified as NEW. Creates a Research Team with Researcher and Analyst roles. Researcher gathers data, Analyst synthesizes and recommends.

11 Ambiguous Multi-Part Request
"Fix the bug and also maybe add some tests"

Classified as CLARIFY. Agent asks: Which bug? What file? What kind of tests? After clarification, re-classifies as TEAM or SINGLE depending on scope.

Workflow Engine

DAG-based execution with real-time control

The workflow engine executes tasks as a directed acyclic graph. Each node is an agent step with defined inputs, outputs, and dependencies. The engine handles orchestration so you don't have to.

  • DAG Execution — Steps execute in dependency order with automatic scheduling
  • Fan-Out — Parallel execution of independent tasks across multiple agents
  • Pause / Resume — Workflows pause when agents need human input, resume on reply
  • 30-Minute Timeout — Per-step timeout prevents runaway agents from burning tokens
  • Retry with Backoff — Failed steps retry with exponential backoff before marking failure
  • Visual Progress — Real-time workflow progress with per-step status indicators
  • Event Bus — Internal pub/sub system coordinates workflow state across components
Parallel Agents

Real concurrency with dependency-aware scheduling

Independent agents execute concurrently across the DAG. Two mechanisms drive parallelism: explicit fan-out (one node spawns agents over a list of work items) and intelligent auto-parallelism (the engine analyzes a step's output and spawns parallel agents when it finds independent items). Nodes advance only when all predecessors complete, with concurrency capped to keep your machine responsive.

  • Dependency-aware layers — The walker discovers entry nodes, spawns them in parallel, then advances as each predecessor completes.
  • Explicit + intelligent fan-out — Route a node to multiple successors with next: [...], or let the engine detect parallelizable items from a previous step's output.
  • Item extraction — The fan-out executor parses numbered lists, sections, and structured outputs to dispatch work to the right number of agents.
  • Wall-clock speedup — A 10-item fan-out with 8 concurrent agents completes in ~2 cycles instead of 10 sequential steps.
  • Crash-safe resume — Each agent step is checkpointed with an idempotency key. Restart and only the unfinished agents re-run.
Plan Mode

Review and approve before any tokens are spent

Plan Mode shows you the agent's complete approach — clarifying questions, step-by-step tasks with time estimates, and identified risks — before any work begins. Approve, send feedback, or reject and iterate. Execution only starts after explicit approval, giving you full control over scope and cost before the agent touches your codebase or burns a single token.

  • Two-phase workflow — Clarifying questions first, then a detailed plan. Execution never auto-starts.
  • Inspect before spend — Per-step time estimates and flagged risks, with totals computed automatically.
  • Targeted questions — Preset options or free-text fields to confirm scope, tech choices, and acceptance criteria.
  • Approve, revise, or reject — Multi-round iteration until the plan matches what you actually want.
  • Team-aware planning — Plans can route to a full team workflow with role assignments. Team execution only triggers after explicit approval.
Goal Mode

Specify outcomes. The agent iterates until they're met.

Goal Mode lets you describe what success looks like instead of dictating the steps. Specify success criteria, deliverables, and acceptance thresholds; the agent executes, then a verifier inspects the outcome against each criterion. On failure, it retries with guided feedback until every check passes or the iteration cap is hit.

  • Declare success upfront — Review extracted criteria and deliverables in a gate before execution begins.
  • Verify-and-retry loop — Every iteration is checked against the criteria. Failures trigger another pass with the verifier's feedback.
  • Human gate, autonomous refinement — Approval forces explicit intent alignment up front. After that, the loop runs on its own (or set goalAutoApprove to skip the gate).
  • Parallel to Plan Mode — One toggle switches the agent between Execute, Plan, and Goal. Pick the mode that matches the work.
  • Open-ended work — Built for tasks where you know the outcome but not the steps: "reduce API latency by 40% and prove it with a load test."
Desktop + Web Parity

One codebase. Identical experience.

Aura Workshop runs as a native desktop app on macOS, Windows, and Linux. Enable the embedded web server and access the exact same interface from any browser on your network.

  • Native Desktop — Tauri 2 app with native performance and OS integration
  • Web UI Server — Embedded axum server on port 18800 (configurable)
  • Identical Streaming — Real-time text streaming, tool calls, and workflow progress in both modes
  • Headless Linux — Run on a server with no display. Access from any browser.
  • Bearer Token Auth — Optional token-based authentication for the web server
  • Shared Core — Business logic in Rust core, thin wrappers for Tauri IPC and REST API
Desktop App
macOS / Windows / Linux
↔
Web UI
Any browser on your network
Same streaming. Same tools. Same workflow progress.
Pixel-perfect parity.
Billing Dashboard

Every token tracked. Every dollar accounted for.

Real-time cost tracking across all providers. Set daily and monthly limits. Auto-fallback when limits are hit. Know exactly what you're spending before the bill arrives.

  • Per-Provider Tracking — Individual expense tracking for each API provider
  • Per-Model Breakdown — Area charts showing token usage by model over time
  • Editable Pricing Table — Customize input/output token prices per model
  • Daily + Monthly Limits — Set spending caps per provider with auto-enforcement
  • Auto-Fallback Chain — DeepSeek → Anthropic → OpenAI → Local model
  • Zero-Cost Local — Local models via Aura Inference as final fallback ($0)
Monthly Usage March 2026
Total Expenses$12.34
$0$50 daily limit
DeepSeek$8.42 | 2.1M tokens
Anthropic$3.12 | 89K tokens
OpenAI$0.80 | 45K tokens
Local (Aura AI)$0.00 | 500K tokens
Skills System

Built-in skills for complex outputs

Agents can invoke specialized skills to generate rich outputs beyond plain text and code.

Mermaid Diagrams

Generate flowcharts, sequence diagrams, ERDs, and state machines rendered as SVG.

PDF Generation

WeasyPrint-powered PDF creation for reports, invoices, and documentation.

DOCX Export

Generate formatted Word documents with headings, tables, and styled content.

PPTX Presentations

Create slide decks with structured content, bullet points, and layouts.

XLSX Spreadsheets

Generate Excel files with data, formulas, and formatted worksheets.

Excalidraw Sketches

Hand-drawn style diagrams and wireframes for architecture and UI planning.

Supported Providers

Any model. Any provider. Your choice.

Aura Workshop is model-agnostic. Use any combination of providers and switch between them at any time.

Anthropic

Claude 3.5, Claude 4, Opus, Sonnet, Haiku

OpenAI

GPT-4o, GPT-4 Turbo, o1, o3, and more

DeepSeek

DeepSeek V3, DeepSeek Coder, R1

Google

Gemini 2.5 Pro, Gemini 2.5 Flash

Moonshot / Kimi

Long-context models for large codebases

Ollama

Run open-source models locally via Ollama

Aura Inference

Bundled engine for GGUF models on your hardware

OpenAI-Compatible

Any endpoint that speaks the OpenAI API format

Custom Endpoints

Point to any base URL with your own API key

Ready to let your agents self-organize?

Download Aura Workshop. Describe what you want. Watch a team of agents deliver it.

macOS (Apple Silicon) Windows Linux