MachinaOS
v0.9 · 2,700+ Tests Passing

Neural Architecture

A layered runtime where intent intake, planning, policy, execution, and tool adapters are separated for reliability — now shipping as a native desktop application with a visual pipeline studio, multi-agent communication graph, and full MCP integration.

Core Runtime

Neural Indexing

Indexes active workspace state so plans can reference real files, tools, and recent execution history before any action is proposed.

Contextual Explorer

Cross-file semantic lookup feeds the planner with concrete evidence. Dynamic tool-name routing detects registered tools in natural-language queries.

Agent Command Center

Delegates subtasks to specialized agents with scoped permissions, approval requirements, capability scoring, fallback strategies, and five orchestration modes.

Neural Vault

Maintains local-first memory partitions, per-agent memory isolation, and execution logs so operators can audit what happened and reproduce any run.

Desktop Application

Native Desktop Shell — Windows, Linux, macOS

Tauri 2-based desktop app ships as a self-contained installer (MSI / NSIS / DEB / AppImage / DMG) with bundled Python runtime and all dependencies. Zero prerequisites on the target machine. Frameless glassmorphic window with custom titlebar, minimize/maximize/close controls, and system-tray hide-to-background support.

Auto-Updater

Built-in updater plugin checks GitHub Releases on startup, displays an install banner, and applies updates with a single click — no manual download required.

14 Integrated Views

Dashboard, Chat, Tools, Memory, Workflows, Studio, Agents, Timeline, Explorer, Processes, Metrics, Events, Settings, Security, MCP — all accessible via keyboard shortcuts Ctrl+1–9 and a fuzzy Command Palette (Ctrl+K).

System Status Bar

Persistent bottom bar shows LLM connection state, agent fleet health, active task count, memory usage, and GitHub connection status — refreshed every 30 seconds.

Visual Workflow Studio

Drag-and-Drop Pipeline Canvas

Build multi-agent pipelines visually on an SVG canvas with a grid background. Drag tools and agents from a searchable palette, reposition nodes freely, connect output ports to input ports with cubic Bezier edges, and load any of 35 built-in chain templates directly onto the canvas. Mouse-wheel zoom (15–200%), pan toggle, fit-view, and shift-drag pan keep large pipelines navigable.

Smart Argument Builder

Per-tool argument schemas replace raw JSON with structured form fields — text, number, textarea, checkbox, and select — covering 30+ native tools with placeholder text, required indicators, and inline hints. MCP tools generate fields dynamically from their ToolSpec.

Studio Debugger — Step-by-Step Execution

Server-side debug sessions with a Step Over / Continue / Stop toolbar. Each canvas node animates in real time: amber pulsing border with spinner while running, green glow with ✓ badge on success, red glow with ✗ on failure, dim ⊘ for skipped. Per-step duration metrics stream into a resizable results tray.

Chain-to-Chain Composition

Treat any saved workflow chain as a reusable building block. Drag chain nodes (rose-colored, link icon) from the Studio palette, embed them in larger pipelines, and execute recursively with depth limiting (max 5) and cycle detection. Composite templates like Master Audit and Daily Ops bundle three sub-chains in one click.

Auto-Assign and Optimize

One-click pipeline negotiation assigns the best-scoring agent to each step by capability confidence. The optimizer detects data-flow dependencies and annotates steps with parallel execution groups for concurrent execution.

Chain Versioning and Rollback

Every workflow update snapshots the previous version. Roll back to any prior configuration, compare version history, and republish — all without leaving the Studio view. Rollback itself creates a new version, making it non-destructive.

Pipeline Lane View

Horizontal lanes per agent show every step assigned to that agent as a draggable badge. Move steps across lanes to reassign agents — the chain auto-saves and increments the version. Conflict resolution lets you pick a winner when capability scores tie.

Security Center

Encrypted Secrets Vault

AES-256-GCM encryption with PBKDF2 key derivation (600k iterations) and per-secret nonces. Falls back to XOR+HMAC when the cryptography package is unavailable. Store and retrieve credentials without ever writing them to disk in plain text.

Immutable Audit Log

SOC 2-compatible append-only audit trail with SHA-256 hash chains. Query by date, export as JSON, verify chain integrity, and review summaries from the Security Center panel.

Vulnerability Scanner

14 built-in patterns covering SQL injection, XSS, SSRF, command injection, path traversal, hardcoded secrets, insecure deserialization, JWT issues, and eval/exec — each finding tagged with a CWE ID and severity level.

Credential Scrubber

20+ regex patterns strip OpenAI, GitHub, GitLab, Slack, Google, AWS, and Azure keys; JWTs; bearer tokens; private keys; and connection strings from tool outputs before they reach the log or UI.

Agent Orchestration

Five Orchestration Strategies

Single, Round-Robin, Capability Score, Fallback, and Pipeline — selectable per-workflow with configurable consensus thresholds and min-voter requirements.

Capability and Reputation Learning

Every tool execution records success rate, latency, and error counts per agent. Reputation scores blend 60% recent with 40% lifetime outcomes, weighting selection toward proven performers.

Agent Health and Auto-Healing

Rolling health records track consecutive failures, recovery thresholds, and degradation onset. A 60-second background sweeper auto-heals DEGRADED or OFFLINE agents and reroutes in-flight tasks.

Consensus and Negotiation

Multi-agent consensus gates risky tool calls (Majority, Unanimous, Weighted). Multi-round proposal/counter/accept lifecycle with auto-counter heuristics and priority-based resolution.

Delegation Chains and Cycle Detection

Scoped delegation with max-depth-5 guards and cycle detection prevents runaway chains. Chain history is visualized with depth indicators directly in the Agents view.

Task Handoffs and Queues

Formal handoff lifecycle (REQUESTED → EXECUTING → COMPLETED/DECLINED) with preconditions and constraints. Priority-sorted task queue with auto-execution, batch drain, and broadcast timeout sweeper.

Agent Communication Graph

The Communications Dashboard provides a read-only monitoring view of all inter-agent traffic across six channels — bus messages, point-to-point requests, broadcast queries, negotiations, handoffs, and collaboration sessions. A live SVG graph renders a circular layout of agent nodes with color-coded directed edges (cyan = requests, amber = negotiations, purple = handoffs, emerald = broadcasts) and edge labels showing message counts. A unified activity feed lists every event chronologically with type filters, and Threads view groups related messages by correlation_id into collapsible cards.

LLM and Tool Connectivity

7 LLM Providers

Ollama (local), OpenAI, Google Gemini, Anthropic Claude, OpenRouter, LM Studio — each with provider-specific headers, split connect/read timeouts, and diagnostic error messages distinguishing connection refusal, timeout, and auth failure. Hot-reconfigure without restart.

Enhanced Neural Link

Conversational command interface with streaming plan cards, RAG-augmented context (ChromaDB vector store), and natural-language delegation. Type intent like delegate to security-auditor scan this project and the planner routes to the right specialist agent. Plan cards expand inline with per-step approvals, tool outputs, and timing.

MCP Integration — Client and Server

MachinaOS operates in a dual MCP role. As a client, it consumes external Model Context Protocol servers via JSON-RPC 2.0 stdio or SSE transport — 5 default servers ship pre-configured (sequential-thinking, memory, fetch, github, sqlite), 3 auto-connect on startup, and a smart registration builder with 14 curated quick-pick packages auto-fills configuration as you type. Discovered tools are namespaced as mcp.<server>.<tool> and behave as first-class entries in the tool registry, usable from chat, Studio, and chains. As a server, MachinaOS exposes all 52 native tools, 4 virtual task tools, 17 resources across 7 domains, and 4 prompt templates to external MCP clients like Claude Desktop and Cursor over an SSE transport at /mcp/serve/sse.

GitHub Integration

Dedicated PAT-based GitHub connection panel, independent of the LLM provider. Validates against api.github.com/user, shows username and connection status dot in the system bar, and persists token in preferences.

40 Native Tools · 7 Domains

filesystem (9), git (10), shell (2), process (4), browser (3), vscode (7), system (4) plus filesystem.copy, filesystem.mkdir, git.tag, git.merge, and system.env with workspace-bounded path safety, overwrite guards, and automatic sensitive-key redaction on env reads.

Template Library

35 Chain Templates

Ready-to-run workflow blueprints: code-review-deep, debug-investigation, security-deep-scan, hotfix-workflow, pr-readiness-check, changelog-generation, performance-profile, git-hygiene, license-compliance, and 26 more.

21 Agent Templates

Preconfigured agent blueprints with scoped tool-filters: code-reviewer, devops, security-auditor, release-manager, debugger, test-engineer, documentation-writer, ci-cd-specialist, frontend-analyst, sre-agent, and more.

24 Workflow DSL Templates

Concise intent-to-step recipes: quick-commit, branch-and-start, full-project-audit, security-sweep, debug-workflow, release-cut, hotfix, feature-start, and 16 others — load any into the Studio or run directly.

Observability and CI/CD

Metrics Dashboard and Event Inspector

Plan success rate, average step count, execution duration, and event count displayed as cards and SVG bar charts. Filterable event stream by type, source, and severity with pagination and newest-first ordering.

GitHub Actions CI/CD Pipeline

Automated cross-platform release builds for Windows (MSI+NSIS), Linux (DEB+AppImage), and macOS (DMG x64+arm64) triggered on version tags. Ruff lint, Rust fmt, and cargo check on every push. Auto-generated update manifest uploaded as a draft release asset.

Execution Principles

No Fake Magic

Every claimed action must map to a tool invocation result, not model-only text.

Human Approval on Risk

Risky operations such as destructive writes or remote actions are gated by policy and explicit approval.

Graceful Degradation

If model services degrade, state, logs, and replayable execution context remain available locally.

Initialize protocol

Use the demo guide to walk through setup and verify your first local orchestration flow, or explore real-world workflows to see how engineering teams use MachinaOS in practice. The interface gallery shows every panel live.

Open Demo Guide

Frequently Asked Questions

What does "intent-native" mean?

Intent-native means the system is designed from the ground up to accept natural-language goals and convert them into structured, validated plans — not just pass prompts to a model. MachinaOS has a dedicated planning layer that decomposes goals into steps, assigns tools, checks policies, and records outcomes as first-class runtime objects.

How is MachinaOS different from LangChain, AutoGen, or CrewAI?

LangChain, AutoGen, and CrewAI are frameworks and libraries — you build orchestration logic in code. MachinaOS is a runtime layer: it ships as a desktop application with a native UI, visual workflow studio, security center, and audit log. You interact with it through natural language or the visual canvas, not YAML files or Python glue code. It also runs fully locally with no cloud infrastructure dependency.

Does MachinaOS run fully locally?

Yes. The runtime, planner, tool registry, memory system, secrets vault, and audit log all run on your local machine. You can use Ollama as a fully local LLM provider with zero network calls. Cloud LLM providers (OpenAI, Gemini, Claude, OpenRouter, LM Studio) are opt-in — the system does not require any cloud infrastructure to operate.

Which LLM providers does MachinaOS support?

MachinaOS supports seven LLM providers: Ollama (fully local), OpenAI, Google Gemini, Anthropic Claude, OpenRouter, LM Studio, and a configurable custom endpoint. You can hot-reconfigure the active provider from the Settings panel without restarting the runtime.

What is the policy gate and approval system?

The policy gate evaluates every proposed step against a risk profile before execution. Destructive operations (file deletion, remote calls, high-impact system actions) trigger an approval requirement — the runtime pauses and presents the plan for human review. You can approve, deny, modify, or request an explanation for any step. Approval decisions are recorded in the immutable audit log.

Is MachinaOS open source?

MachinaOS is currently in pre-release and the repository is private. The full source — runtime, policy engine, hash-chained audit log, UI, and the 2,700+ test suite — is available for review to serious evaluators (investors, security reviewers, prospective partners) on request: reach out via the contact page. Licensing and open-source plans will be announced with the v1.0 public launch.