The AI Application Server

Use major cloud providers, local Ollama/llama.cpp models, or any OpenAI-compatible endpoint through OneLLM.

Select models at the formation, agent, SOP, trigger, skill, or workflow-step level with aliases and fallback chains.

MUXI's orchestrator automatically breaks down complex requests into executable subtasks. No predefined workflows – agents analyze complexity, identify dependencies, and create execution plans on the fly.

Assess task requirements in real-time. Execute in optimal order with parallel processing when possible.

Package instructions, references, and scripts using the open SKILL.md specification with progressive disclosure.

Execute skill scripts and the bundled compute skill in an isolated RCE sandbox with resource and timeout controls.

Isolate sessions, memory, and credentials per user while sharing one formation across your product.

Group-based RBAC with inheritance and agent, MCP, and tool allow/deny rules keeps every request inside its permitted surface.

Delegate work across formations and external A2A services with capability-aware routing.

Secure inbound and outbound traffic with API keys, OAuth 2.0, HMAC, and OpenID authentication.

Connect large tool catalogs without injecting every schema into every prompt.

MUXI indexes definitions once, selects relevant tools at runtime, and filters the surface with layered allow/deny rules.

Store OAuth tokens and API keys per user with encryption at rest and automatic credential resolution.

One formation can access the same MCP tools on behalf of many users without crossing credential boundaries.

Connect any formation directly to Claude Desktop, Cursor, or another MCP client – no adapter code required.

The same agents, memory, tools, and authentication become a secure, AI-native tool surface.

Every memory write carries provenance in an immutable event stream, making memory explainable and rebuildable.

User, group, and formation scopes add selective forgetting, decay, contradiction handling, knowledge graphs, and Captain's Logs.

Embeddings run locally by default – no data leaves your infrastructure. FAISS-backed vector search powers memory, knowledge retrieval, and semantic routing.

Zero per-request embedding costs. Swap models without changing agent code. Your data stays yours, and your bill stays predictable.

Navigate hierarchical document trees instead of retrieving disconnected chunks, so agents can reason across large sources.

Combine vector, tree, and hybrid retrieval with synchronized HTTP, cloud-storage, rsync, FTP, and SFTP knowledge.

Handle images, PDFs, audio, and video natively. Vision models (GPT, Claude, Gemini) process screenshots, diagrams, and photos. Extract text from documents. Analyze charts and graphs.

OCR, image classification, and visual reasoning out of the box. Multimodal context flows through entire agent system.

Generate documents, spreadsheets, presentations, charts, code, and other downloadable outputs.

Every artifact is versioned, encrypted, user-scoped, and recallable in later conversations.

Let formations return typed options, actions, links, and MCP resources that applications render as interactive interfaces.

All 12 SDKs surface widgets idiomatically, while response text provides a universal fallback.

“Check email every hour” or “remind me tomorrow at 2pm” – agents understand natural time expressions. Recurring and one-time tasks with intelligent datetime parsing and timezone awareness. No cron jobs required.

Schedule once, runs everywhere.

Activate agents from schedules, webhooks, and external events through secure formation endpoints.

Transform results for Slack, Telegram, Discord, email, or custom webhooks with templates, authentication, retries, and fallbacks.

Combine SOP-guided execution with dynamic task decomposition, dependencies, and parallel steps.

Select models per workflow step and optionally replan when intermediate results change the best path forward.

Delegate coding tasks to Claude Code, Droid, OpenCode, Pi, or a custom headless adapter.

Jobs run asynchronously in isolated workspaces, while deterministic monitoring tracks long-running work without spending tokens on polling.

Let formations initiate contact through per-user notification channels instead of waiting for a prompt.

Heartbeats, active hours, soul documents, channel transformers, and built-in slash commands keep outreach useful and controlled.

Distill enterprise knowledge on-premises, then push only pre-processed memory through a signed ingestion channel.

Ed25519 authentication, scoped registrations, trust levels, revocation, and quotas keep every submission controlled.

Formations analyze operational evidence, detect patterns, and distill what they learn into candidate revisions of MUXI.md.

Review experiments and watched metrics, then inspect, apply, or dismiss every suggested behavioral change.

Define organizational knowledge that agents reason over, not rigid workflows they follow. Agents interpret procedures contextually and adapt them to specific situations. Version-control your organizational procedures.

Agents understand why procedures exist, not just what they say.

Official SDKs for Python, TypeScript, Go, Ruby, PHP, C#, Java, Kotlin, Swift, Dart, Rust, and C++.

Chat, stream, manage formations, parse UI widgets, and handle typed responses using conventions native to each language.

Trace model calls, tool use, memory, workflows, triggers, and lifecycle decisions through hundreds of typed events.

Regex and entity-based PII redaction runs before events reach pluggable monitoring and logging exporters.

Circuit breakers, exponential backoff, fallback models, and graceful degradation contain upstream failures.

Idempotency keys make chat, trigger execution, and scheduled-job creation safe to retry without duplicate work.

Review version-controlled formations through pull requests, validate them in CI, and promote drafts to production.

Updates and rollbacks preserve persistent memory plus MUXI.md and PENDING-MUXI.md self-tuning state.

muxi deploy validates, bundles, uploads, and starts your formation through the MUXI Server.

Develop against draft runtimes, update live formations, and roll back to the previous server-managed version.

Run the Go server as a binary or container on your own Linux, macOS, or Windows infrastructure.

Formation runtimes stay isolated through Apptainer/SIF on Linux and the Docker-based runtime runner on macOS and Windows.

Licensed under Elastic License 2.0 and Apache 2.0, with complete source on GitHub. Fork, modify, and use commercially without paying any licensing fees – no vendor permission required. Community-driven roadmap with transparent development.

No black boxes, no proprietary lock-in.

MUXI V1 coordinates stable contracts across the Server, Runtime, CLI, Formation API, and Server API.

All 12 SDKs share the same 1.x generation for consistent authentication, streaming, errors, idempotency, and UI widgets.