AutomatosX – Multi-agent framework with persistent memory for developers

1 month ago 4

Provider-Agnostic AI Agent Orchestration

A CLI-first tool for orchestrating specialized AI agents with persistent memory, intelligent delegation, and cross-provider support (Claude, Gemini, OpenAI).

Status: ✅ Production Ready · v5.1.0 · October 2025

Looking for answers? See the FAQ.

AutomatosX extends Claude Code with specialized AI agents that remember context, delegate tasks, and collaborate autonomously.

# In Claude Code, simply use /ax:agent /ax:agent Paris, design authentication system with JWT /ax:agent Bob, implement the auth design # Bob auto-receives Paris's design from memory

The result: Claude Code becomes a learning, coordinated team instead of a stateless assistant.

The Problem with Stateless AI

Traditional AI assistants (ChatGPT, vanilla Claude):

❌ No memory between conversations
❌ You repeat context every time
❌ No coordination between tasks
❌ Knowledge disappears after each session

AutomatosX changes this:

✅ Persistent memory: < 1ms search, zero cost, 100% local
✅ Multi-agent delegation: Agents coordinate automatically
✅ Context retention: Never explain the same thing twice
✅ Knowledge accumulation: Your team gets smarter over time

Without AutomatosX:

Day 1: You explain architecture to Claude → Response lost Day 2: You ask to implement → You re-explain architecture Day 3: Different task → You re-explain everything again

With AutomatosX:

Day 1: Product designs architecture → Saved to memory Day 2: ax run backend "implement auth" → Backend finds Product's design automatically Day 3: ax run security "security audit" → Security has full context from Day 1-2

Time saved: Hours per week. Quality: Consistent. Cost: $0.

📋 Full Changelog | 🎉 Release Notes

🧠 Core Value: Persistent Memory

AutomatosX remembers everything. Every agent conversation is automatically saved and searchable.

# Automatic memory saving ax run product "Design calculator with add/subtract" → Task + Response saved to SQLite FTS5 # Automatic memory retrieval ax run backend "Implement the calculator" → Memory searches "calculator" automatically → Backend receives: "# Relevant Context from Memory: Product's design..." → Backend implements WITHOUT you repeating the spec

SQLite FTS5: Built-in full-text search
< 1ms search: 62x faster than v3.x vector search
$0 cost: No embedding APIs, no cloud calls
100% local: Your data never leaves your machine
Automatic injection: Relevant context added to every agent

✅ Cross-day continuity: Pick up where you left off ✅ Cross-agent knowledge: All agents share the same knowledge base ✅ Learning from history: Agents avoid past mistakes ✅ Zero cost scaling: 10,000 entries = ~10MB, still < 1ms search

Learn more: Memory System Guide | Memory Tutorial

🤝 Core Value: Multi-Agent Orchestration

Agents coordinate automatically. Natural language delegation creates complex workflows without manual orchestration.

// Product Manager analyzes and delegates ax run product "Build authentication feature" Product response: "I'll design the auth system with JWT + OAuth2. @backend Please implement the JWT authentication API based on this design. @security Please audit the implementation for security issues." // AutomatosX automatically: // 1. Backend receives full spec, implements code // 2. Security receives spec + code, performs audit // 3. Results aggregated back to Product

7 delegation syntaxes: @mention, DELEGATE TO, Please ask, etc.
Cycle detection: Prevents infinite loops
Depth limits: Default 2 levels (configurable)
Session tracking: Who did what, when
Workspace isolation: No file collisions

✅ Automatic coordination: No manual task switching ✅ Parallel execution: Multiple agents work simultaneously ✅ Transparent workflows: Full delegation chain visible ✅ Context preservation: Every agent has complete context

Learn more: Multi-Agent Orchestration Guide

🎭 12 Specialized Agents with Clear Governance

v5.0.12 introduces strict role ownership and delegation controls to eliminate cycles:

💻 Engineering Team (Implementers)

maxDelegationDepth: 1 - Can delegate once for cross-domain needs, no re-delegation

Bob (backend) - API design, database modeling, caching strategies
- Can delegate to: frontend, data, security, quality, devops
Frank (frontend) - Component architecture, state management, accessibility
- Can delegate to: backend, design, security, quality, devops
Oliver (devops) - Infrastructure as code, CI/CD pipelines, observability
- Can delegate to: backend, frontend, security, quality
Daisy (data) - Data modeling, ETL pipelines, SQL optimization
- Can delegate to: backend, security, quality
Steve (security) - Sole owner of security-audit, threat modeling, secure coding review
- Can delegate to: backend, frontend, devops, quality

🎯 Quality Team (Coordinator Role)

maxDelegationDepth: 1 - Can delegate fixes back to implementers, no re-delegation

Queenie (quality) - Sole owner of code-review and debugging, testing strategies
- Can delegate to: backend, frontend, security, devops, data

🎨 Content Team (Implementers)

maxDelegationDepth: 1 - Can delegate once for cross-domain needs, no re-delegation

Debbee (design) - UX research, wireframes, design systems
- Can delegate to: frontend, writer, quality
Wendy (writer) - API documentation, ADRs, release notes
- Can delegate to: backend, frontend, design, quality

📊 Leadership Team (Coordinators)

maxDelegationDepth: 1 - Delegate to implementers, focus on strategy, no re-delegation

Paris (product) - Product strategy, feature planning, roadmap
- Can delegate to: backend, frontend, design, writer, quality
Eric (ceo) - Business strategy, organizational leadership
- Can delegate to: paris, tony, all agents
Tony (cto) - Technology strategy, technical leadership
- Can delegate to: backend, frontend, devops, security, quality

🔬 Research Team (Specialist)

maxDelegationDepth: 0 - Execute research work directly, no delegation

Rodman (researcher) - Idea validation, feasibility analysis, research reports
- Specializes in: logical reasoning, risk assessment, literature review
- Produces: executive summaries, feasibility studies, long-form research reports

New in v5.0.12: Each agent has role-specific workflow stages, smart ability loading (abilitySelection), and explicit delegation scopes. Most agents have maxDelegationDepth: 1 to allow cross-domain collaboration while preventing delegation cycles.

📖 Complete Agent Directory

🚀 Two Ways to Use AutomatosX

AutomatosX offers two powerful modes to fit your workflow:

1️⃣ Claude Code Integration (Recommended)

Use AutomatosX agents directly inside Claude Code conversations with the /ax:agent slash command.

# In Claude Code, use the slash command /ax:agent Paris, design a REST API for user authentication /ax:agent Bob, implement the auth API from Paris's design /ax:agent Steve, security audit the authentication code

Perfect for:

💬 Interactive development workflows
🔄 Seamless context switching within Claude Code
🤝 Collaborative coding sessions
🎯 Quick agent delegation while coding

How it works: Claude Code executes AutomatosX commands behind the scenes, brings results back into your conversation, and maintains full context.

2️⃣ Terminal/CLI Mode (Power Users)

Use AutomatosX as a standalone CLI tool for automation, scripting, and direct control.

# In any terminal (Bash, Zsh, PowerShell) ax run Paris "Design REST API for user authentication" ax run Bob "Implement the auth API" # Auto-receives Paris's design from memory ax run Steve "Security audit the auth code" # Auto-receives design + implementation # Full CLI power ax memory search "authentication" ax agent list --by-team engineering ax session list --active

Perfect for:

⚙️ CI/CD pipelines and automation scripts
🔧 Custom workflows and integrations
📊 Batch processing and reporting
🎛️ Advanced configuration and debugging

How it works: Direct command-line execution with full control over providers, memory, sessions, and configuration.

Scenario Recommended Mode

Coding in Claude Code	Claude Code Integration (/ax:agent)
Automation scripts	Terminal Mode (ax run)
CI/CD pipelines	Terminal Mode
Quick questions during dev	Claude Code Integration
Memory management	Terminal Mode
Agent creation/management	Terminal Mode
Multi-agent workflows	Both work great!

Using Terminal Mode? → Complete Terminal Mode Guide
Using Claude Code? → Continue reading below for slash command examples
Want both? → They work together seamlessly! Memory is shared across both modes.

npm install -g @defai.digital/automatosx

# In Claude Code, initialize (first time only) # Ask Claude: "Please run ax init" # Or use terminal mode: ax init # Then use the slash command for agents /ax:agent Paris, design REST API for users /ax:agent Bob, implement the API /ax:agent Queenie, write tests for the API

What happens:

Claude Code executes AutomatosX behind the scenes
Paris designs the API → Saved to memory
Bob reads Paris's design from memory → Implements code
Queenie reads everything → Writes comprehensive tests
Results flow back into your Claude Code conversation

# Initialize your project ax init # Run agents from any terminal ax run Paris "Design REST API for users" ax run Bob "Implement the API" # Auto-receives Paris's design from memory ax run Queenie "Write tests for the API" # Auto-receives design + implementation # Manage memory and agents ax memory search "API design" ax agent list --by-team engineering ax agent create myagent --template developer

That's it! Agents now remember everything and coordinate automatically across both modes.

MCP Server Mode (Advanced) ✨ NEW in v5.1.0

Use AutomatosX as a native MCP server for direct Claude Code integration via Model Context Protocol.

# Start MCP server ax mcp # Add to Claude Code's claude_desktop_config.json { "mcpServers": { "automatosx": { "command": "ax", "args": ["mcp"] } } }

What you get:

✅ 16 native MCP tools for Claude Code
✅ 90% faster than CLI execution (shared state, < 300ms p50 latency)
✅ Persistent services across requests (MemoryManager, SessionManager)
✅ First-class integration with Claude Desktop

Available MCP Tools:

Agent execution: run_agent, list_agents
Memory operations: search_memory, memory_add, memory_list, memory_delete, memory_export, memory_import, memory_stats, memory_clear
Session management: session_create, session_list, session_status, session_complete, session_fail
System info: get_status

Performance:

No subprocess overhead (3-5s → 300ms)
< 1.5s cold start
Shared services across requests
Native JSON-RPC 2.0 protocol

📖 Terminal Mode Guide | Installation Guide | Quick Start Tutorial

Quick Start Guide - Get up and running in 5 minutes
Terminal Mode Guide - Complete CLI usage tutorial
Core Concepts - Understand agents, memory, providers
Installation Guide - Detailed setup instructions
FAQ - Common questions and troubleshooting

Agent Communication & Memory - How agents communicate and remember
Multi-Agent Orchestration - Natural language delegation
Team Configuration - Team-based agent organization
Agent Templates - Quick agent creation

Memory Management - Hands-on memory system guide
Creating Your First Agent - Build custom agents

CLI Commands - Complete command reference
Agent Directory - All available agents

🔬 The Technical Advantage

Feature Traditional AI Chat Claude Code Claude Code + AutomatosX

Memory	No	No	✅ SQLite FTS5 (< 1ms)
Cost	$20/month	Included	✅ $0 (100% local)
Multi-Agent	No	No	✅ 12 specialized agents
Coordination	Manual	Manual	✅ Automatic delegation
Context Retention	Copy-paste	Session only	✅ Persistent (days/weeks)
Knowledge Sharing	No	No	✅ Cross-agent memory
Privacy	Cloud	Claude servers	✅ 100% local data
Speed	Web UI	Terminal	✅ Instant CLI

# Using friendly agent names in terminal ax run Paris "Design user authentication feature" # Paris creates spec → Saved to memory ax run Bob "Implement auth based on spec" # Bob auto-receives spec → Implements code ax run Steve "Security audit the auth implementation" # Steve auto-receives spec + code → Performs audit ax run Wendy "Document the auth system" # Wendy auto-receives everything → Creates docs # Or in Claude Code: # /ax:agent Paris, design user authentication feature # /ax:agent Bob, implement auth based on spec

Result: 4-step workflow, zero context re-explanation, complete audit trail

# Mix of names and roles (both work!) ax run Queenie "Debug the payment timeout issue" # Queenie analyzes, saves findings to memory ax run backend "Fix the issue Queenie found" # Backend reads Queenie's analysis → Implements fix ax run quality "Test the payment fix" # Quality knows the bug + fix → Comprehensive testing

Result: Coordinated debugging with full context preservation

# Using agent names for clarity ax run Daisy "Analyze user behavior patterns" # Daisy analyzes patterns → Findings in memory ax run Paris "Design features based on Daisy's analysis" # Paris reads analysis → Creates product spec ax run Eric "Business case for Paris's proposal" # Eric has analysis + spec → Strategic evaluation

Result: Data-driven decision making with complete context

# Single command triggers automatic multi-agent coordination ax run Paris "Build a user dashboard with real-time metrics" # Paris analyzes and delegates automatically in its response: # "I've designed the dashboard architecture: # # @Bob Please implement the REST API endpoints for user metrics # @Frank Please create the React dashboard components # @Steve Please review the data access security # # All specs are in my workspace." # AutomatosX automatically: # ✓ Bob implements backend API → Saves to workspace # ✓ Frank builds frontend UI → Reads Bob's API spec # ✓ Steve audits security → Reviews both implementations # ✓ Results aggregated → Complete dashboard delivered

Result: One command orchestrates 4 agents with automatic coordination

Delegation Syntaxes:

@Bob Please implement this # Direct mention DELEGATE TO Frank: Create the UI # Explicit syntax Please ask Steve to audit this # Polite request I need Daisy to analyze the data # Need expression

🎯 Why Teams Choose AutomatosX

Extend Claude Code with persistent memory
Never repeat context - agents remember everything
Coordinate complex tasks with multi-agent workflows
100% local - your data stays private

Shared knowledge base - export/import memory across team
Consistent quality - agents learn from past work
Faster onboarding - new members inherit team knowledge
Audit trail - complete history of all decisions

For Claude Code Power Users

Slash command integration - /ax:agent for instant access in Claude Code
Terminal-native - no context switching
CLI-based - scriptable and automatable
Zero latency - local memory = instant search

✅ 1,136 tests passing (100% pass rate) ✅ TypeScript strict mode (zero errors) ✅ ~56% test coverage (comprehensive testing) ✅ 458KB bundle (99.9% smaller than v3.x) ✅ < 1ms memory search (62x faster than v3.x)

Memory Search: < 1ms (10,000 entries) Bundle Size: 458KB (down from 340MB in v3.x) Dependencies: 19 packages (down from 589 in v3.x) Test Coverage: ~56% (1,136 tests passing, 100% pass rate) Memory Cost: $0 (no API calls)

Runtime: Node.js 20+
Language: TypeScript 5.3 (strict mode)
Memory: SQLite + FTS5 (built-in full-text search)
Testing: Vitest 2.x (1,136 tests)
Build: tsup/esbuild
Providers: Claude CLI, Gemini CLI, Codex CLI (OpenAI)

Enhanced Claude Code integration
Visual workflow builder
Advanced memory analytics
Cross-project knowledge sharing
Plugin system for custom providers

We welcome contributions! AutomatosX is built in the open.

Contributing Guide - How to contribute
Development Setup - Local setup
Architecture Guide - Understand the codebase

Join the community:

🐛 Report Issues
💡 Feature Requests

AutomatosX is Apache 2.0 licensed.

📦 npm: @defai.digital/automatosx
🐙 GitHub: defai-digital/automatosx
📖 Documentation: docs/
❓ FAQ: FAQ.md
🎉 Releases: GitHub Releases
📋 Changelog: CHANGELOG.md

Transform Claude Code into an intelligent, coordinated team with AutomatosX. 🚀

Built with ❤️ by the AutomatosX team

Read Entire Article