AgentX docs

AgentX Documentation

AgentX is a self-hosted, glassbox AI agent platform: persistent multi-type memory, Agent Teams that delegate, six thinking patterns, MCP tools, and your choice of model — all behind a Django REST API you run yourself, with every step of the agent loop observable.

Current release: v0.21.222 (the v0.20 milestone — “Mobile-Ready Alpha”).

Architecture at a Glance

graph TB
    Client[Tauri Client<br/>React 19 + Vite]

    subgraph API["Django API (port 12319)"]
        Agent[Agent Core<br/>planner · session · context]
        Reasoning[Reasoning<br/>CoT · ToT · ReAct · Reflection]
        Drafting[Drafting<br/>speculative · pipeline · candidate]
        Providers[Providers<br/>LM Studio · Anthropic · OpenAI · OpenRouter · Vercel]
        MCP[MCP Client<br/>stdio · SSE · HTTP]
        Prompts[Prompt System<br/>profiles · composition]
        Translation[Translation Kit<br/>NLLB-200 · 200+ languages]
        Memory[Agent Memory<br/>episodic · semantic · procedural · working]
    end

    subgraph Data["Data Layer (Docker)"]
        Neo4j[Neo4j<br/>entity graphs]
        Postgres[PostgreSQL + pgvector<br/>vectors · episodic · audit]
        Redis[Redis<br/>working memory cache]
    end

    Client -->|HTTP| API
    Agent --> Reasoning & Drafting & Providers & MCP & Prompts & Memory
    Memory --> Neo4j & Postgres & Redis
    MCP -->|stdio/SSE| ExtServers[External MCP Servers]

Key Features

FeatureDescriptionDocs
Agent ChatConversational AI with streaming, tool use, and session managementChat
ReasoningThinking patterns in chat + an offline CoT/ToT/ReAct/Reflection kit, plus multi-model draftingReasoning
MCP ClientConnect to external tool servers via stdio, SSE, or HTTPMCP
Agent TeamsA lead delegates subtasks to specialist teammatesAgent Teams
ProvidersUnified interface for LM Studio, Anthropic, OpenAI, OpenRouter, VercelProviders
PromptsProfile-based prompt composition with global prompt layerPrompts
Memory4-type persistent memory with recall, extraction, and consolidationMemory
TranslationTwo-level detection + NLLB-200 translation for 200+ languagesTranslation
Quick Start — Install and run AgentX in minutesAPI Reference — All REST API endpoints with examples
Architecture — System design, module layout, request lifecycleDevelopment — Setup, contributing, and testing
Database Stack — Neo4j, PostgreSQL + pgvector, RedisRoadmap — Development history and future plans

Technology Stack

LayerTechnologyPurpose
FrontendTauri v2 + React 19Desktop application shell
BuildVite + TypeScriptFast development and bundling
BackendDjango 5.2REST API framework
AI/MLHuggingFace Transformers · sentence-transformersTranslation (NLLB-200) + local embeddings
Graph DBNeo4j 5.15Entity relationships and knowledge graphs
Vector DBPostgreSQL + pgvectorSemantic search and episodic memory
CacheRedis 7Working memory and session state
Task RunnerTask (Taskfile)Development automation
PythonuvFast dependency management
ClientbunClient package management

Project Status

Completed (Phases 1-14, 17):

  • Django API with 70+ REST endpoints (chat, agent, memory, MCP, providers, prompts, profiles, jobs, auth, multi-agent, config)
  • Tauri desktop app: multi-page layout, browser-style conversation tabs, drawer panels, agent profiles
  • Mobile-ready: Tauri v2 Android target (v0.20.0 — “Mobile-Ready Alpha”)
  • Two-level translation system (200+ languages)
  • Database stack (Neo4j, PostgreSQL + pgvector, Redis)
  • MCP client with stdio/SSE/HTTP transports
  • Model provider abstraction (LM Studio, Anthropic, OpenAI, OpenRouter, Vercel AI Gateway)
  • Drafting framework (speculative, pipeline, candidate)
  • Reasoning framework (CoT, ToT, ReAct, Reflection)
  • Agent core with task planning and goal tracking
  • Memory system: 4 types, recall layer (5 techniques), extraction pipeline, consolidation
  • Context gating: task-aware compression, intent-based retrieval, trajectory compression
  • Agent identity: Docker-style IDs, self-memory channels, assistant self-extraction
  • Three-layer fact verification pipeline (hash → semantic → LLM adjudication)
  • Server management (Phase 17): optional session auth, Docker production stack, multi-cluster deploy, client/API version matching
  • 190+ backend tests

In Progress:

  • Phase 15: Plan execution (Core Complete — executor, Redis-tracked state, streamed progress, cancellation)
  • Phase 16: Multi-agent — Agent Teams v1 + routing, ad-hoc delegation, and @-mention routing shipped (~65%)
  • Phase 18: UX improvements + memory tuning (~93%)

See the Roadmap for detailed phase history.

License

This project is licensed under the MIT License.