Open-Source · Self-Hosted · Apache 2.0

Local Multi-Agent Chat UI &
Event-Driven AI Agent Runtime

Run a collaborative team of AI agents on your own Mac or Linux machine. Bring multiple specialized clones into one group chat with Ollama, vLLM, OpenAI, Anthropic Claude, or Google Gemini — complete with persistent memory, workspace file tools, Model Context Protocol (MCP), and local SDXL image generation.

$curl -fsSL https://raw.githubusercontent.com/UClone-AI/uclone-x/main/install.sh | bash

Or install from PyPI with pip install "uclone-x[cli,http]" or uv tool install --managed-python --python 3.12 "uclone-x[cli,http]", then run ucx start.

View on GitHub Read Documentation PyPI Package
Python 3.11 – 3.13 macOS & Linux Ollama · vLLM · Gemini · Claude · OpenAI MCP · A2A · ACP
No-Code Multi-Agent Workspace

One timeline. Six specialized AI agents. 100% local privacy.

Ask scout to research, writer to draft, guardian to review risks, and artist to illustrate — all inside one shared conversation room, with your files staying on your machine.

UClone-X local multi-agent chat UI running a research task on Ollama qwen3:8b
UClone-X v0.4.1 web dashboard (`http://127.0.0.1:5180`) running on a local Ollama model (`qwen3:8b`) and generating images in-process.
clonedefault collaborator
scoutweb & file research
writerlong-form drafting & stories
artistlocal SDXL image generation
guardiancritical review & risk analysis
pioneerarchitecture & strategy

Multi-Agent Group Chat UI

Bring multiple clones into one conversation room. Use @mention to route the next turn to a specific specialist, or turn on Auto discuss so agents debate and refine answers together.

Local Ollama & Cloud LLMs

Run 100% offline with Ollama or vLLM with zero API cost, or connect OpenAI, Anthropic Claude, and Google Gemini keys from the Settings UI.

Model Context Protocol (MCP)

Connect local stdio or remote HTTP MCP servers directly from the dashboard Settings, or paste any standard mcpServers JSON configuration.

Persistent Memory & Ontology

Every clone records, recalls, and retracts facts across conversations, backed by ranked embedding/lexical retrieval and a LinkML-validated knowledge graph.

In-Process SDXL Image Engine

Generate illustrations locally on Apple Silicon (MPS), NVIDIA CUDA, or CPU via diffusers without configuring a separate server, or connect a running ComfyUI daemon.

Event-Driven Core, A2A & ACP

Built on a non-blocking 6-stage reactive state machine over an in-memory priority event bus, with HTTP+JSON A2A (ucx a2a serve) and stdio ACP (ucx acp serve) servers.

Ecosystem Comparison

How UClone-X compares to other local AI & agent tools

Most open-source AI tools force a choice between a single-model chat interface and a code-only Python SDK. UClone-X gives you both a ready-to-use multi-agent workspace and a modular event-driven runtime.

Capability
Single-Agent Local UIs
(Open WebUI, AnythingLLM)
Code Frameworks
(AutoGen, CrewAI, LangGraph)
UClone-X
Out-of-the-box Web UI & CLI
Web UI included
No (requires custom frontend)
Yes — React 19 dashboard (ucx start) & terminal REPL (ucx run)
Multi-agent group chat rooms
Single active model per turn
Programmatic scripts only
Built-in rooms with @mention routing & autonomous multi-agent discussion
Per-agent memory & ontology
Shared document RAG
Bring-your-own storage
Per-clone persistent memory + LinkML semantic knowledge graph
Local image generation
Requires external server
Requires custom tool code
Built-in in-process SDXL engine + automatic ComfyUI detection
Agent protocols (MCP, A2A, ACP)
MCP client in some
Library adapters
Native MCP client + A2A HTTP server + ACP editor server
For Developers & Integrators

An event-driven Python 3.11+ agent runtime under the hood.

Every surface — the React web dashboard, the ucx CLI, the multi-agent room orchestrator, and the A2A/ACP protocol servers — sits cleanly over one strictly typed kernel.

Non-Blocking Reactive State Machine

Agents transition cleanly across IDLE → INGESTING → REASONING ⇄ CALLING_TOOL → EMITTING_RESPONSE over an in-memory priority event bus with backpressure policies.

Workspace Sandbox & Dynamic Skills

File and shell tools are confined to your workspace directory by default. Agents can load, synthesize, audit, and promote reusable SKILL.md skill packages.

OpenTelemetry Native

Built-in tracing, token accounting, context compaction, and OTLP export so you can inspect every LLM call and tool invocation in detail.

Frequently Asked Questions

Common questions about UClone-X

Can I run multiple Ollama agents in one group chat?

Yes. Create a group conversation in the dashboard (or run ucx room in your terminal) and add any combination of built-in or custom clones. Mention a clone with @name or enable Auto discuss so they collaborate on your local Ollama model.

Does UClone-X work completely offline without an API key?

Yes. When paired with Ollama or vLLM and a local SDXL checkpoint, every conversation, memory record, and generated image stays 100% on your computer with zero cloud dependency.

What hardware is recommended for local models?

For cloud API keys (Gemini, Claude, OpenAI), any Mac or Linux machine works. For local models via Ollama, 8–16 GB RAM runs qwen3:1.7b, and 16–24+ GB unified memory or VRAM runs qwen3:8b alongside local SDXL image generation.

How do I connect MCP servers or use A2A / ACP?

Add local or remote MCP servers from Settings → MCP in the dashboard. To expose a clone to other agents or editors, run ucx a2a serve (HTTP+JSON Agent-to-Agent) or ucx acp serve (stdio Agent Client Protocol).