Who lumina is for#
Privacy-focused developers who need a fully local AI coding agent
Developers who rely on AI for coding assistance but refuse to send their codebase to cloud APIs get a complete agentic framework with sandboxed code execution, codebase indexing, and browser automation, all running locally. Lumina's persistent memory means context from last week's debugging session is still available today without re-explaining the project.
Skip if:
If your primary need is team collaboration or shared agent state across multiple users, Lumina has no multi-user support. It is a single-user desktop agent.
Power users building a personal AI operating system
Users who want an AI that learns their preferences, maintains long-term project context, and runs scheduled background tasks will find the MemPalace, dreaming, and My Human profile closer to what they want than any cloud assistant. The 14-backend abstraction lets you switch models or providers without reconfiguring memory or tool settings.
Skip if:
If you want a managed, zero-maintenance AI assistant, Lumina requires Python environment management and a GPU. The project is in public beta, so some features require tolerance for rough edges.
Users migrating off cloud AI automation platforms
People paying for Kimi Work, Gumloop, or similar AI automation platforms and wanting to eliminate per-workflow or per-seat fees can move to Lumina for a self-hosted alternative with persistent memory and approximately 70 agentic tools. The Apache-2.0 license means no future licensing surprises.
Skip if:
If you need visual no-code workflow building or team pipeline management, Lumina does not have a workflow builder. It is a personal agent, not an automation orchestration platform.
The problem it solves#
Most AI assistants reset on every session. Your context, projects, and preferences start from scratch each time, forcing you to re-explain yourself or maintain external notes to compensate. Cloud-based platforms solve persistence by storing context on their servers, which introduces its own challenges: your data lives under someone else's retention policies, you pay per seat or per workflow run, and any change in the provider's pricing or service terms affects your automations immediately.
The deeper challenge is that most 'local' AI tools are not truly local. They support a local model backend but sync memory and session history to the cloud. Running inference on your GPU while your conversations live on a remote server is not data ownership. Users who want an agent that genuinely learns their preferences and operates entirely within their own infrastructure have no mainstream option.
How it solves it#
Multi-tier MemPalace with temporal decay
Lumina's memory system is a three-layer SQLite-backed architecture: Layers 0 and 1 hold permanent identity and structural facts; Layer 2 holds episodic memory with a decay constant of lambda 0.0083, giving roughly 78% retention after 30 days and 47% after 90. All three layers inject automatically into the system prompt each turn. AAAK compression keeps token overhead low on constrained hardware.
Approximately 70 pre-installed agentic tools
Lumina ships with around 70 tools covering file operations, sandboxed code execution, browser automation, web search, codebase indexing, and more. You can write and register custom tools from within Lumina. Non-owner sessions (Discord bots and future public integrations) receive only a named restricted tool profile, with tool creation structurally absent from those sessions.
Runtime backend switching across 14 providers
The LLM layer is a shared interface with 14 implementations: local options include llama.cpp (primary), Ollama, LM Studio, and vLLM; cloud options include Anthropic, OpenAI, Google Gemini, Moonshot Kimi, and Alibaba Qwen, plus OpenRouter, DeepSeek, Groq, and a generic custom slot. Each backend tracks its own context window independently. Switching takes effect immediately, no restart required.
Voice cloning and local TTS
Lumina supports voice cloning with local text-to-speech synthesis, enabling a personalized audio output profile that runs on your hardware. Combined with nine swappable AI personas (including Lumina, HAL 9000, and Rick Sanchez), the agent's identity is fully customizable without cloud voice APIs.
Remote access via Telegram and Discord
Telegram integration gives full access to the same toolset as the desktop interface, locked to a single authorized chat ID. Discord integration creates a public-safe bot with a restricted tool profile (web search and skill recall, no filesystem or memory access), rate-limited per user. Both channels are configured from the Communications tab in Settings with no JSON editing required.
Dreaming and idle memory synthesis
When a session goes idle, Lumina runs a synthesis pass over the conversation and writes a compact summary to a separate nightstand memory space. Dream entries are tagged with their own provenance and are fully reviewable before promotion. The same mechanism drives context compaction, summarizing rolled-off turns into Layer 2 of the MemPalace so nothing is silently discarded.
Strengths and trade-offs#
Strengths
- Genuine local-first architecture with no silent cloud fallbacksEvery component (inference, memory, voice, tools, browser automation) runs on your hardware. The only network requests are ones you explicitly initiate. Unlike platforms with telemetry or cloud fallbacks baked into onboarding, Lumina has no account to create and no data leaving your machine by default. Apache-2.0 license with no usage caps.
- Designed for consumer GPUs, not data center hardwareLumina was built and tested on a 4GB Nvidia Quadro T1000. Feature decisions were filtered through that constraint, producing a system that fits mid-range consumer hardware rather than requiring server-grade VRAM. If your machine can run a mid-range local model, Lumina is designed to run on it.
- Explicit trust model with deterministic security guardrailsEvery session carries an explicit owner flag; non-owner sessions start with everything disabled. Tool creation is structurally absent from non-owner sessions, not just toggled off. A deterministic shell-command denylist (regex-based, not an LLM judgment) blocks catastrophic operations like wildcard deletes and force-pushes even for trusted sessions. Prompt injection resistance tags external content as data, never instructions.
- Memory that grows without manual curationThe dreaming system and My Human profile update automatically from conversation patterns. My Human synthesizes an evolving picture of you from idle-sweep review, without overriding what you have explicitly stated. The MemPalace injects all three layers into every system prompt, so long-term context is always available without manual note management.
Trade-offs
- -Requires a GPU and comfort with Python environment setupLocal inference via llama.cpp or Ollama requires a GPU (4GB VRAM is the tested baseline). Setup assumes terminal familiarity and Python environment management. The README explicitly states that Lumina is not trying to be the easiest AI assistant. Users who want a zero-configuration cloud product are not the target audience.
- -macOS and Windows support is in betaThe README describes macOS and Windows compatibility as in the testing phase, with beta testers needed. Linux is the primary development platform. Teams relying on macOS or Windows for production use should expect rough edges and possible feature gaps until those platforms stabilize.
- -Single-user only, no team collaborationLumina is designed as a personal desktop agent for one user. There is no shared workspace, no multi-user memory, and no team permission system. Use cases that require multiple people collaborating through a shared AI agent are not supported.
lumina vs alternatives#
Lumina vs Kimi Work
Kimi Work is a cloud-hosted AI workspace product from Moonshot AI, with no self-hosting option and subscription-based access. Lumina is a local-first desktop agent where all inference, memory, and tool execution run on your own hardware under Apache-2.0.
| Feature | Lumina | Kimi Work |
|---|---|---|
| License | Apache-2.0 | Proprietary |
| Deployment | Self-hosted desktop | Cloud only |
| Memory persistence | Local MemPalace | Cloud-managed |
| Data residency | Your machine only | Moonshot AI servers |
| Team collaboration | Not supported | Yes |
| Cost | Free (self-hosted) | Subscription |
Lumina is the better choice when data residency is a hard requirement, when you want persistent memory that never expires without a subscription, and when you need an extensible agentic tool framework on your own hardware. Kimi Work is the better choice for teams that need shared workspaces and do not want to manage hardware, Python environments, or GPU setup.
Lumina vs Gumloop
Gumloop is a cloud-hosted AI workflow automation platform focused on visual no-code pipeline building for teams. Lumina is a single-user desktop agent with approximately 70 built-in tools and a custom tool framework, designed for personal agentic workflows rather than team automation pipelines.
| Feature | Lumina | Gumloop |
|---|---|---|
| License | Apache-2.0 | Proprietary |
| Deployment | Self-hosted desktop | Cloud only |
| Tool framework | ~70 tools, custom creation | Visual workflow builder |
| Primary interface | Desktop GUI and CLI | Web browser |
| Data privacy | Fully local | Cloud-managed |
| Team automation | Not supported | Core use case |
Lumina is the better choice when you want a persistent personal AI agent with local tool execution and full data privacy, with no per-workflow billing. Gumloop is the better choice when your workflows involve team handoffs, when you need a visual no-code interface for multi-step automations, or when you need to manage pipelines across an organization without any infrastructure overhead.
What it's built on#
- Languages
- Python
FAQ#
Is Lumina free to use?
Yes. Lumina is licensed under Apache-2.0 and is free to download and run on your own hardware. There is no cloud subscription, no usage-based billing, and no account to create. The only costs are your own compute (a consumer GPU with at least 4GB VRAM is the tested baseline) and your time setting up a Python environment. Optional cloud backends like OpenAI or Anthropic require their own API keys and incur their standard costs.
What hardware does Lumina require?
Lumina was built and tested on a 4GB Nvidia Quadro T1000 and is designed to run on consumer GPUs at that VRAM level. A GPU is needed for local inference via llama.cpp, Ollama, or similar backends. If you use a cloud backend (OpenAI, Anthropic, Gemini), local inference is not required, but memory, tools, and the UI still run locally. Linux is the primary supported platform; macOS and Windows support is currently in beta testing.
How does Lumina's memory system work?
Lumina's MemPalace is a three-layer SQLite-backed architecture. Layers 0 and 1 hold permanent identity and configuration facts. Layer 2 holds episodic memory with temporal decay: roughly 78% retention after 30 days and 47% after 90 days (decay constant lambda 0.0083, tunable). A dreaming feature synthesizes idle-session summaries into memory automatically. My Human is a separate profile that Lumina updates from conversation patterns without overriding what you have explicitly stated about yourself. All layers inject into every system prompt.
Can Lumina be accessed remotely from a phone or other device?
Yes. Lumina has two live remote channels. Telegram integration gives full access to the same toolset as the desktop interface, locked to your specific chat ID so others cannot interact with your instance. Discord integration creates a public-safe bot with a restricted tool profile (web search and skill recall, no filesystem or memory access), rate-limited per user. Both are configured from the Communications tab in Settings.
How does Lumina compare to cloud-based AI platforms like Gumloop or Kimi Work?
Gumloop and Kimi Work are cloud-hosted platforms where your workflows, memory, and agent state live on their servers. Lumina is a self-hosted desktop agent where all computation and data stay on your machine under Apache-2.0 with no subscription fees. The tradeoff is setup complexity: Lumina requires a GPU and Python environment, while cloud platforms work in a browser. Lumina is the better choice when data privacy and eliminating per-seat or per-workflow costs matter more than zero-configuration ease.
Similar open-source tools#
maka
Local-first AI agent workspace with append-only execution logs
Flare
Graph-first IDE for agentic coding and real-time oversight
guaardvark
Self-hosted AI workstation with agents, video, voice, and RAG.
berd
Self-hosted desktop app for AI agent work, model-agnostic by design.
rakazo
AI teammates you own: your keys, your model, your machine.
mex
Persistent memory and code graph for AI coding agents