Mycelis

Product Introduction

1. Definition: Mycelis is a private AI infrastructure and orchestration platform (technical category: AI/ML Ops and API Gateway). It provides a unified, OpenAI-compatible API layer to deploy, manage, and route requests between open-source models on dedicated GPUs, third-party managed API keys (BYOK), custom fine-tuned models, and integrated agents.
2. Core Value Proposition: Mycelis exists to provide data-sovereign, cost-optimized AI application development by abstracting hardware complexity and offering intelligent routing. Its primary value is enabling businesses and developers to deploy private AI with full control over their data, significantly reduce inference costs through smart model selection, and eliminate server maintenance overhead.

Main Features

1. Unified OpenAI-Compatible API Gateway: Mycelis acts as a single, drop-in replacement endpoint for the OpenAI API. Developers can point existing OpenAI SDKs (Python, JavaScript, etc.) to Mycelis's base URL without changing client code. This gateway virtualizes access to all underlying compute resources, including GPU instances and third-party API keys.
2. Conditional Smart Routing & VirtualModels: This is a core intelligence layer. Users configure a "VirtualModel" (identified by a slug, e.g., my-assistant). Behind this slug, multiple real models (e.g., Llama 3.1 8B, GPT-4o) are defined. Mycelis analyzes incoming requests (e.g., complexity, token count) and uses rule-based routing to forward the query to the cheapest capable model. Simple tasks go to cost-efficient open-source models, while complex ones are escalated to more powerful (and expensive) models automatically.
3. Integrated RAG Pipeline & Semantic Vector Cache: The platform includes a native Retrieval-Augmented Generation (RAG) system for building Knowledge Bases. Users can upload documents (automatic chunking and embedding), enabling models to query private data. The integrated semantic cache stores and retrieves semantically similar past queries and completions, slashing costs and latency by avoiding redundant model calls for identical or similar prompts.
4. MCP (Model Context Protocol) Agent Integration: Mycelis natively supports MCP, allowing agents to integrate with external tools and data sources (e.g., GitHub, Slack, Discord, web search, databases, custom APIs). These tools are centrally configured in an MCP library and can be made available to any VirtualModel or agent, enabling AI assistants to take actions and access real-time data.
5. Flexible Compute Options: GPU Instances & Managed Keys: The platform offers two primary compute paths. GPU Instances provide dedicated, hourly-billed GPUs (e.g., RTX 4090, A100) via partners like RunPod for running open-source models. Managed Keys (BYOK) allow users to bring their own API keys for services like OpenAI, Anthropic, or Google, or use Mycelis's managed tokens, paying only for usage with no infrastructure setup.

Problems Solved

1. Pain Point: Prohibitive and unpredictable costs of using large language model APIs (e.g., GPT-4) for high-volume applications.
2. Target Audience: Cost-conscious SMEs, enterprise teams scaling AI pilots, and indie developers needing predictable billing. Also targets developers using tools like OpenCode/OpenClaw who require private, on-premise AI agents.
3. Use Cases: Deploying a customer support chatbot using RAG on internal documentation, routed to a small local model for FAQs and a powerful model for complex issues. Creating a coding assistant (e.g., for OpenCode) that runs on private infrastructure with access to internal codebases via MCP tools. Finetuning a small model on proprietary industry data for precise, domain-specific Q&A.

Unique Advantages

1. Differentiation: Unlike pure cloud AI providers (OpenAI, Anthropic), Mycelis offers hybrid compute (self-hosted + cloud keys) and data sovereignty. Unlike raw GPU marketplaces (RunPod, Vast.ai), it adds a sophisticated orchestration layer (routing, caching, RAG, MCP). Compared to other API gateways, its deep integration with open-source models and conditional routing is a key differentiator.
2. Key Innovation: The conditional Smart Routing tied to VirtualModels is a significant technical innovation. It automates cost-optimization at the API call level based on request analysis, which is typically a manual, application-level task for developers. The combination of a semantic cache, native RAG, and MCP tool integration within a single, unified API endpoint creates a highly efficient "private AI agent" factory.

Frequently Asked Questions (FAQ)

1. How does Mycelis Smart Routing save costs compared to using the OpenAI API directly? Mycelis Smart Routing analyzes each request and automatically directs it to the most cost-effective model capable of handling it. For example, a simple text summary might be routed to a local Llama 3.1 8B model costing €0.002/1k tokens, instead of GPT-4o at €0.015/1k tokens. The company claims average savings of 62-80% while maintaining output quality for equivalent tasks.
2. Is my data secure with Mycelis, and does it comply with GDPR? Yes, Mycelis emphasizes data sovereignty. Data processed on dedicated GPU instances or via on-premise deployments never leaves your designated infrastructure. As a provider not based on US cloud infrastructure by default and offering EU hosting options, it facilitates GDPR and DSGVO compliance for European enterprises, a key advantage over many US-based AI cloud services.
3. What is the difference between a GPU Instance and Managed Keys on Mycelis? A GPU Instance is a dedicated, isolated virtual machine with a GPU (e.g., RTX A6000) that you rent by the hour to run open-source models like Llama or Mistral. Managed Keys involve using Mycelis as a proxy for third-party API services (OpenAI, Anthropic); you either provide your own key (BYOK) or use Mycelis's pooled keys, paying per token with no GPU management. The former offers full control and data isolation; the latter offers zero setup and serverless scaling.
4. Can I use Mycelis to run AI agents like OpenClaw or custom agents privately? Absolutely. Mycelis is designed as the ideal backend for private AI agents. You can deploy the model, equip it with RAG knowledge bases and MCP tools (like GitHub access), and expose it via a single OpenAI-compatible API slug. This allows tools like OpenClaw or OpenCode to connect to a fully private, powerful assistant without sending data to external APIs.
5. Does Mycelis support fine-tuning models, and how does it work? Yes, Mycelis offers a wizard-guided fine-tuning workflow, specifically for LoRA (Low-Rank Adaptation). This allows you to train models on your proprietary data without requiring deep machine learning expertise. The fine-tuned model can then be deployed on a GPU instance and integrated into your VirtualModels and routing strategies alongside other models.