OpenLLM serves any large language model as an OpenAI-compatible API endpoint from a single CLI command, handling model download, backend selection, quantization, and port binding automatically. It supports the full spectrum of popular models including Llama 3.3, Qwen2.5, DeepSeek, Mistral, and Phi3, choosing between vLLM and PyTorch inference backends based on hardware capabilities. When vLLM is available, continuous batching with PagedAttention achieves up to 23x throughput improvement over naive serving, while GPTQ and bitsandbytes quantization reduces memory requirements for GPU-constrained deployments. The server exposes a RESTful API on port 3000 with full OpenAI client library compatibility, enabling drop-in replacement for commercial providers in any application using the standard chat completions format. A built-in web chat UI at the /chat endpoint provides immediate interactive testing without external clients. Custom model repositories allow teams to maintain private catalogs of fine-tuned models alongside the default repository that tracks the latest releases. Deployment workflows generate production-ready Docker images automatically, with Kubernetes manifest support for orchestrated scaling. Native integration with LangChain and LlamaIndex supports RAG pipelines, Transformers Agents enables tool-calling workflows, and HuggingFace Hub handles model discovery. Server-Sent Events enable real-time token streaming across all API endpoints. Backed by BentoML's production ML infrastructure. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache 2.0 licensed.

ODS turns a GPU machine or server you already own into a complete private AI stack with a single install command, replacing a shelf of separate tools with one system that serves models, chat, agents, automation, and search from your own hardware. The installer reads your GPU, VRAM, memory, and disk, then downloads a GGUF model matched to that hardware tier, and a compact bootstrap model starts answering within minutes while the full model finishes downloading. From the web dashboard you watch every service's health, ports, and resource use, restart what misbehaves, and check storage and updates without touching a terminal. You chat with local models in Open WebUI, hand background work to the Hermes and Portal agents, and build automations in n8n that connect the stack to more than 400 other apps. Documents you add become searchable knowledge through Qdrant vector storage, so answers cite your files instead of the model's memory. SearXNG and Perplexica answer web questions without trackers or accounts, Whisper transcribes speech, Kokoro reads text aloud, and ComfyUI generates images, all on the same box. Every bundled service is an extension you can enable, disable, or replace, and a Privacy Shield scrubs personal data before anything reaches an optional cloud model through the LiteLLM gateway. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.

With over 48,000 GitHub stars and monthly releases since March 2023, LocalAI is the self-hosted AI engine that replaces every OpenAI endpoint with a single Docker container running on your own infrastructure — serving chat completions, image generation, text-to-speech, speech-to-text, embeddings, vision, video generation, and function calling through identical API schemas that require zero application code changes. The composable backend architecture isolates each inference engine as a separate gRPC service running in its own OCI container, so llama.cpp, vLLM, SGLang, transformers, whisper.cpp, diffusers, MLX, Stable Diffusion, and Flux install on demand without touching the core, can run on separate machines, and a fault in one never affects others. Hardware acceleration spans NVIDIA CUDA 12 and 13, AMD ROCm, Intel oneAPI/SYCL, Apple Silicon Metal, Vulkan, and NVIDIA Jetson L4T — or runs entirely on CPU without any GPU. Built-in AI agents support autonomous tool use, retrieval-augmented generation, Model Context Protocol integration, and skill-based workflows directly in the web interface. The model gallery provides curated YAML configuration files for hundreds of models that install with a single command, while P2P federated inference distributes model shards across multiple machines for running models larger than any single node's memory. Multi-user API key authentication with quotas and role-based access enables team deployments. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

Your prompts never have to leave the building: TextGen runs open large language models on your own hardware behind a chat interface and an API that speak fluent OpenAI. Load GGUF or Hugging Face weights through llama.cpp, ik_llama.cpp, Transformers, ExLlamaV3, or TensorRT-LLM, and switch models or backends from the interface without restarting the server. Chat mode gives you conversations with characters and system prompts, notebook mode behaves like a familiar completion editor, and both accept image attachments for vision models and documents for context. Models that support tool calling can search the web, fetch pages, and do math through single-file Python tools, and MCP servers plug in for larger tool catalogs. Start it with the API flag and the same model answers at chat, completions, and messages endpoints that mirror OpenAI and Anthropic, complete with streaming, logprobs, embeddings, and optional API key authentication. Portable builds for Linux, Windows, and macOS bundle every dependency, so the distance from download to a running model is an unzip and one shell command. Because everything is local, prompts, documents, and outputs stay on machines you control, and parallel request handling lets a single loaded model serve a whole team's tools at once. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. AGPL-3.0 licensed.

Imp runs large AI chat models privately on one NVIDIA RTX 50 or RTX PRO graphics card, exposing them through OpenAI and Anthropic compatible APIs plus a built-in browser chat. Start it with a single Docker command, point it at a model, and the chat opens in your browser with saved conversations and a live readout of generation speed. Coding agents such as Claude Code and Codex connect by changing one base URL, so your editor talks to the model on your own hardware instead of a rented cloud endpoint. Tool calls, structured JSON output, streaming responses, and reasoning models all work over the wire, and requests can be protected with an API key, per-IP rate limits, and concurrency caps. Load GGUF files you already have or NVFP4 and FP8 SafeTensors checkpoints straight from Hugging Face with no conversion step before the first answer. When a model is bigger than the card, mixture-of-experts weights stay in system RAM while the busiest experts are cached on the GPU, so a 123.6 GiB model still answers from a 32 GB card. Supported models can also read images alongside text, and tokenize, health, and model listing endpoints make the server easy to wire into scripts and monitors. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

A single GPU can only hold so many models at once, and when a chat server, an image pipeline, and a voice engine all grab VRAM together, somebody runs out of memory and crashes. Berth fixes the scramble by making every engine lease the card: wrap a launch in berth run and the daemon grants memory in turn, queues the rest with an honest place in line and a start estimate, and drains a holder cleanly before the next load begins. The broker learns each engine's real footprint from its runs, so grants are sized by measured peaks instead of guesses, and processes that grab memory without a lease are reported as stowaways, never killed, with contested grants flagged in plain sight. A collisions board counts what the scheduling prevented: avoided collisions, VRAM handed off, drains, stuck releases, and refusals, and every count links back to a gap free event log you can follow live. The Harbor web UI turns that into a departures board of leases, a six hour tide of memory by holder, engine cards with decode throughput and a token ledger, plus a per model leaderboard and thirty days of stored history. Alerts reach your desktop or a webhook, Prometheus scrapes the metrics endpoint, and multi GPU machines can even pass one card through to a VM while the others keep serving. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.

SmarterRouter stops you from playing model roulette with your local AI setup, choosing the right model for every prompt so you never have to. Point your applications at one familiar endpoint and send the request; the router studies the prompt, weighs category and complexity against benchmark results gathered on your actual hardware, and hands the job to the model most likely to answer well. It discovers the models your backend already serves, profiles each one with standard tests the first time it meets them, and keeps learning from the feedback your usage generates. When the first choice fails, a backup model takes over without your application noticing. Repeated questions come straight back from cache instead of burning another generation. Your graphics memory is watched continuously across one card or several, and models are unloaded before an ambitious request runs the machine dry. Cloud providers can join the same routing pool when a question deserves them, while everything local stays local and free. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

With over 126,000 GitHub stars and adoption across professional studios, research labs, and independent creators, ComfyUI has become the most widely used node-based interface for generative AI workflows — supporting image, video, audio, and 3D content creation through a single visual canvas. The graph editor natively supports Stable Diffusion 1.5, SDXL, SD3.5, Flux.1, Flux.2, HunyuanDiT, Lumina Image 2.0, HiDream, Qwen Image, and Pixart for image generation, plus Wan 2.1 and 2.2, LTX-Video, HunyuanVideo 1.5, CogVideoX, and Mochi for video, ACE-Step and Stable Audio for audio, and Hunyuan3D 2.0 for 3D models. Built-in tools handle inpainting, outpainting, ControlNet conditioning, LoRA and Hypernetwork loading, ESRGAN upscaling, area composition, model merging, and GLIGEN spatial control without writing code. The execution engine implements asynchronous queue processing with partial graph re-execution, running only changed nodes between iterations, and smart VRAM management that offloads models on GPUs with as little as 1 GB of memory. API nodes optionally connect to closed-source models through Comfy API while the core runs fully offline. Reusable subgraphs and App Mode expose complex workflows as simplified interfaces for non-technical users. The V3 custom node schema enables stateless execution with async support and process isolation. The TypeScript and Vue frontend ships as a PyPI package with stable releases every two weeks. Workflows save as JSON and embed in generated PNG, WebP, and FLAC files for reproducibility. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. GPL v3.0 licensed.

With 12,800 GitHub stars and backing from the same developer who created ControlNet, Stable Diffusion WebUI Forge replaces Automatic1111's inference backend with a dynamic GPU memory management system that runs SDXL 30-75% faster while consuming significantly less VRAM — enabling 1024x1024 generation on 6GB cards where A1111 requires 8GB or more. The Gradio 4 interface provides txt2img, img2img, inpainting, and outpainting workflows with a Forge Canvas supporting pressure-sensitive input from Wacom tablets and Microsoft Surface devices. Native Flux.1 model support loads Flux Dev and Schnell checkpoints using BitsandBytes NF4 and FP8 quantization for deployment on consumer GPUs without model splitting. Built-in ControlNet integration includes all preprocessors — Canny, Depth, Normal, OpenPose, MLSD, Scribble, Segmentation, Tile, and IP-Adapter — without requiring separate extension installation. The extension ecosystem maintains full compatibility with popular Automatic1111 extensions including Adetailer for face enhancement, After Detailer, Regional Prompter, and Dynamic Prompts. LoRA loading supports standard, LyCORIS, and DoRA formats with automatic weight detection. The API provides RESTful endpoints for txt2img, img2img, extra single/batch processing, and progress monitoring enabling headless batch generation. Deploy via one-click installer package, Python virtual environment, or Docker with NVIDIA GPU passthrough. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. AGPL-3.0 licensed.

One endpoint, every model you run, and a government-grade paper trail for each request. OpenGateLLM sits in front of your AI infrastructure as the single door: applications call one OpenAI-compatible API, and the gateway decides which backend answers, a self-hosted vLLM cluster, an Ollama box in the corner, a Hugging Face TEI embeddings service, or a hosted provider when you allow one. You publish friendly model names as routers and attach real provider APIs behind them, then load-balance across those backends by shuffle or least-busy so a second GPU node starts earning its keep the moment you register it. Chat completions, embeddings, reranking, OCR, and audio transcription each get their standard endpoint, so existing OpenAI SDK code, LangChain, and LlamaIndex projects point at a new base URL and keep working. Governance is built in rather than bolted on: organizations, users, and roles control who may call which model, API keys authenticate every request, and rate limiting keeps one enthusiastic team from eating the whole cluster. Usage monitoring follows every inference, and carbon footprint estimates attach to models through their hosting zone and parameter counts. Retrieval-augmented generation against Elasticsearch or Qdrant plugs into the same surface. No chat history is stored, which is rather the point when the data is yours. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

Backed by 56,000+ GitHub stars and over 240 million Docker pulls, LiteLLM delivers the open-source AI gateway trusted by Netflix, Lemonade, Rocket Money, and thousands of engineering teams to route every LLM request through one unified API. The Rust-core gateway adds sub-millisecond overhead per request with 8ms P95 latency at 1,000 RPS, 15x throughput improvement and 11x lower memory footprint compared to Python-only proxies. A single OpenAI-compatible endpoint connects to 100+ providers and 1,800+ models spanning OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Vertex AI, Hugging Face, vLLM, Nvidia NIM, Ollama, and Mistral with day-zero support for new model releases. The Auto Router V2 classifies request complexity across four tiers using rule-based scoring, semantic keyword matching, and adaptive Thompson sampling to route each request to the most cost-effective model without API calls or training data. Virtual API keys enable multi-tenant governance with per-team, per-user, and per-project cost tracking, budget caps with automatic fallback rerouting, and role-based access control. Built-in guardrails provide PII masking, prompt injection detection, and model-graded evaluation before requests reach providers. The Agent Gateway extends routing from model calls to agent workflows with MCP server integration. Observability integrates with Langfuse, Arize Phoenix, OpenTelemetry, and MLflow for complete request tracing. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

Bifrost is an open-source AI gateway that unifies 23+ LLM providers into a single OpenAI-compatible endpoint with automatic failover, semantic caching, and built-in cost governance, so one provider going down never takes your production AI application with it. Point your existing OpenAI or Anthropic SDK at Bifrost's local endpoint and gain access to OpenAI, Anthropic, AWS Bedrock, Google Vertex, Azure, Groq, Mistral, and Ollama without changing application code. Define fallback chains that automatically switch providers when one returns errors or exceeds latency thresholds, keeping response times stable during outages. The built-in web dashboard at port 8080 lets you configure providers, create virtual API keys, monitor live request traffic, and review analytics without editing configuration files. Semantic caching combines exact hash matching with vector similarity search via Weaviate, serving cached responses for identical or paraphrased prompts in sub-millisecond time to cut costs on repetitive workloads. The MCP gateway connects AI agents to external tools like filesystems, databases, and web APIs, exposing them to clients such as Claude Desktop and Cursor with per-key allow-lists. Four-tier budget hierarchy at customer, team, virtual key, and provider levels enforces spend caps, rate limits, and model restrictions across your organization. Extend functionality through custom Go plugins for analytics, monitoring, or security middleware. Native Prometheus metrics and OpenTelemetry distributed tracing give operations teams full production observability. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache 2.0 licensed.

Instead of burying your documents in a vector index you cannot inspect, LLM Wiki v2 files them into a Markdown wiki you can read, browse, and question. Drop text, Markdown, or PDF files anywhere on the page and each one is catalogued automatically, filed into a folder that says what it holds, and linked to the notes it relates to. Ask a question and the answer cites the exact notes it used, with clickable sources you can open, verify, and correct. New documents update the note on their subject instead of piling up as disconnected chunks, so what you teach it stays taught. A small classifier routes filing and questions through the folder tree, keeping the wiki quick and cheap as it grows, and you can browse the tree, follow see-also links, move or delete notes, and watch the tokens spent by both the model and the classifier. Connect OpenRouter, OpenAI, Gemini, a local Ollama, or any compatible endpoint, and expose the same wiki to Claude Code or Codex as long-term memory they can recall from and write to. Because every note is plain Markdown in a folder you own, the whole knowledge base can live in git beside your other files. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. AGPL-3.0 licensed.

Backed by Y Combinator and trusted by over 2,300 companies processing billions of observations monthly, Langfuse is the most widely adopted open-source platform for building, monitoring, evaluating, and debugging LLM applications. The hierarchical tracing engine captures every LLM call, tool invocation, retrieval step, and agent action as nested spans based on OpenTelemetry, with automatic cost calculation, latency tracking, and token usage attribution across sessions and users. Prompt Management separates prompts from code with versioned artifacts, label-based deployments, one-click rollbacks, and runtime SDK fetching with server-side caching, while linking every generation back to its exact prompt version for attribution analytics. The evaluation system supports LLM-as-a-judge scoring, heuristic code evaluators, user feedback collection, and manual annotation workflows that run automatically on production traces or against curated datasets. The Playground enables interactive prompt testing on real production inputs with side-by-side model comparison across providers. Datasets and Experiments define test cases for systematic benchmarking with comparative result visualization. Native SDKs for Python and TypeScript provide decorator-based instrumentation, while 100+ integrations cover LangChain, LlamaIndex, OpenAI SDK, LiteLLM, Vercel AI SDK, and any OpenTelemetry-instrumented framework. The analytics dashboard surfaces cost breakdowns, quality scores, latency percentiles, and usage trends across models and prompt versions. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

Your AI application is burning through API tokens faster than you can refresh the billing page, and you have no idea which prompt template is responsible. OpenLIT plugs that visibility gap with a self-hosted observability platform built specifically for LLM workloads. Add one line of code to instrument 90+ LLM providers, agent frameworks, and vector databases, then watch every request flow through a tracing dashboard that shows tokens consumed, latency measured, and dollars spent per call, per model, per environment. The requests view lists every LLM interaction with provider, model, cost, and token breakdown in a filterable table, while the trace detail panel lets you drill into individual spans to read the exact prompt sent and response received. Prompt Hub turns prompts into versioned artifacts you deploy, rollback, and A/B test without touching application code. OpenGround compares models side by side on the same input, so you can evaluate cost-versus-quality tradeoffs before committing to a provider. Automated evaluations run LLM-as-a-judge scoring on live production traces, flagging hallucinations, bias, and toxicity in real time. The Vault stores and rotates API keys centrally so secrets stay out of your codebase. Custom dashboards let you build drag-and-drop monitoring views with charts, stat cards, and tables backed by SQL queries against ClickHouse. GPU utilization, memory, temperature, and power metrics feed into the same platform for end-to-end infrastructure visibility. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache 2.0 licensed.