Every prompt your application sends is a small buying decision, and OrcaRouter Lite makes it a smart one instead of a habit. Point your app at one OpenAI-compatible endpoint on your server, and each request is shopped across the providers whose keys you hold: OpenAI, Anthropic, Google, xAI, DeepSeek, and dozens more. Send model set to auto and the router picks the cheapest model that can actually do the job, checking whether the request needs tools, vision, or JSON mode before it commits your money. Prefer a different tradeoff and you can switch the whole gateway between balanced, cheapest, fastest, and quality strategies, with quality picks driven by real benchmark scores rather than vibes. Clients that speak Anthropic's or Gemini's native protocols connect directly too, so Claude Code, the OpenAI SDK, and google-genai all land on the same routing pipeline. Streaming responses flow through untouched, and every answer reports which model really served it. The built-in dashboard shows recent requests, spend, latency percentiles, unreachable providers, and a savings tile that tells you what your traffic would have cost if you had blindly used a flagship model for everything. Your provider keys stay encrypted on your own box, your analytics never leave it, and if every local provider has a bad day, the hosted OrcaRouter can step in as a failover safety net. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

BitRouter is a context-aware LLM router that learns which model delivers the cheapest successful outcome per workflow step, cutting agent costs by up to 80% while maintaining 96% quality versus all-frontier baselines. Point any agent runtime at http://localhost:4356 with a one-line OPENAI_BASE_URL change and BitRouter routes to OpenAI, Anthropic, Google, Groq, DeepSeek, Mistral, Moonshot, MiniMax, Nvidia, and any OpenAI-compatible endpoint simultaneously, normalizing authentication, streaming, and cross-protocol translation between wire formats. The act-observe-evaluate-learn loop traces every hop with cost, tokens, and latency attribution, scores each decision against a versioned policy-lock.yaml, then tightens routes automatically with no LLM judge in the path. Native MCP gateway auto-discovers tools from connected servers and makes them routable and governed alongside model calls. Agent Client Protocol integration enables the TUI to manage Claude Code, Codex, OpenCode, OpenClaw, Gemini, and Copilot sessions in real time with inline tool-call approval and live streaming. Built-in guardrails inspect, redact, or block risky content before requests leave your network. Virtual keys scope API access per agent or user without exposing upstream credentials. Per-agent spend caps and loop guards contain runaway cost automatically. Multi-account failover reroutes mid-run so rate limits never re-pay completed work. Ships as a single Rust binary via npm or Cargo. 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 scattering provider credentials and vendor-specific SDK logic across dozens of microservices, engineering teams route every LLM call through GoModel, an open-source AI gateway that unifies 34 model providers behind OpenAI-compatible and Anthropic-compatible endpoints. Pointing existing OpenAI or Anthropic SDKs at the gateway unlocks chat completions, the Responses API, Anthropic Messages, embeddings, batch jobs, and realtime audio across OpenAI, Anthropic, Gemini, Bedrock, Azure, Groq, DeepSeek, Ollama, and vLLM. Platform operators publish virtual model aliases that distribute traffic via weighted round-robin or lowest-cost routing while rotating multiple API keys per provider and failing over through circuit breakers during upstream outages. Sticky session routing keeps multi-turn agent conversations on the same target to preserve warm prompt caches, while exact-match and vector semantic caching via Qdrant, pgvector, Pinecone, or Weaviate cut duplicate token costs. Versioned workflows and pre-dispatch guardrails enforce prompt injection rules, rate limits, and hard USD spend budgets across hierarchical tenant user paths before requests are dispatched. The built-in admin dashboard offers live audit logs, per-tenant cost analytics, LiteLLM key migration, an interactive request playground, Model Context Protocol tool aggregation, and OpenTelemetry and Prometheus metrics. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT 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.

One OpenAI-compatible endpoint, 1,600+ models behind it: Portkey Gateway sits between your app and every major LLM provider, deciding in real time where each request should go. When a provider times out or returns errors, automatic retries with exponential backoff kick in, and fallback rules quietly move traffic to your second-choice model before users notice anything. Load balancing spreads calls across multiple API keys and providers by weight, conditional routing sends each request down a different path based on its content or metadata, and per-request timeouts cut off providers that hang. Guardrails inspect inputs and outputs as traffic passes through: block banned words, screen for sensitive content, and enforce the checks your compliance team asked for, with dozens of pre-built rules ready to attach. Text is only the start: the same API shape reaches vision models, speech-to-text, text-to-speech, image generation, and realtime websocket APIs, so multimodal features stop multiplying your integration work. Routing policies live in JSON configs you attach per request or per client, which means experiments, cost controls, and provider migrations ship as configuration changes, not redeploys. 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.

Your apps keep using the OpenAI SDK while Antirion, a self-hosted AI gateway, decides which provider answers, what each team may spend, and what gets logged. Clients speak the OpenAI, Anthropic, Gemini, or Bedrock APIs, and Antirion translates between them, so a Claude deployment can answer an OpenAI-style call, streams included. Add providers such as Amazon Bedrock, Vertex AI, Mistral, Groq, DeepSeek, or a vLLM server, then publish them as named Models. Routing rules send traffic by team or project with fallback chains, weighted splits, least-cost or least-latency targets, retries, hedging, and circuit breakers. Budgets and rate limits in dollars or tokens apply per organization, team, project, and key, with projections and alerts before money runs out. Guardrails check for personal data, secrets, keywords, and regex matches, then block, redact, flag, or divert. Exact and semantic caching answers repeated prompts without calling a provider. The request log records every attempt, route, token count, cost, and guardrail verdict, with Prometheus metrics and OTLP traces. An MCP gateway and A2A proxy put tools and agents behind the same keys. Sign in with OIDC, SAML, or GitHub, provision users with SCIM, and review every change in the audit log. Changes you choose can require a change request or a second person's approval, and the Playground sends requests from the console. 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.

Respan Gateway is a single-binary LLM gateway that lets the SDK you already use call any AI model from any provider, translating between OpenAI, Anthropic, and Gemini API formats on the way through. Point an OpenAI, Anthropic, or Gemini client at it and you reach chat models on sixteen providers, plus any OpenAI-compatible endpoint of your own, with streaming, tools, images, and reasoning translated. Traffic control works the way operators think: name a route, send it to the cheapest or fastest provider, retry on failure, fall back to a second vendor, or hedge a slow call by racing another provider. Models that never had tools get them anyway, because MCP servers and built-in search, browsing, code, document, and memory tools run inside ordinary chat requests behind the gateway. Realtime voice crosses vendors too, with an OpenAI Realtime client able to talk to Gemini Live, alongside images, speech, transcription, embeddings, and video through one endpoint. A live request log in the web UI shows the client key, model, provider, status, tokens, cost, and latency of every call, so runaway agents and surprise bills are visible immediately. Everything is a plugin on one public Rust API, from routing and pricing to tracing, with examples for guardrails, PII masking, and billing. One TOML file configures keys, providers, routes, and tools, validates before it starts, and reloads on signal. 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.

Inference Gateway puts a single OpenAI-compatible API endpoint in front of OpenAI, Anthropic, Groq, Cohere, Ollama, DeepSeek, Google, Mistral, MiniMax, Moonshot, Nvidia, and llama.cpp, so your application code never changes when you switch models or providers. The Go binary starts on port 8080 and normalizes authentication, streaming protocols, and response formats across all backends transparently. Native Model Context Protocol support auto-discovers tools from connected MCP servers and injects them into LLM requests without client-side management, enabling server-side tool execution across any provider that supports function calling. Agent-to-Agent protocol integration allows distributed agent communication through a declarative Agent Definition Language that generates production-ready Go or Rust servers from a single YAML manifest. The dedicated Kubernetes Operator manages Gateway, Agent, MCP, and Orchestrator custom resources with automatic HPA scaling, OIDC authentication, and service discovery that rebuilds MCP configurations when the discovered server set changes. Prometheus metrics and OpenTelemetry tracing provide full request-level observability across the entire inference pipeline. Middleware controls enable per-request provider selection, model routing, and fallback strategies. Official SDKs in Go, Python, TypeScript, and Rust provide typed client interfaces with streaming support. Docker Compose deployment requires only environment variables for API keys. A CNCF Sandbox applicant. 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.

GPT Load has become the go-to self-hosted AI gateway for teams managing multiple LLM provider credentials behind a single URL. The Go-built proxy transparently preserves native API formats for OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Google Gemini — applications swap one base URL and keep their existing SDK integration untouched. Twenty built-in channels cover official APIs (OpenAI, Anthropic, Gemini, xAI), cloud platforms (Azure OpenAI, AWS Bedrock, Google Vertex AI), model services (DeepSeek, Moonshot AI, SiliconFlow, Zhipu AI, Alibaba Cloud, Volcengine, OpenRouter, Groq), and subscription accounts (Codex, Claude, Antigravity, Grok) using the same credential management, scheduling, and health system. The intelligent key pool rotates among valid credentials using atomic counters for fair high-concurrency distribution, automatically blacklists failing keys after configurable thresholds, and recovers them via scheduled health checks — all transparent to clients. Weighted load balancing distributes traffic across multiple upstream endpoints while per-credential proxies route individual keys through different egress paths. The Vue 3 management dashboard provides real-time statistics, group configuration, key testing, request logs with full debugging context, per-model cost estimates, and route inspection. Docker deployment exposes the management interface and proxy on port 3001 with SQLite by default, MySQL or PostgreSQL optional, and Redis for distributed leader-follower cluster scaling. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

Juggling separate API keys, incompatible payload schemas, and unpredictable rate limits across OpenAI, Anthropic, and Google Gemini slows down every AI workflow, and Octopus solves it by consolidating all your upstream model providers into one self-hosted API gateway. Point Claude Code, OpenAI Codex CLI, or any SDK client at a single gateway URL and translate requests transparently between OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Google Gemini endpoints without modifying client code. Group multiple provider channels under a unified model alias so your coding agents automatically fail over to backup providers and shield active tasks from upstream HTTP errors or rate limits. Watch live end-to-end request routing on the interactive topology canvas as prompts flow from client applications through model groups to specific upstream channels. Synchronize available model catalogs directly from connected providers and pull up-to-date token pricing from models.dev, or define custom per-million input, output, and cache-read rates for private deployments. Inspect detailed request logs, monitor token consumption trends, and audit daily API spend across every channel from the web dashboard. 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.

New API is a Go-powered LLM gateway that funnels over 40 AI providers, from OpenAI and Anthropic to Google Gemini, Azure, AWS Bedrock, DeepSeek, and Mistral, through a single OpenAI-compatible endpoint with intelligent routing and enterprise cost controls. Bidirectional format conversion translates between OpenAI Chat Completions, Claude Messages, and Gemini GenerateContent APIs transparently, so clients call any provider using their preferred format while the gateway handles the rest. Routing logic supports weighted random channel selection, priority-based failover, and automatic retry on provider errors to maximize uptime. The React admin dashboard shows usage charts, token consumption tracking, per-model cost breakdowns, and real-time request monitoring. Multi-tenant architecture includes three role levels, token-based authentication, per-user quota management, and a three-phase billing system with tiered pricing via a custom expression language. Multimodal support covers text completion, vision, embeddings, text-to-speech, speech-to-text, image generation, and creative task providers like Midjourney-Proxy and Suno-API. The interface ships in five languages including English, Chinese, Japanese, and French. Docker deployment runs with SQLite or MySQL for persistence. 45,000+ GitHub stars. AGPL-3.0 licensed.

Anyone who pays for Claude Code, Codex, and Gemini CLI separately ends up juggling three endpoints, three sets of keys, and three kinds of rate-limit surprise, and CliRelay collapses all of that into one self-hosted gateway with a control panel. Clients speak the OpenAI Chat or Responses, Anthropic Messages, or Gemini protocol, and CliRelay translates each request for whichever upstream account is healthy. Channel groups limit where each key may send traffic, round-robin, fill-first, and session-sticky scheduling spread it, and cooldowns with automatic failover move requests off accounts that have run out of quota. Add upstreams from the /manage console through browser sign-in, device codes, or bulk credential import, covering Claude, Codex, Gemini, Antigravity, Vertex AI, AWS Bedrock, Grok, Qwen, Kimi, and any OpenAI-compatible service. Portal accounts own client keys with daily, period, and lifetime quotas, RPM and TPM limits, and per-model pricing. A monitor center scores health and charts live traffic, P50 to P99 latency, and time to first token, while every request lands in a searchable log with tokens, latency, and cost. Tenants, roles with resource.action permissions, content moderation profiles, an outbound proxy pool, and an audit trail make shared deployments workable, and the console speaks English, Simplified Chinese, and Russian. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

FreeLLMAPI collapses the chaos of 29 free LLM providers — Google AI, Cerebras, Groq, Mistral, OpenRouter, GitHub Models, Cohere, Cloudflare Workers AI, NVIDIA NIM, HuggingFace, SiliconFlow, Reka, Z.ai, and more — into a single /v1 endpoint that speaks both OpenAI and Anthropic protocols. The smart router selects the best available model for each request, automatically fails over to the next provider when rate limits hit, and tracks per-key token consumption so you never exceed a free-tier cap. Keys are stored with AES-256-GCM encryption and clients authenticate using a single unified bearer token, never exposing upstream provider credentials to downstream applications. The catalog tracks 251 model families across 358 provider/model endpoints with approximately 4 billion tokens per month of aggregate free-tier capacity, auto-refreshing from a signed manifest at freellmapi.co twice daily without requiring git pulls. Beyond chat completions, the proxy handles embedding, image generation, and audio/TTS endpoints, plus structured outputs with JSON schema forwarding, JSON healing, and format-ignore failover. An integrated MCP server at /mcp provides gateway introspection for coding agents, while the self-hosted OpenAPI reference at /v1/docs documents every route. Compatible with OpenAI SDKs, LangChain, LlamaIndex, Continue, Claude Code, and Hermes — just change base_url. Deploy via Docker, npm, or build from source. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.

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.

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.

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.

Large language models get a polished front end that can run fully offline: Open WebUI is the self-hosted front end of choice. It talks to local model runners, primarily Ollama, and to any OpenAI-compatible API, so LM Studio, vLLM, Groq, Mistral, OpenRouter, and cloud providers all plug into the same chat interface and can be mixed per conversation. RAG is built in: upload files to knowledge bases or reference them in chat with the # command, backed by a choice of nine vector databases (ChromaDB and PGVector officially maintained) and multiple extraction engines including Tika and Docling, with hybrid BM25-plus-vector search and cross-encoder reranking. Web search results from providers like SearXNG, Brave, and Tavily inject directly into conversations. Extensibility comes from Python tools and functions that run inside the chat, a Pipelines plugin framework, and native MCP support. Multi-user features include RBAC, SSO, and group permissions, and the instance itself exposes an OpenAI-compatible API your own apps can call.