Getting Started
Start Here
Everything on this server runs locally: your prompts, your documents, and your data stay on hardware you control. This page is the map: what a fresh installation can do, organized by what you are trying to achieve, with the fastest first step for each.
1The first ten minutes#
- Open the server's address in a browser. The landing page lists every destination this installation offers; on first run, the setup assistant walks the essential choices.
- Get a model. Admin Console, Models section: pull one from the built-in catalog (a compact chat model downloads quickly and proves the pipeline). Downloads run in the background with live progress.
- Talk to it. Open the Playground and chat. If that works, everything works: the same engine serves every API below.
By default the server listens on loopback only and accepts local calls without a key, so the first ten minutes need zero credential ceremony. Before anything reaches it from the network, read Going Live. And if any of these first steps misbehaves, Troubleshooting First Runs is the symptom-indexed triage.
2Pick your use case#
| You want to... | Do this | Go deeper |
|---|---|---|
| Ship a complete workflow from a recipe | The cookbook implements real workloads request by request: RAG over your own corpus, invoice automation, compliance redaction, a support assistant, a searchable archive | The Cookbook |
| Chat with a local model | The Playground, or any chat client pointed at this server | API Compatibility |
| Keep using your existing AI apps and SDKs | Point their base URL here: OpenAI, Anthropic, and Ollama dialects are served side by side | API Compatibility |
| Serve your coding assistant locally | Continue, Cline, and their peers point here through their provider settings; source code never leaves the machine | Local Coding Assistants |
| Move an Ollama setup here | The Ollama API is served natively: change the host and existing clients keep working, with keys and teams added | Coming from Ollama |
| Run Claude Desktop on your models | Four fields in the app's gateway dialog; the /claude page hands you each one and confirms the first request | Claude Desktop |
| Chat with your documents | Playground: attach a document and ask; answers cite the source | Grounded Answers |
| Build a searchable knowledge base | Search works out of the box on a built-in store: create a tenant, index documents, query full-text, semantic, or hybrid | The Search Engine |
| Get cited answers from your corpus | POST /lmkit/v1/search/answer returns an answer with citations, and refuses when the corpus does not support one |
Grounded Answers |
| Prove retrieval quality with numbers | Golden sets (hand-written or generated from your corpus) and persisted evaluation runs are built into the Search API | Measuring Search Quality |
| Automate document workflows (IDP) | Split scanned batches into labeled documents, classify each against your own taxonomy, extract validated fields, and route exceptions to review: every stage is an endpoint | The IDP Pipeline |
| Extract structured fields from documents | POST /lmkit/v1/extract-structured-data: your JSON Schema in, verified JSON out with per-field confidence, coordinates, and review flags |
Structured Extraction |
| Process documents programmatically | The native API (/lmkit/v1/...) covers the document lifecycle: conversion to Markdown, OCR, thumbnails, archival validation, and a full PDF suite (merge, split, redact, search, PDF/A) |
Document Processing, The PDF Toolbox |
| Analyze text at scale | Native endpoints for classification, sentiment, entities, PII detection, keywords, language detection, translation, correction, and rewriting | Text Analysis |
| Transcribe audio and video | POST /lmkit/v1/audio-transcription: recordings and video files become transcripts with a local speech model |
Transcription |
| Run OpenAI Responses-API agents locally | Stateful turns, function tools, structured output, and file_search grounding over local vector stores |
Responses and Vector Stores |
| Bring your own vector database | Embed text, documents, and images here, rerank your own results; your store keeps its role | Embeddings and Reranking |
| Define and serve your own agents | Named bundles of prompt, skills, tools, and memory: adopted by one field on any chat endpoint, discoverable by API, testable end to end | The Agent Platform |
| Give Claude and agents document tools | Connect any MCP client (Claude Desktop, Claude Code, IDEs) to the MCP endpoint: on one machine, over the stdio bridge, or hosted with keys | The MCP Server |
| Fine-tune a model on your data | The Training workbench: dataset in, live loss curves, a deployable artifact out | Fine-tuning |
| Run it for a team | Mint least-privilege keys, expose it safely, watch it run | Keys and Authentication, Going Live, Observability |
| Scale it to a farm | Identical nodes behind a load balancer: shared keys and stores, warm joins, lossless drains, autoscaling on a built-in signal | Scaling Out, Running in Containers |
| Build the case before committing | The decision framework, the cost arithmetic, and the procurement dossier each have a guide | When to Run AI Locally, The Cost Math, Passing the Security Review |
3The four doors#
Every capability above is reachable through whichever surface fits the consumer:
- The Playground, for people: chat, documents, rendered answers, zero setup (The Playground).
- The compatibility APIs, for existing software: OpenAI, Anthropic, and Ollama clients connect unmodified; moving off a hosted API is a short runbook.
- The native API, for builders: the full breadth, documented live in the API Reference.
- MCP, for agents: document tools over the Model Context Protocol (The MCP Server), governed by the same policies as everything else.
One engine and one model pool serve all four, under one key model. What a caller may do is decided by its key, never by which door it came through.
4Stated plainly#
- Nothing leaves the machine: models, prompts, documents, and indexes are local.
- The default posture is safe (loopback only); opening up is one deliberate, guided step.
- Start with the Playground to validate, then pick the door that fits what you are building.