You control
Private by architecture
Source documents, indexes and processing stay inside the environment you control; outbound access is a governed decision.
The Private AI Application Server
Run your AI stack privately. Not just the model
LM-Kit One serves models, agents, RAG, search and document intelligence behind the APIs your applications already speak.
Signed packages land September 15.
Four outcomes, stated plainly.
You should understand these without knowing what RAG or MCP mean. The depth comes after.
Extract
Turn documents into validated, structured data
PDFs, scans and photos become schema-valid records, with a confidence figure on every field so review goes only where it is needed.
Structured data extractionKnow
Search company knowledge, get cited answers
A search engine over everything you ingest. Ask in plain language; every answer names the document, the page and the passage.
Private company knowledgeAct
Run agents and tools under your policies
Function calling against a governed tool catalog: allow and deny rules, risk levels, and human approval in front of the changes that matter.
AI agentsIntegrate
Point existing AI apps at your private server
LM-Kit One serves the OpenAI, Anthropic and Ollama dialects plus MCP, so existing clients and coding agents connect by changing a base URL.
The compatibility matrixThe whole private AI stack, engineered as one.
What teams otherwise assemble from a dozen components ships as one engine: built, versioned and improved together, 150+ releases since 2024.
AI Agents
Planning, orchestration, memory, tools, MCP
Document Intelligence
OCR, parsing, extraction, classification, conversion
Vision & Multimodal
Image understanding, VLM chat, image embeddings
RAG & Knowledge
Vector search, hybrid retrieval, citations
Text Analysis
NER, PII, sentiment, classification
Speech & Audio
Local transcription, 100+ languages
Text Generation
Chat, translation, rewriting, summaries
Local Inference
The local runtime they all share: CPU, CUDA, Vulkan, Metal
Serving & Operations
The layer LM-Kit One adds: four API dialects, authentication, policies, audit, console, telemetry, fine-tuning jobs
The documents worth automating are the ones you cannot send.
Contracts, claims, invoices, policies, personnel files: the work that would pay for AI is exactly the work you cannot, or should not, hand to a hosted service.
Why hosted AI is out
Policy
Internal rules forbid handing customer or personnel material to a third party.
Contract
A customer agreement or DPA names who may process the data. Not a hosted model.
Regulation
Residency, sector rules and disclosure obligations decide where processing happens.
Cost
Per-page and per-token billing scales with exactly the volume you automated.
Connectivity
Air-gapped or intermittent sites cannot depend on an external service being up.
What remains is manual work or a dozen glued components. One server is the third path →
A server you deploy, or a runtime you embed.
One technology, two products. LM-Kit One serves every application and team; LM-Kit.NET compiles the same engine into the product you ship.
The private AI application server
Deploy LM-Kit One
Serve models, agents, search and document intelligence to every application over OpenAI, Anthropic, Ollama and MCP APIs, governed from one admin surface.
See LM-Kit OneThe embedded private AI runtime
Embed LM-Kit.NET
Add the same capabilities directly to your .NET product, without a Python sidecar or a cloud document dependency. One NuGet, in process.
See LM-Kit.NETBoth are reached from any stack: governed MCP tools, HTTP APIs, or in-process .NET.
You should not have to trust a black box.
LM-Kit is built the other way: control where it runs, verify what it produces, and govern what it is allowed to do.
The old way
Send the documents to the intelligence, and trust what comes back.
The new standard
Bring the intelligence to the documents, and verify what comes back.
You verify
Verifiable by output
Structured data carries per-field confidence, answers carry citations, and document changes can be checked against the original.
You govern
Governed in operation
Models, tools, access and disclosure are yours to set, with human review in front of the actions that matter.
Fully local, or selective disclosure.
Both keep source material inside your infrastructure. They differ in which model does the reasoning, and that difference deserves an exact answer.
Mode A
Fully local
A local model does the reasoning as well as the processing. Documents, OCR, indexes, embeddings and output stay inside the perimeter.
Mode B
Local processing, external reasoning
An external assistant holds the conversation; the sensitive work runs locally through governed MCP tools. Only the result of an allowed tool leaves.
In Mode B, anything a tool returns is disclosed to the external model. The claim is that source documents stay local and administrators choose which derived results are shared. Security posture and data-flow detail live in the Trust Center.
Judged by the people who shipped with it.
Rated 4.8 / 5 across 29 verified SourceForge reviews, quoted verbatim: read every review.
1 / 3
Straight answers, before the call.
The questions every evaluation asks, answered the way we answer them on calls.
What can leave our network?
In fully local mode, nothing. When an external assistant uses LM-Kit through MCP, source documents stay local and only the result of an allowed tool is returned; administrators choose the available tools and so control the disclosure boundary.
What does it run on?
Windows, Linux x64 and ARM64, and macOS. CPU works out of the box; CUDA, Vulkan and Metal are used when present.
Which models can we use?
Curated open-weight models across chat, vision, OCR, embeddings and speech, downloaded once and cached locally. Your own fine-tunes load the same way.
How do AI assistants connect?
Through governed MCP tools on LM-Kit One. The assistant calls a tool, the work happens locally, and only the allowed result is returned.
How is it licensed?
Free below the published company thresholds (revenue, headcount, funding), commercial above them, and one licence covers both products. The licensing section below has the exact terms.
Do we need Python or a GPU cluster?
Neither. LM-Kit is one NuGet in .NET or one server deployment, and a single machine with a modern CPU or GPU is enough to start.
Free below the thresholds. Commercial above them.
One license, both products, no activation key. Nothing checks a license at runtime. Evaluation and development are free at any company size, with no time limit.
Free
$0no key, no expiry
The complete SDK and the complete server, including commercial use and redistribution, for small companies. Evaluation and development stay free at any size.
- Under $1M USD annual gross revenue
- 10 or fewer employees
- No more than $3M USD raised from outside investors
- Always free: personal, education, nonprofits, open source
Professional
Customannual, scaled to scope
Required above the thresholds. Scaled to deployment size, never metered by tokens, seats or end users. Carries the assurance a production deployment needs.
- Commercial use and redistribution at any scale
- Long-term support builds and security patches
- Support with response-time commitments
- Unlimited developers and end users
Get started