🛰️LM-Kit.NET Goes Agentic: MCP, Tools, Orchestration, All On-Device

We Built What They Said Couldn't Be Done

Let's be direct: the AI industry told .NET developers they had two choices. Use cloud APIs and accept the trade-offs (latency, costs, data exposure, vendor lock-in). Or settle for limited local capabilities that couldn't match production requirements.

We refused to accept that.

LM-Kit.NET exists because a team of engineers with decades of experience in high-performance systems, SDK development, and enterprise software decided to solve the hard problems nobody else would tackle. We built a complete local AI stack from the ground up: native inference engines, optimized backends for CPU/GPU/Hybrid execution, intelligent document processing pipelines, semantic RAG with built-in vector storage, and now, a comprehensive agentic framework.

This isn't a wrapper around someone else's Python library. This isn't bindings to a foreign runtime. This is a native .NET SDK architected by engineers who understand what enterprise software demands: reliability, performance, observability, and complete control.

The Local Agentic Era

Late 2024 marked a turning point. Anthropic introduced the Model Context Protocol (MCP) in November 2024, establishing an open standard for connecting AI agents to tools and data. Throughout 2025, the ecosystem accelerated rapidly: thousands of MCP servers emerged, and adoption spread across the industry.

But here's what most people missed: the real revolution isn't just about agents. It's about where those agents run.

Cloud-based agents have fundamental limitations:

• Latency: Network round-trips add hundreds of milliseconds per interaction
• Cost: Per-token pricing means expenses scale unpredictably with usage
• Privacy: Your data, your users' queries, your business logic, all flowing through third-party servers
• Availability: Rate limits, outages, and API deprecations outside your control
• Compliance: HIPAA, GDPR, air-gapped requirements become architectural nightmares

LM-Kit.NET is the platform that makes this possible for .NET developers. Not someday. Today.

A Unique Position in the .NET Ecosystem

There is no other solution that delivers what LM-Kit.NET delivers. Most AI toolkits force you to stitch together disparate components: one library for inference, another for embeddings, a third for vector storage, a fourth for document processing, a fifth for orchestration. Each with different APIs, different dependencies, different failure modes.

LM-Kit.NET provides everything in one SDK:

Intelligent Document Processing

Enterprise AI isn't just chat. It's processing invoices, extracting entities from contracts, analyzing reports, managing knowledge bases. LM-Kit.NET treats document intelligence as a first-class capability:

• Structured Extraction: Pull typed data from unstructured documents with JSON schema validation
• Named Entity Recognition: Custom and pre-built entity types
• PII Extraction: Identify and handle sensitive information
• Classification: Categorize documents at scale
• Summarization: Condense long documents intelligently
• VLM-powered OCR: Extract text from images and scanned documents using vision models

Intelligent Information Management

RAG isn't a feature we bolted on. It's a core architectural pillar: semantic chunking, hybrid search, reranking, agent memory, and multi-source RAG across different storage backends.

Enterprise-Grade Engineering

We don't ship demos. We ship production software: 100+ pre-configured model cards, cross-platform support (Windows, macOS, Linux including ARM64), hardware acceleration (CUDA, Vulkan, Metal), .NET Framework 4.6.2 through .NET 10 compatibility, MAUI support, and OpenTelemetry instrumentation built in from day one.

What's New: The Agentic Framework

Today's release represents a massive expansion of LM-Kit.NET's capabilities. We're delivering a complete framework for building autonomous AI agents that operate locally with full enterprise requirements in mind.

Complete MCP Client Implementation

Our McpClient supports the full Model Context Protocol specification: tools, resources, prompts discovery and invocation, sampling, roots, elicitation, progress tracking, cancellation, logging, completions, resource templates, and subscriptions. Transport options include HTTP/SSE for remote servers and Stdio for local subprocess servers.

  

    
    
    
  
  C#

// Connect to any MCP server
var client = await McpClient.ForStdio(new StdioTransportOptions
{
    Command = "npx",
    Arguments = ["-y", "@modelcontextprotocol/server-github"]
});

// Import the entire tool catalog
chat.Tools.Register(client);

Your local agents now have access to thousands of community MCP servers while keeping all inference on-device.

Agent Framework Core

Build agents with clear abstractions: Agent, AgentBuilder, AgentExecutor, AgentRegistry, AgentIdentity, AgentCapabilities, and AgentExecutionOptions.

  

    
    
    
  
  C#

var agent = new AgentBuilder()
    .WithModel(model)
    .WithIdentity(new AgentIdentity
    {
        Name = "DocumentAnalyst",
        Description = "Extracts insights from enterprise documents",
        Instructions = "Focus on actionable findings..."
    })
    .WithCapabilities(c => c
        .AddTools(documentTools)
        .AddSkills(extractionSkills)
        .EnableMemory(agentMemory))
    .Build();

var result = await agent.ExecuteAsync("Analyze the Q4 contracts for risk factors");

56 Built-In Tools

Production-ready tools with no external dependencies:

  

    
    
    
  
  C#

chat.Tools.Register(BuiltInTools.All());

Five Planning Strategies

Four Orchestration Patterns

Compose multiple agents with PipelineOrchestrator (sequential), ParallelOrchestrator (concurrent), RouterOrchestrator (dynamic routing), and SupervisorOrchestrator (coordinated delegation).

  

    
    
    
  
  C#

var pipeline = new PipelineOrchestrator()
    .AddStage(extractorAgent)
    .AddStage(analyzerAgent)
    .AddStage(reporterAgent);

18 Agent Templates

Start fast with pre-configured specialists:

Agent Skills Protocol

Reusable task specialists with progressive loading:

  

    
    
    
  
  C#

var registry = new SkillRegistry();
registry.RegisterSkill("/skills/invoice-processor/SKILL.md");

// Keyword and semantic matching
var matches = await registry.FindSemanticMatchesAsync(query, embedder);

// Slash command invocation
// User: /invoice-processor analyze attached document

Resilience and Observability

Resilience Policies: Retry, CircuitBreaker, Timeout, RateLimit, Bulkhead, Fallback, Composite

OpenTelemetry GenAI Conventions: Token usage histograms, operation duration metrics, agent and tool span attributes, conversation correlation

We're Just Getting Started

What you see today is the foundation. Our ambition goes far beyond this release.

We believe Local Agentic AI will transform how enterprises build intelligent systems. Not by replacing human judgment, but by augmenting it with autonomous capabilities that respect privacy, deliver predictable costs, and operate under complete organizational control.

Our roadmap:

• LM-Kit Server REST API: Expose your local agents via HTTP endpoints (coming in weeks)
• Advanced multi-agent collaboration: Sophisticated team dynamics for complex workflows
• Persistent agent checkpointing: Pause and resume long-running agent tasks
• Visual workflow designer: Build agent pipelines graphically
• Expanded model support: More architectures, more optimizations, more capabilities

Let's Talk

Here's our promise: we achieve what others consider impossible.

If you have a use case that requires local inference, intelligent document processing, semantic RAG, or agentic orchestration, and you've been told it can't be done without cloud dependencies, talk to us.

We've spent years solving the hard problems in native AI inference and enterprise document intelligence. We understand the constraints of real-world deployments: compliance requirements, latency budgets, cost pressures, integration complexity.

Get Started Today

Install:

  

    
    
    
  
  Shell

dotnet add package LM-Kit.NET

Build your first local agent:

  

    
    
    
  
  C#

using LMKit.Model;
using LMKit.Agents;
using LMKit.Agents.Tools.BuiltIn;
using LMKit.Agents.Templates;

var model = LM.LoadFromModelID("gptoss:20b");

var agent = AgentTemplates.Assistant()
    .WithModel(model)
    .WithTools(BuiltInTools.All())
    .Build();

var result = await agent.ExecuteAsync(
    "Extract all vendor names and amounts from the attached invoices, " +
    "then calculate the total payable this month."
);

Resources:

Agentic Samples:

PS: We're forming Calico IIM Group, a new entity positioned to accelerate our mission with significant investment. More details coming soon.