LM-Kit One · Coding agents

Your coding agent, your server.

Coding assistants see your whole repository. Point them at LM-Kit One and every prompt, diff and file stays on infrastructure you control.

Anthropic and OpenAI dialects Guided Claude Desktop setup Free to build and evaluate

Claude Desktop gets a guided lane.

The server ships a setup wizard that provisions the gateway, sizes the serving profile, and connects the app step by step.

Wizard

Plan, apply, connect, verify

Each step explains what it changes and why. Nothing is written to configuration without an explicit action.

Sized right

Serving profile, audited

Desktop assistants carry very large system prompts, so the server audits its context and slot shape against what the client actually needs and flags a misfit before you hit it.

Governed

Same boundaries as any caller

The assistant arrives through identities, policies and audit like every other client, and its traffic shows up attributed in the console.

Four clients, one endpoint swap.

Every major coding agent accepts a custom endpoint; here is the exact knob for each.

Claude Code follows the standard Anthropic environment overrides.

terminal

export ANTHROPIC_BASE_URL=http://your-server:5189
export ANTHROPIC_AUTH_TOKEN=your-server-token
claude

Continue takes an OpenAI-compatible block in its configuration.

config.yaml

models:
  - name: LM-Kit One
    provider: openai
    model: qwen3.5:9b
    apiBase: http://your-server:5189/v1
    apiKey: your-server-token

Cline connects through its OpenAI Compatible provider.

Cline settings

API Provider : OpenAI Compatible
Base URL     : http://your-server:5189/v1
API Key      : your-server-token
Model ID     : qwen3.5:9b

Zed overrides the OpenAI endpoint in its settings file.

settings.json

{
  "language_models": {
    "openai": {
      "api_url": "http://your-server:5189/v1"
    }
  }
}
// Then add a model from your server's catalog in Zed's model picker.

MCP-capable assistants can additionally mount the server's governed document and knowledge tools from /mcp; which tools are reachable stays an operator decision.

Sized for the job, stated up front.

Two realities decide whether a local coding agent feels great or terrible.

Context

Agent prompts are enormous

A coding assistant's system prompt and tool schemas alone can exceed tens of thousands of tokens. The serving profile reserves a wide context window per chat slot, and the console prices that against your hardware before you commit.

Model

Pick a coding-capable model

Agentic coding rewards strong tool-calling models; the catalog marks them. Evaluate on your own repository, and size the model to your GPU rather than hoping.

LM-Kit One

Keep the repository where it belongs.

Download LM-Kit One See the coverage matrix