Open-source AI coding agent built in pure C#/.NET.
⭐ Star | 🚀 Quick Start | 📖 Architecture | 🌐 litosai.dev
A minimal, transparent AI coding agent written in pure .NET. Multi-provider (Anthropic, OpenAI, Google Gemini, OpenRouter), multi-face (a VS Code extension, a desktop GUI, and a REST API with web UI), with tool use (file edit, shell, web search) and MCP client support.
Built to learn how a coding agent actually works, from the inside, in C#/.NET — the stack I've worked in for 23+ years — rather than treating one as a black box.
Primarily inspired by Mario Zechner's pi, particularly its session-as-working- directory model, and secondarily by alejandro-ao's Tau and its "separate the brain, the environment, and the face" architecture.
Litos is one agent core with several interchangeable faces. These three are working end to end today — pick whichever fits how you work; all talk to the same providers, tools, and MCP servers.
Chat with Litos in a docked panel inside VS Code, with file/shell/web tools and MCP support, without leaving your editor. Works on Windows and macOS.
Install from the
VS Code Marketplace
(search for "Litos" in the Extensions view, publisher litosai), or run:
ext install litosai.litos-vscode
The extension bundles a small local backend it manages for you — no separate service to run. On
first use, if no provider API key is found, it prompts you to enter one (or use /keys).
A standalone desktop chat window (Avalonia), for when you'd rather not tie the agent to an editor window. Works on Windows and macOS.
Windows (PowerShell):
irm https://raw.githubusercontent.com/nitinmms/LitosAiCodingAgent/master/deploy/install.ps1 | iexInstalls to %LOCALAPPDATA%\Programs\Litos, adds it to your PATH, and creates a Start Menu
shortcut. No admin rights needed. The build is unsigned, so SmartScreen may warn on first launch
— click More info → Run anyway.
macOS (Terminal):
curl -fsSL https://raw.githubusercontent.com/nitinmms/LitosAiCodingAgent/master/deploy/install.sh | bashInstalls Litos.app to /Applications (Apple Silicon and Intel both supported). This build is
signed and notarized, so Gatekeeper won't block it.
Re-running either command upgrades an existing install in place. On first launch, if no provider
API key is found, Litos opens a dialog to enter one — see
Environment variables below. Keys can be added or changed later from
inside the app with /keys.
Litos running as a headless REST API, for building your own clients or integrations against the agent, or running it self-hosted / in Docker as a service. It ships with a web UI and a Telegram bridge as two ready-made clients — Telegram is just one way in; any client can talk to the same Agent API. See Getting started below to run it from source, and ReadMe_HeadlessServiceTool.md for Docker/deployment.
- Built an AI Coding Agent in Pure C#/.NET (No Python!) — Watch It Build a WinForms 15-Puzzle Game
- Built an AI Agent in Pure C#/.NET— Here it is working through Telegram creating Mermaid Diagrams
- Litos Coding Agent Building Features into an Existing Postgres Query Application
Slash commands are typed directly into the chat input. Each face (Litos.VsCode, Litos.Gui,
and Litos.Api's Telegram bridge) implements its own command dispatch, so support varies:
| Command | Description | VsCode | Gui | Api |
|---|---|---|---|---|
/new |
Start a new session | Yes | Yes | Yes |
/resume [id] |
Resume a previous session (picker if no id given) | Yes | Yes | Yes |
/attach <path|url> |
Attach a file, folder, or URL to the conversation | Yes | Yes | No |
/provider [name] |
Switch the active LLM provider | Yes | Yes | No |
/model [id] |
Switch the active model | Yes | Yes | No |
/skills |
List available skills | Yes | Yes | Yes |
/skill <name> |
Load a skill by name | Yes | Yes | Yes |
/branch <msgIndex> |
Rewind the session to an earlier message | Yes | Yes | Yes |
/compact |
Force-summarize/compact the conversation to free context | Yes | Yes | Yes |
/reflect |
Distill the session into an AGENTS.md |
Yes | Yes | No |
/mcp |
Manage MCP servers (add/browse/run prompts) | Yes | Yes | No |
/keys |
Add or update provider API keys | Yes | Yes | No |
Litos.Gui also supports dynamic /mcp__<server>__<prompt> commands for running prompts exposed
by a connected MCP server; Litos.VsCode doesn't yet surface these as commands (its /mcp panel
still covers full server management).
Three faces are working end to end today, on both Windows and macOS:
Litos.VsCode— the VS Code extension. Published on the Marketplace, with near-full slash-command parity withLitos.Guiand MCP server support (add/manage servers, browse their tools, run their prompts). Linux isn't supported yet. See ReadMe_VsCodeExtension.md.Litos.Gui— the Avalonia desktop app. Includes MCP server support (add/manage servers, browse their tools, run their prompts).Litos.Api— the AI agent exposed as a REST API, with a web UI, Docker support, MCP server support, and a working Telegram bridge as one example client. Other clients can connect to the same Agent API directly.
The WhatsApp/Rocket.Chat/email channel bridges are a work in progress — those three bridges currently exist only as design docs, not shipped features.
Of the four LLM providers, only OpenAI and OpenRouter have been explicitly exercised end-to-end. Anthropic and Google Gemini adapters exist and should work, but are unconfirmed.
Already installed Litos.VsCode from the Marketplace or Litos.Gui via the one-line command
above? Just launch it. The rest of this section is for building from source or running Litos.Api.
Requires the .NET 10 SDK.
git clone https://github.com/nitinmms/LitosAiCodingAgent.git
cd LitosAiCodingAgent
dotnet buildSee ReadMe_VsCodeExtension.md for building and debugging the
extension and its bundled Litos.VsCodeHost backend from source.
dotnet run --project src/Litos.Guicp src/Litos.Api/.env.example src/Litos.Api/.env # fill in your provider API key(s)
dotnet run --project src/Litos.ApiThen open http://localhost:8080 and sign in with the ADMIN_TOKEN you set in .env. For
Docker/Telegram deployment, see ReadMe_HeadlessServiceTool.md and
ReadMe_TelegramIntegrationTool.md.
At least one LLM provider API key is required. Configure these via the .env file (Litos.Api,
copied from src/Litos.Api/.env.example) or, for Litos.VsCode and
Litos.Gui, the first-run dialog or /keys command (writes a Windows user environment variable
on Windows, or ~/.litos/config.json on macOS). An env var always wins over any value stored in
Litos.Api's mounted config file or Litos.VsCode's/Litos.Gui's config.json.
| Variable | Purpose |
|---|---|
ANTHROPIC_API_KEY |
Anthropic (Claude) provider |
OPENAI_API_KEY |
OpenAI provider |
GEMINI_API_KEY (or GOOGLE_API_KEY as fallback) |
Google Gemini provider |
OPENROUTER_API_KEY |
OpenRouter provider |
TAVILY_API_KEY |
Web search tool (optional — not a chat provider) |
ADMIN_TOKEN |
Login password for Litos.Api's admin UI (Bearer/cookie) |
TELEGRAM_BOT_TOKEN |
Bot token from @BotFather; enables the Telegram bridge (also settable from the /telegram admin page — the env var always wins). Off by default even when set. |
dotnet test- ReadMe_Architecture.md — brain/environment/face layering, project list, and how context is assembled each turn
- ReadMe_AgentDesign.md — full architecture and design rationale
- ReadMe_Extensibility.md — adding providers, tools, and faces
- ReadMe_VsCodeExtension.md —
Litos.VsCodeextension design, Marketplace publishing, and status - ReadMe_HeadlessServiceTool.md — running
Litos.Apias a headless/Docker service, admin UI - ReadMe_TelegramIntegrationTool.md — Telegram bridge (
Litos.Api) - ReadMe_LitosApi_Mcp.md — MCP server integration in
Litos.Api - ReadMe_MCPSupportInLitosGUI.md — MCP server integration in
Litos.Gui - ReadMe_DeployToAWS.md — AWS deployment guide
Design docs only — no implementation exists yet, read as proposals:
- ReadMe_WhatsAppIntegrationTool.md — WhatsApp bridge
- ReadMe_RocketChatTool.md — Rocket.Chat bridge
- ReadMe_EmailIntegrationTool.md — Email integration
Apache License 2.0 — see LICENSE.txt. Third-party dependency licenses are listed in NOTICE.