Multi-session development
Run multiple AI coding sessions in parallel, allowing you to work on different tasks within the same project.
The local coding agent
AI software engineer
for your local codebase.
- Reads code.
- Runs commands.
- Fixes bugs.
- Ships features.
NoirCode is a coding agent that works in your repository, executes tools, and stays under your control. The terminal, the desktop app, and noircode pair share one server on your machine.
Your repository stays on your machine.
01 About
What is NoirCode?
NoirCode is a local-first AI coding agent that helps you build, debug, and maintain software from your terminal, IDE, or desktop.
Work directly with your codebase, connect your preferred AI models, and control how agents interact with your files, tools, and development environment.
GitHub Copilot integration
Connect your GitHub Copilot account to access supported AI models and development workflows.
OpenAI integration
Connect supported OpenAI services and models using available authentication methods.
Your models, your choice
Work with a broad range of AI providers, including OpenAI, Anthropic, Google Gemini, DeepSeek, Qwen, OpenRouter, and locally hosted models through Ollama or LM Studio.
Work across environments
Use NoirCode through its terminal interface, desktop application, or supported editor integrations.
Local-first architecture
Keep your development environment under your control, with a local server that manages sessions, tools, and permissions.
Agent-driven development
Plan changes, explore repositories, generate code, execute commands, run tests, and review results through configurable agents.
Permission-based execution
Define which files, commands, and tools agents can access, with options to allow, deny, or require approval.
Your codebase. Your models. Your environment. NoirCode.
02 Product
Why NoirCode
The model answers. NoirCode is the part that holds the session, runs the tools, and asks before it touches the machine.
Local-first
The repository stays on disk. Sessions live in a local database. A provider only sees the prompt assembled for that step.
How control worksAny model
Connect Claude, GPT, Gemini, DeepSeek, OpenRouter, or a model you host. /models picks it for the session.
Real tools
Not a chat box. The agent can read files, search the tree, run the shell, and write a patch. Permissions allow, deny, or ask.
PermissionsExtensible
Add agents, plugins, commands, and MCP servers. Themes, keybindings, and permissions live in config you control.
Agents guide03 Session
How it works
You choose the agent and the model. NoirCode does not route the prompt to a different vendor on its own.
- YouA prompt in the terminal, the desktop app, or the web UI.
- Local serverOne process owns the session, the config, and the permission check.
- AgentBuild can edit. Plan will not. Or an agent you wrote.
- ToolsRead, grep, shell, and edit run on the machine after the rules pass.
- Your modelThe provider you connected. A retry repeats that same call.
- The repoA patch in the working tree. Tests run if the shell is allowed.
04 Agents
Agents
A session starts on an agent. Primary agents hold the conversation. Subagents take one task in a child session.
Primary
Build
- Edits project files
- Runs commands
- Can execute tests
Primary
Plan
- Reads the codebase
- Writes the plan, not the code
- Shell still asks first
Subagent
Explore
- Searches and reads
- No file edits
- Used by the primary agent
Yours
Reviewer
- Add
.noircode/agents/ reviewer.md - Deny edit and shell
- Review on the model you name
- Primary agents Build and Plan. Switch between them with Tab, or the key you bind to
switch_agent. - Subagents General and Explore, plus agents you add. The primary agent can launch them, or you can
@mentionone. - Built-in Compaction, title, and summary run on their own. They stay out of the agent picker.
- Use plan when you want a reading of the repo and a plan. Plan does not edit ordinary project files. Shell commands still follow your permissions.
05 Platform
What an agent can reach
The local server is the center. Protocols, skills, and tools attach to it, and permissions still decide what runs.
Server
One process owns the session, the config, and the permission check. The terminal, the desktop app, and noircode pair share it. noircode serve runs that process in the foreground.
MCP servers
Connect a Model Context Protocol server and the agent can use its tools, prompts, and resources. Add only the servers a project needs.
MCPACP support
noircode acp speaks the Agent Client Protocol, so a client such as Zed can run NoirCode. Sign in from a terminal before the client starts.
Agent skills
A skill is a SKILL.md the agent loads when the task matches. Keep scripts and references beside it under .noircode/skills.
Custom tools
A plugin registers a tool with a name, a description, and an input schema. The agent calls it under the same permission rules as read, edit, and shell.
PluginsLSP servers
Language-server settings are kept in config. This release does not start those servers or expose LSP tools. Use the typecheck and lint commands you already allow.
Config notes06 Models
Works with the model you already run
Connect a provider with /connect, or point a baseURL at a server on your network. Ollama and LM Studio are discovered locally. Model availability depends on the provider, your credentials, and the current integration.
Provider directory
- 302.AI
- Amazon Bedrock
- Anthropic
- Atomic Chat
- Azure OpenAI
- Azure Cognitive Services
- Baseten
- Cerebras
- Cloudflare AI Gateway
- Cloudflare Workers AI
- Cortecs
- DeepSeek
- DeepInfra
- DigitalOcean
- Eden AI
- FrogBot
- Fireworks AI
- GitLab Duo
- GitHub Copilot
- GMI Cloud
- Google Gemini
- Google Vertex AI
- Groq
- Hugging Face
- Helicone
- IO.NET
- llama.cpp
- LM Studio
- Moonshot AI
- MiniMax
- Mistral AI
- Modal
- NVIDIA
- Nebius Token Factory
- Ollama
- Ollama Cloud
- OpenAI
- OpenRouter
- LLM Gateway
- OVHcloud AI Endpoints
- Poolside
- SAP AI Core
- Scaleway
- SCX.ai
- Snowflake Cortex
- STACKIT
- Together AI
- Venice AI
- Vercel AI Gateway
- xAI
- Z.AI
- ZenMux
07 Control
Your machine. Your call.
Local server
One backend for the CLI, desktop app, and noircode pair. --standalone keeps a private process.
Permission gates
Allow, deny, or ask for a path or a command. Build already asks before it reads sensitive environment files.
Provider choice
Use a cloud model, or a host you run. A policy can deny every provider except the one you allow.
08 Start
Install. Open the repo. Prompt.
01
Install
02
Run
03
Prompt
Add retry logic to the API client.
09 Release
NoirCode is coming soon
A local-first AI coding environment built for developers who want control over their code, models, and tools.
The NoirCode CLI and desktop application are being prepared for public release on Windows, macOS, and Linux.
Distribution coming soon.
Follow development on GitHub and explore the documentation while the first release is being prepared.
NoirCode — Your code. Your models. Your environment.
The manual
CLI, agents, config, permissions, and the local server API.