Miii

Miii is a privacy-first, local AI assistant that pairs fast streaming chat with pluggable skills, semantic tool routing, and optional live web search — all running on your own machine.

Why Miii exists

Traditional assistants ship every prompt to the cloud. Miii keeps your messages local while still giving you the flexibility of agentic tools, document-aware answers, and a modern UI.

• ⚡ Streamlined UX — browser chat backed by a local server and skills.
• 🔌 Composable intelligence — define JSON skills, chain LangGraph actions, and optionally pull live Tavily search results.
• 🧠 Document-aware context — Chroma-based RAG per conversation with Ollama embeddings.
• 📎 File attachments — attach text, code, or PDF files straight into the composer for one-off context.
• 🛠️ Developer-friendly — streaming NDJSON, model picker, slash commands, and a CLI that installs globally.

Quick start

From source (recommended for contributors)

npm install
npm run dev

The web UI launches at http://localhost:3000, and you can build your prompts, enable tools, and see streaming Markdown responses instantly.

One-line installer (install globally as miii)

curl -fsSL --proto '=https' --tlsv1.2 https://raw.githubusercontent.com/akshaymaru-61/miii/main/scripts/install.sh | bash

After installation you can run miii web or miii help from any terminal. Add ~/.npm-global/bin (or wherever your npm global bin lives) to PATH if commands are missing.

Architecture at a glance

User Input
    │
    ▼
Next.js App Router
    │
    ▼
/api/chat (streaming NDJSON)
    │
    ▼
LangGraph Agent ────┬──► Ollama streaming (no tools)
                    │
                    └──► Agent loop when tools / web search active
                         │
                         ├── Custom skills (`customTools/*.json` → LangChain)
                         └── Tavily web search (live results)

Response streams back to the React renderer.

Stack

Built with Next.js App Router, React 19, Tailwind CSS 4, LangGraph, and Ollama for the local models.

Features

Agentic tool routing

Miii uses LangGraph to decide when to run a tool (custom skill, web search) and when to stream directly from Ollama, giving you the best of both speed and flexibility.

Custom skills

Drop JSON tool definitions into customTools/ and the UI will treat them as LangChain tools. Extend lib/custom-tools.ts to hook in actual execution logic if you need real-world actions.

Example definition:

{
  "name": "my_skill",
  "description": "Explain how this skill works",
  "createdAt": "2025-01-01T00:00:00.000Z"
}

Add, edit, or delete tools via the slash menu (/tools) or the dedicated UI panels.

Optional live web search

Flip on Tavily web search from the UI (or set TAVILY_API_KEY in env). Miii will pull live search snippets before composing an answer so you can stay current despite model cutoffs. Keys stay in the browser unless you share them server-side.

File attachments

Attach text, code, or PDF files from the composer (the paperclip button). Contents are parsed client-side and sent inline with your next message, so you can drop in a snippet or document without a full RAG ingest. Attachment chips show file names, flag truncated files, and can be removed before sending.

Model picker and slash menu

Switch models on the fly via the picker in the browser. The composer supports / commands for clearing chats, managing tools, toggling web search, and more.

Shared memory

The browser client talks to the backend, persists conversations, and streams Markdown (GFM) token-by-token, including tables, code blocks, and lists.

Configuration

Variable	Meaning
OLLAMA_BASE_URL	Override Ollama server URL (defaults to local ollama serve).
TAVILY_API_KEY	Tavily key for server-side web search. Browser keys stay client-local.
CHROMA_URL	Chroma HTTP API (http://127.0.0.1:8000 by default).
CHROMA_API_KEY, CHROMA_TENANT, CHROMA_DATABASE	Optional Chroma defaults.
OLLAMA_EMBED_MODEL	Embedding model for RAG (defaults to nomic-embed-text).

Store overrides in .env.local or configure through the web UI (Chroma and Tavily headers persist in browser local storage).

Running Chroma

The repo bundles chromadb as a dependency. Start the Chroma service from the project root:

npm run chroma

This runs node_modules/.bin/chroma on port 8000. You can also use npx chroma run or point to a remote server with CHROMA_URL.

Slash commands

Send these from the composer for quick actions:

Command	Action
/clear	Clear current conversation messages.
/clear all	Reset the browser to default chats.
/tools	Open Add Tool dialog (custom skills).
/delete-tool	Open Delete Tool dialog.
/web	Open Tavily key dialog.

Custom skills wiring

1. Create or edit a JSON file under customTools/ with name, description, and metadata.
2. Update lib/custom-tools.ts if the tool needs to call real code (e.g., shelling out, querying local data, etc.).
3. Navigate to the Tools menu in the UI or use /tools to activate it for the agent loop.

Custom skills behave like LangChain tools and can be combined with LangGraph reasoning for rich workflows.

Contributing to Miii

We welcome contributors!

1. Fork the repo and open a pull request against main.
2. Run npm run lint and npm run test (if tests exist) before submitting.
3. Propose new features (new tools, UI improvements, docs updates) by opening an issue first if the scope is large.
4. Help triage issues, answer questions, or write docs — every contribution is valuable.

Project layout

Path	Role
app/api/chat	Streaming chat backbone (Ollama + LangGraph).
app/api/models	Lists local Ollama models.
app/api/ollama/pull	Streams ollama pull progress.
app/api/rag/collections	List Chroma collections.
app/api/rag/ingest	Ingest text/files for RAG.
app/api/custom-tools	CRUD for JSON skills.
lib/chat-graph.ts	LangGraph agent setup.
lib/chat-tools.ts	Tool merging logic.
lib/custom-tools.ts	Tool definitions loader.
lib/rag-chroma.ts	Chroma client helpers.
lib/tavily-tool.ts	Tavily web search integration.
lib/attachments.ts	Parse attached text/code/PDF files for the composer.
lib/messages.ts	System prompts.
lib/conversation-storage.ts	Local storage helpers.
customTools/	JSON skill definitions.
components/chat/	Browser chat client.

Roadmap highlights

• Persistent skill workers (real execution logic behind custom tools).
• ChromaDB semantic skill retrieval (auto tool selection via embeddings).
• Document ingestion (PDFs, notes, and local files) through the UI.
• Multi-turn memory across conversations.
• Docker Compose for single-command bootstrapping.

Need help?

• Ask questions in the repository discussions or open an issue with the support label.
• Bring feature ideas and bug reports to GitHub — we track them in Issues.

License

Miii is open source under the MIT License.