MLC Artifact Service

License: MIT Platforms: Windows, macOS, Linux

Don't route large data through the LLM — let MCP servers exchange artifacts by ID. Built for local LLM setups.

Download Source on GitHub →

// Features

Why choose mlcartifact.

Virtual File System (VFS)

Hierarchical path-based storage — /projects/report.md instead of flat blob IDs.

Surgical Patching

Line-based partial updates and appending — no expensive re-uploads of large artifacts.

Multi-Protocol Access

MCP 2026-07-28 (stdio, Streamable HTTP, SSE), gRPC/Connect and a dedicated CLI — one store, every client.

TTL-based Cleanup

Automatic expiry keeps the store lean. Explicit deletion also supported.

Multi-Language Clients

Official clients for Go, TypeScript, Python, and Rust, source on GitHub — drop-in for any MCP server.

Lightweight & Self-Hosted

Single Go binary, Docker-ready, no external dependencies.

// The Shared Memory Layer for MCP Ecosystems

The Shared Memory Layer for MCP Ecosystems

MCP Server A writes a 2 MB CSV to mlcartifact and returns only the artifact ID to the LLM. MCP Server B reads it directly — the large data never flows through the context window.

Only IDs flow through the LLM — not the data itself
Virtual File System with path-based lookup
Surgical line-based patching for large files

The Shared Memory Layer for MCP Ecosystems

Persistent Memory for the Agentic Era

MLC Artifact Service is more than just a storage backend. It provides a structured, hierarchical Virtual File System (VFS) designed specifically for AI agents, distributed tools, and complex automation workflows.

Why use MLC Artifact Service?

In many agentic systems, data is passed back and forth as large strings or flat files. MLC Artifact Service introduces a stable, path-based workspace:

Hierarchical VFS: Organize artifacts into directories like /projects/alpha/src/main.py.
Token Efficiency: Instead of re-uploading a 100KB file for a 1-line change, agents use VFSPatch for surgical edits.
Protocol Agnostic: Whether you use MCP, gRPC, or the CLI, your data is accessible and consistent.
Smart Lifecycles: Automatic TTL management ensures your storage doesn't clutter while allowing critical data to persist.

Who Is It For?

mlcartifact is primarily meant for local LLM setups, for example via Ollama, llama.cpp or LM Studio, and for custom harnesses. These usually have no built-in way for tools to hand files to each other: every intermediate result would have to pass through the model's context, which is exactly what is scarce and slow on local hardware.

Professional platforms such as Claude (Anthropic) or Gemini (Google) usually ship a comparable mechanism of their own, such as a sandbox with a file system or a files API. Large files stay outside the model's context between tool calls and never have to travel over the network to the model and back. If you use such a platform, check its built-in solution first. There, mlcartifact mainly pays off when your own MCP servers should exchange data directly through a shared store.

Key Tools

vfs_ls & vfs_find

Navigate your virtual workspace with full glob support and directory-style listing.

vfs_patch

Perform surgical updates. Append logs, replace specific code blocks, or update JSON metadata objects without high overhead.

Integrated MCP Prompts

The service comes with its own usage guidelines to "train" agents on how to best manage their persistent memory.

Getting Started

Download the binary for your platform from the releases.
Start the server: ./artifact-server
Connect via MCP to use it with your favorite LLM or via gRPC for your custom backend.

The server speaks MCP protocol 2026-07-28 over stdio, Streamable HTTP (/mcp) and, for older clients, SSE (/sse). Start it with -addr :8082 for HTTP mode.

Client Libraries

To let your own MCP servers write and read artifacts directly, there are clients for Go, TypeScript, Python and Rust. All of them cover the full API, including virtual paths, patch and find. Source code, examples and guides are in the GitHub repository:

Go: go get github.com/hmsoft0815/mlcartifact/client (guide)
TypeScript (Node, browser, edge): client-ts
Python: client-python
Rust: client-rust

Quick Setup (MCP)

Claude Desktop

Add the following to your claude_desktop_config.json:

{
  "mcpServers": {
    "mlc-artifact": {
      "command": "artifact-server"
    }
  }
}

Gemini-CLI

Add to your ~/.gemini/settings.json:

{
  "mcpServers": {
    "mlc-artifact": {
      "command": "artifact-server"
    }
  }
}

MCP-Tester

Add a new profile:

mcp-tester profile add artifact -c "artifact-server"