Artifacta MCP Server is the official MCP server for Artifacta. Artifacta is an artifact storage service designed specifically for AI agents, used to persist, list, and hand off run outputs, with session metadata, deduplication, and expiring share links. The project provides both TypeScript and Python implementations, with identical tool surfaces, error contracts, and path restriction engines.
Purpose and Positioning
This server is aimed at AI agent scenarios, helping agents save run artifacts (files, reports, datasets, build results, etc.) and supporting cross-session handoff and sharing. Content-hash deduplication means that re-storing identical bytes incurs no additional cost.
Main Capabilities
- Persist, list, and hand off agent run outputs.
- Include session and agent metadata.
- Content-hash deduplication.
- Share artifacts via download links with expiration times.
- Provide MCP resources, including whoami, artifact metadata, artifact bytes, and sessions.
- Secure defaults: local file uploads are restricted to an explicit --allow-path allowlist; destructive tools (public share links, deletion, session sealing) are hidden from clients whose writes cannot be confirmed unless --allow-destructive is passed. publish_artifact and unpublish_artifact are idempotent write operations rather than destructive operations and are not gated by --allow-destructive.
- Managed OAuth connections (mcp.artifacta.io) add a second layer of control: the consent screen grants one of three scopes—artifacts:read ⊆ artifacts:write ⊆ artifacts:destroy. The tool list always shows all 13 tools; calling a tool for which permission has not been granted returns an error with code insufficient_scope and indicates the missing scope, and the resolution is to reauthorize with a broader scope. Scope gating applies only to managed OAuth; ak_live_ API keys and local stdio retain full access, using --allow-destructive / confirmation flags instead.
Usage
Install as a Claude Code plugin
For Claude Code, the fastest path is to use the plugin marketplace that this repository also serves as—it will configurethe managed server as well as a skill that automatically persists run outputs:
This bundles the same managed MCP connection as the quick start below, plus the persisting-outputs skill (/artifacta:persisting-outputs, or automatically triggered when there are outputs worth saving in the session). Use /plugin marketplace update artifacta to update the plugin. See the plugin setup guide.
Quick Start (Managed Server)
The fastest way to connect is the managed server—no installation, no API key:
On first use, the client self-registers via OAuth Dynamic Client Registration (PKCE) and opens a browser for authorization—no need to copy or store an ak_live_ key. See the managed setup guide.
Local / CI (stdio)
For headless, air-gapped, or CI environments where the browser OAuth flow is unavailable, you can run the package locally via stdio using an API key. Get a key at app.artifacta.io/dashboard/keys, then add it to your MCP client configuration (Claude Desktop, Claude Code, Cursor, or any MCP client):
Or run the Python implementation with pipx run artifacta-mcp.
The README for each package contains configuration file setup, path restrictions (--allow-path), destructive tool gating (--allow-destructive), and troubleshooting: TypeScript · Python.
Framework Integrations
The Python package ships with optional adapters for the OpenAI Agents SDK (pip install 'artifacta-mcp[openai-agents]') and LangChain/LangGraph (pip install 'artifacta-mcp[langchain]').
Full documentation is at docs.artifacta.io/mcp/overview.
This repository is published from the Artifacta monorepo; issues and PRs are welcome here.
MIT — see LICENSE.
