GitHub Repository

Declarative AI-native application framework — schema-driven entities, Markdown skills, MCPB packaging

9 starsPython

Upjack – Declarative framework for building apps over MCP

Name: Upjack – Declarative framework for building apps over MCP
Availability: InStock
Author: barefootsanders

by barefootsanders·Feb 25, 2026·1 point·0 comments

Visit Project View on HN

AI Analysis

●●SolidBig BrainShip It

Schema-to-server codegen for MCP, but targets the crowded AI app layer.

Strengths

•Removes boilerplate by auto-generating CRUD tools, validation, and search from declarative schemas
•Works across Claude Desktop, Claude Code, and any MCP client without framework lock-in
•Event hooks and scheduled jobs make it useful beyond CRUD—real state machine logic possible

Weaknesses

•MCP app frameworks are multiplying fast; unclear differentiation from Anthropic's own examples or other MCP boilerplate generators
•JSON file storage won't scale past hobby projects; no migration path to real databases

Post Description

Open-sourced Upjack today.

I built three apps with it: a CRM, a research assistant, and a todo app. Each one started the same way. I told Claude Code what I wanted. The app-builder skill generated schemas, domain skills in Markdown, a server, seed data. I pointed Claude Desktop at it. Working app, running locally.

The CRM has five entities (contact, company, deal, pipeline, activity) and three skills (lead qualification, deal forecasting, follow-ups). Hooks fire on events (new contact -> auto-score). Scheduled jobs run on cron (pipeline review weekdays 9 AM). Named views filter data (hot leads: score >= 80).

The research assistant has four entities (topic, source, note, report) and a synthesis skill that describes how to evaluate sources and cross-reference findings. Completely different domain, same framework.

Apps are MCPB bundles. They run in Claude Desktop, Claude Code, Codex, or any MCP client. Any data app you can describe, you can build.

The idea: we've always defined data models for developers to code against. LLMs don't need that. They reason over JSON Schema directly. They read Markdown and apply it. Give them a well-defined schema and domain rules, they do the rest.

My sales lead wrote the CRM's qualification rubric. Not a developer. C-suite: +25 pts, corporate email: +10. The agent reads it and follows it.

Storage is flat JSON files, one per record, backed by git. Grep your CRM, diff a contact, branch your pipeline. The storage layer is pluggable. Flat files today, other backends as the project grows.

Tools are generated from schemas, six per entity. Built on FastMCP in Python, TypeScript library too. The server:

from upjack.server import create_server mcp = create_server("manifest.json") mcp.run()

Three tiers: (1) schemas + skills, no code. (2) create_server() one-liner. (3) Python/TypeScript escape hatch for custom logic.

Looking for developers and businesses who want to experiment with AI-native applications. If you're interested in collaborating, open an issue or ping me directly.

https://github.com/NimbleBrainInc/upjack Docs: https://upjack.dev