fullstack-ai-agent-roadmap: a 12-to-18-month Chinese curriculum that ends at mini-agent-sdk

A structured curriculum in a vault, not another awesome list

The repository's own contrast is with link collections. What Karovia/fullstack-ai-agent-roadmap ships is a curriculum: 110 Markdown tutorials totalling about 580,000 Chinese characters, organized as an Obsidian vault rather than a rendered docs site, aimed at learners starting from zero. The README is explicit about the audience and the cost: no prerequisites, but three hours a day plus persistence, with the overall cycle given as 12 to 18 months. A pace table breaks that down further, from 8 to 10 months at five-plus hours daily to 18 to 24 months for someone with a full-time job at 1.5 hours a day. That framing matters for adoption. This is not a reference you consult during work; it is a study commitment closer to a part-time degree, and every design choice in the repository, from per-chapter verification projects to learning contracts, assumes you are following it in order.

The module ladder ends at mini-agent-sdk and a self-built MCP Server

Ten numbered modules carry the structure. Module 00 covers learning methodology in one week. Modules 01 and 02 spend ten weeks each on Python and JavaScript, ending in a published PyPI package called pyspider-cli (an async crawler) and a hand-written mini-lodash with an online sandbox. Module 03 is four weeks of HTML and CSS with a pixel-level rebuild of the Linear, Stripe and Apple landing pages. Module 04 reaches React, and its capstone pairs an online collaborative whiteboard with a hand-written mini-react implementing Fiber and Hooks. Module 05 covers FastAPI and Node through the RealWorld Conduit backend plus real-time chat, module 06 builds a high-concurrency mini-twitter on top of database and engineering practices, and module 07 implements GPT from zero alongside an AskMyDocs enterprise document QA project. The starred module 08 runs 12 weeks on agent frameworks and source code, finishing with mini-agent-sdk and a self-built MCP Server, and module 09 is an 8-to-12-week graduation project the README describes as something that can launch, earn and go on a resume.

Obsidian is the delivery mechanism, with Canvas as the map

The recommended path is to open the repository as an Obsidian vault. The quick start is concrete: install Obsidian, git clone https://github.com/Karovia/fullstack-ai-agent-roadmap.git, open the folder as a vault, start with the Canvas mind map file for the global view, then begin at 全栈学习路线/README.md. The Canvas file is interactive only inside Obsidian, which makes the tool a soft dependency rather than an option if you want the roadmap graph. Browsing on GitHub works too, and the README claims the Chinese filenames render fine there. Beyond reading, the vault ships what it calls a forced feedback system: templates for a learning contract, weekly reports and milestones. Together with the rule that a chapter is not finished until its verification project is done, the templates turn the vault into something closer to a course with homework than a wiki, and that is the real mechanism of the project, more than any single tutorial.

Curated GitHub projects at three depths, with a how-to-learn note

Each module ends with a curated GitHub project list, over 400 projects across the roadmap, and the curation has a stated method. Projects are tagged at three depths, entry-level items meant to run in 30 minutes, intermediate items read for their design, and source-level items showing real production code, plus a fourth Easy-vibe category for small but complete projects. Four deep-dive supplements extend this: a Chinese companion guide to build-your-own-x, an index of awesome lists, a teardown collection of real open-source SaaS products, and the 30-minute starter set. The difference from a plain awesome list is the stated intent: every entry comes with guidance on how to use it for learning, not just a link. From the material supplied, the actual quality of these annotations cannot be verified, but the structure itself, depth labels attached to each chapter, is more usable than the undifferentiated lists it positions itself against.

Read the outcome table as the author's claims, not guarantees

The README's learning-outcome table is where skepticism belongs. It promises production-grade Python and TypeScript without framework dependence, an independently launched web app, the ability to design and deploy multi-agent systems, reading React, LangGraph and parts of CPython source, full engineering practices around Docker, CI/CD, monitoring and testing, job offers as an AI Agent engineer, and a GitHub project with 100-plus stars. None of that is verifiable from the repository, and some of it, like star counts and offers, is outside any curriculum's control. The footer states the project was made in Obsidian and powered by Claude, so a learner should assume substantial AI assistance in the writing and calibrate trust accordingly, sampling chapters early rather than assuming uniform depth. Maintenance looks present but thin: the last push was 2026-09-07, four days before this review's reference date, there are no releases, and the project reads as a single-author effort with an issues tracker and discussions enabled for feedback.

CC BY-SA 4.0 declared in prose, absent from the metadata

Licensing needs one careful look. The README declares the content under CC BY-SA 4.0, with free sharing, adaptation and commercial use permitted, provided the original author Karovia is credited and derivative works carry the same license. The repository metadata, however, reports no detected license, which usually means no LICENSE file at the root. Before redistributing or adapting the tutorials, confirm the declaration with the maintainer or wait for an actual license file; the share-alike clause in particular means a translated or reworked derivative must itself be CC BY-SA 4.0, which affects any team wanting to fold these chapters into internal proprietary training. For learners rather than reusers this is moot. Alternatives named in the material itself are narrower but English-first: build-your-own-x, which lists self-build tutorials without a sequenced 18-month path or feedback system, and Anthropic's courses repository, which teaches agents directly without the Python-to-React foundation. Datawhale's ecosystem, credited in the acknowledgements, covers similar Chinese-language ground module by module rather than as one integrated vault.

Editorial conclusion

Chinese-speaking beginners who can commit roughly three hours a day for over a year should clone the vault, open the Canvas mind map, and work through modules 00 and 01 before judging the rest. It is the wrong choice for English-only learners, since only the README is translated while all 110 tutorials are Chinese, and for engineers who already ship production code, who should jump straight to module 08 and its mini-agent-sdk plus MCP Server capstone instead of starting from Python. Verify two things first: whether an actual LICENSE file matching the README's CC BY-SA 4.0 declaration exists, and whether the early chapters' teaching style suits you, because the repository states it was built with Claude and the outcome promises, from job offers to a 100-star project, are the author's claims rather than anything a reader can confirm from the material.

Official sources

1. Issues
2. Karovia/fullstack-ai-agent-roadmap on GitHub
3. README