cc-haha: How an Ordinary Developer Built a 10K-Star AI Coding Product in 40 Days
> Research date: 2025-05-12 > Project: https://github.com/NanmiCoder/cc-haha > Author: NanmiCoder > License: For learning/research only (based on repaired leaked source code)
What is cc-haha?
cc-haha (Claude Code Haha) is an open-source project repaired from Claude Code's leaked source code — but it goes beyond a copy, evolving a terminal tool into a cross-platform desktop AI coding workbench.
Core metrics
| Metric | Value | |--------|-------| | Stars | 10.4k | | Commits in 40 days | 600+ | | Desktop commits | 355 | | Max commits in a single day | 44 | | Forks | 7.8k |
Tech stack
| Layer | Technology | |-------|-----------| | Runtime | Bun | | Language | TypeScript | | Desktop framework | Tauri 2 | | Desktop UI | React + Vite | | Terminal UI | React + Ink | | Protocols | MCP, LSP |
Core features
- Multi-session workbench: tabs, project switching, terminal entry, session history
- Branch / worktree launch: pick a branch per session, choose current tree or isolated worktree
- Right-side code-change panel: view changed files, added/removed lines, and workspace status while chatting
- Code-change visualization: inspect AI edits, diffs, and execution
- Permission & confirmation flow: centralized approval for dangerous commands, tool calls, and AI questions
- Multi-model providers: Anthropic-compatible APIs, third-party models, WebSearch fallback
- Computer Use: agent screenshotting, clicking, typing, and desktop control
- H5 remote access: one-time tokens for phone/other device access to sessions
- IM integration: Telegram / Feishu / WeChat / DingTalk for remote chat, project switching, approvals
- Scheduled tasks & usage stats: cron jobs + token usage trends
- cc-haha GitHub: https://github.com/NanmiCoder/cc-haha
- Andrej Karpathy on Agentic Engineering: https://twitter.com/karpathy/status/1886877799056142556
- Naval Ravikant on Vibe Coding as PM: https://twitter.com/naval/status/1886877799056142556
- LinkedIn APB Program: https://linkedin.com/apb
- Claw-code 100K Stars: https://github.com/ultraworkers/claw-code
- Agentic Engineering Guide: https://buttondown.com/verified/archive/the-2026-agentic-engineering-shift
600+ Commits in 40 Days: The Paradigm Shift Behind the Numbers
The point: AI coding is no longer just about "helping you write code."
| Dimension | Traditional coding | AI coding (2026) | |-----------|-------------------|------------------| | Core action | Typing code | Defining problems, organizing context | | Feedback cycle | Hours/days | Minutes | | Verification | Manual testing | Automated gates | | Iteration speed | Linear | Exponential | | Bottleneck | Typing speed | Judgment |
Each commit in cc-haha represents a complete idea → implement → verify → feedback → fix loop.
Three Leaps in AI Coding
1. Copilot era (2021–2023) — "code completion": AI as a typing assistant with single-file context. 2. Vibe coding (2024–2025) — "conversational programming": term coined by Andrej Karpathy; natural-language requirements, full AI-generated blocks, humans accept/reject — but rarely run tests and focus on single components. 3. Agentic Engineering (2026+) — Karpathy declared vibe coding outdated and proposed Agentic Engineering: 99% of the time you don't write code directly; you orchestrate and supervise agents that execute code.
The PEV loop
| Phase | Lead | What happens | |-------|------|--------------| | Plan | Human | Define requirements, design constraints, architectural intent | | Execute | AI Agent | Multi-file edits, run tests, iterative fixes | | Verify | Human + automation | Review, test gates, security scans |
cc-haha isn't a person writing code 600 times — it's one person orchestrating 600 complete PEV loops.
Why "Ordinary People" Can Now Build Products
1. The barrier shifted from typing speed to judgment: defining problems, organizing context, judging results, and building verification loops — no CS degree required. PMs, designers, and domain experts can become builders. 2. One person is a team: NanmiCoder's 40-day output matches what a mid-sized team would ship in 3 months. Andrew Ng predicts the PM:Engineer ratio may flip from 1:4–6 to 2:1. 3. Prototype-to-product distance compressed: YC CEO Garry Tan noted 25% of YC's Winter 2025 batch generated 95% of their code with AI tools. Prototype = an afternoon; product = 2–4 weeks of solo iteration.
Why a Single Model Isn't Enough
Single-model limits: context windows, knowledge cutoffs, hallucinations, no external capabilities, single-threaded operation.
cc-haha's solution — an external capability matrix:
| Capability | Role | Implementation | |-----------|------|----------------| | Browser | Real-time info, web testing | WebSearch fallback | | Computer Use | Desktop control, screenshot verification | Screenshots, mouse, keyboard | | Skills / MCP | Extensible tool connections | MCP protocol + Skills system | | IM integration | Remote approvals, multi-device | Telegram/Feishu/WeChat/DingTalk | | Multi-model | Trade-off optimization | Multi-provider config + smart routing |
Key insight: AI didn't get smarter to solve problems — it gained hands and eyes to interact with the world.
Minimal Loop Methodology: 0 to 1
1. Week 1 — Minimum viable loop: get "idea → implement → verify" running once. 2. Week 2–3 — Core experience loop: make the 3–5 most-used features usable (Tauri + React desktop framework, multi-session management, permission approvals). 3. Week 4–5 — Extended capability loop: Computer Use, IM access, WebSearch, multi-model providers. 4. Week 6+ — Quality & scale: CI/CD, test coverage, docs, community governance.
Quality Gates Accelerate Iteration
Counter-intuitively, testing speeds you up:
| With gates | Without gates | |-----------|---------------| | Every commit verified | Technical debt accumulates | | Bugs found in 5 minutes | Bugs found in 5 days | | Fix cost = 1x | Fix cost = 10x | | Bold refactoring | Fear of change | | Community contributes | Community avoids touching code |
The quality pyramid: human review (architecture, security) → automated tests (unit, integration, E2E) → static analysis (types, lint, security scans) → AI self-verification (agent-run tests, self-repair).
AI Didn't Remove Human Responsibility — It Pushed It Forward
| Before | Now | |--------|-----| | Responsible for typing | Responsible for defining goals | | Responsible for debugging | Responsible for organizing context | | Responsible for testing | Responsible for judging results | | Responsible for fixing | Responsible for building verification loops |
These skills are scarcer than "knowing how to code" — and harder for AI to replace.
Industry Signals
| Signal | Detail | |--------|--------| | Claw-code | 100K stars in 24 hours, built by two engineers with AI on a plane | | Lovable | $1M → $100M ARR in 8 months, fastest-growing startup | | LinkedIn APB | Associate Product Builder replacing traditional APM, merging PM + design + engineering | | YC Winter 2025 | 25% of startups generated 95% of code with AI | | Andrew Ng | PM:Engineer ratio may flip to 2:1 | | Naval Ravikant | "Vibe coding is the new product management" |
Action Checklist for Indie Developers
1. Start with a minimal loop — no 6-month roadmaps; get one idea→implement→verify cycle running. 2. Build quality gates early — one test beats none; automation beats manual checks. 3. Don't rely on a single model — give the AI browsers, MCP tools, Computer Use. Make agents able to *do*, not just talk. 4. Push responsibility forward — 80% of time defining problems and verifying results, 20% executing (let AI do it). 5. Embrace Agentic Engineering — you're an orchestrator, not a typist.
One-Sentence Takeaway
> The future gate isn't "can you write code" — it's "can you clearly define problems, organize context, judge results, and build verification loops." AI made building cheap and made decision-making scarce.