30天衝破八千星！這款 crm 示範了什麼叫不瞎猜的 AI Agent 實作｜開源好物週報

Introduction

An open-source project called crm, which racked up an incredible 8,917 GitHub stars in just 30 days, has completely shattered our imagination of what AI applications can be. Instead of relying on fancy chat boxes, it runs Agents directly as the main characters in background scheduling. In this issue, besides breaking down this amazing Agent implementation architecture, I will also share the Mojo programming language—which was just announced as fully open-source under Apache 2.0 and acquired by Qualcomm—as well as a mysterious website that helps you painlessly "debloat" your setup using open-source software. I'll tell you all about them in a moment.

Open Source Picks of the Week

1. crm: A Practical Open-Source CRM Architecture That Treats Agents as First-Class Citizens and Never Guesses Facts

What it is: It solves a major pain point of most CRMs on the market when integrating AI—simply slapping on a useless, awkward chat window. It is specifically designed for engineers who want to self-host their tools and are actively building AI Agents. This project is not just a ready-to-run product, but also a highly valuable reference template for a reliable Agent architecture.
This Week's Buzz: A brand-new project launched within the last 30 days, its GitHub stars quickly surged to 8,917. It is licensed under the MIT license, with TypeScript as its primary development language.
Key Features:
- Job Queue Scheduling: The Agent is not a passive chat box; instead, it runs continuously on its own job queue. It independently decides which contact to research next and when to schedule follow-ups. Users can view its reasoning process and questions in a dedicated Agent tab.
- "Never Guess" Evidence Rules: It only records observed facts and attributes their sources (e.g., crm.signature-block, github.account-identity). When the evidence is weak, it only offers suggestions to humans; it only auto-fills database fields when the evidence is strong.
- Highly Customizable & Sandbox Security: It provides 18 custom-written tools and 4 plain-text prompt configuration files (such as evidence.md). The prompting logic is completely transparent, and the Agent sandbox blocks all outbound connections by default (deny-all egress—no internet access, no direct database connections), keeping only bash and grep for extremely high security. The tech stack is very developer-friendly, utilizing Next.js App Router, NestJS, tRPC, Postgres, Prisma, and Bun.
Quick Start: As long as you have Bun and Docker, you can spin it up locally with just a few commands:
```
git clone https://github.com/trycompai/crm.git && cd crm
# 複製 .env.example 後執行：
bun install
docker compose up -d
bun run db:deploy
bun run dev
```
The frontend is at localhost:3000, and the API is at localhost:3001. Integrations like Perplexity and LinkedIn are optional; it runs fine even without entering API keys.
Comparison: Open-source CRMs like Twenty and EspoCRM are traditional databases with AI chat windows tacked on. Comp AI does the opposite: the Agent is the permanent protagonist, and the UI is merely its work log. Compared to HubSpot and Salesforce, which lock AI features behind paywalls, this is completely open-source under the MIT license with no separate paid cloud version.
Who it's for: Backend and full-stack engineers who are developing AI Agent applications and want to learn how to design a "reliable, no-hallucination" architecture.
Link: https://github.com/trycompai/crm

2. Mojo: An AI-Era Programming Language with Compiler and Toolchain Fully Open-Sourced under Apache 2.0

What it is: After shouting "open source" for three years without actually releasing the compiler, Mojo has finally open-sourced its compiler and entire toolchain under the Apache 2.0 license, alongside releasing the stable 1.0.0 version. This addresses a major pain point for AI application developers who want to write GPU kernels but want to avoid dealing with complex C++.
This Week's Buzz: It garnered 233 points and 115 lively comments on Hacker News, with half of the buzz coming from the news that Modular was just acquired by chip giant Qualcomm.
Key Features:
- True Open Source & Stable Release: All tools, including the compiler, are now hosted on GitHub, and the version has reached Stable 1.0.0 (no longer a preview). Additionally, daily nightly builds are available for tracking.
- One Language for Everything from CPU to GPU: The Modular Platform supports hardware from multiple vendors, including NVIDIA and AMD GPUs, AWS Trainium, Google TPU, and Qualcomm Cloud AI 100 Ultra. Furthermore, the team has promised that following the Qualcomm acquisition, they will continue to support hardware that directly competes with Qualcomm's platforms.
- Painless Python Ecosystem Integration: Installation is as simple as installing a standard Python package, and an official VS Code extension is also provided.
Quick Start: On macOS or Linux, you can install it with just two lines of commands:
```
curl -LsSf https://astral.sh/uv/install.sh | sh
uv pip install mojo
```
Then you can start writing .mojo files. If you prefer the conda ecosystem, you can use pixi init with pixi add mojo.
Comparison: Compared to Triton, which is a Python-bound, NVIDIA-centric kernel DSL, Mojo is a complete, independent language where you write everything from CPU programs to various accelerators using the same syntax. Compared to Julia or C++/CUDA, Mojo's installation and configuration complexity is extremely low, eliminating the need to wrestle with complex toolchains first.
Who it's for: Python and machine learning engineers who want to develop high-performance AI applications or GPU kernels but don't want to write complex C++ code.
Link: https://www.mojolang.org

3. debloat.dev: An Open-Source Alternative Guide to Help You "Debloat" Bloated Closed-Source Software

What it is: It solves a major pain point for self-hosting enthusiasts and developers who spend hours searching and filtering for "clean, lightweight" open-source alternatives. It is a clean, ad-free directory of open-source alternatives with the tagline "replace the junk," ready to use right out of the box.
This Week's Buzz: It garnered 402 points and 128 comments on Hacker News.
Key Features:
- Intuitive Comparison & Rating Tables: Currently featuring 213 open-source projects and 168 posts, it clearly indicates "which closed-source software is being replaced" (e.g., using Jellyfin to replace Plex, Home Assistant to replace SmartThings). It also directly displays the license type, user ratings, and discussion counts, so you don't have to click through just to find out if it can be used commercially.
- Interactive Forums & Wishlist: Each item listing functions as a forum where anyone can leave comments. Once logged in, users can post new items, rate them, or request alternatives in the wishlist section.
- Machine-Friendly Data Formats: It provides a sitemap, RSS feed, llms.txt, and API documentation, making it easy for developers to access programmatically or even feed directly to an AI Agent to build custom dashboards.
Quick Start: No installation required—just open https://debloat.dev in your browser. To automatically track new items, simply subscribe to its RSS feed or connect to its API.
Comparison: awesome-selfhosted is merely a static Markdown list with links but no ratings or discussions, while alternativeto.net is cluttered with closed-source options and heavy advertising. debloat.dev focuses strictly on open-source, provides community ratings, and additionally offers llms.txt and API endpoints for machine consumption.
Who it's for: Developers and system administrators who love self-hosting, want to streamline their daily digital tools, and are looking for clean alternatives.
Link: https://debloat.dev

Closing Thoughts

Looking at this week's trends, the open-source world is moving toward "greater practicality" and "more thorough openness." Whether it is crm, which makes Agent implementation transparent, or Mojo, which is finally fully open-sourced, both have injected powerful momentum into the developer community. I hope the tools compiled in this issue are helpful to you. This is Mu-Yen, and I'll see you next week in the "Open Source Picks Weekly"!