Every agent I've built starts the same way. Paste an API key into .env, export it, hope it doesn't end up in a log or a subprocess env dump. token expires and something quietly breaks. We've all been there
so I wrote authsome. The bit I think is actually interesting is the run command:
authsome run -- python my_agent.py
It launches the child behind a local auth proxy and the proxy intercepts outbound HTTPS and injects Auth headers at request time. the child process never has the secret in its environment, so it can't leak through os.environ, ps -e, or anything that dumps a subprocess env and the agent code doesn't change as well.
the tokens are stored locally, encrypted, and refreshed before they expire. Oauth flows for interactive and headless, plus a browser bridge for API-key providers. There is a cli for pulling headers directly when you don't want the proxy.
the proxy only sees traffic that goes through it, so libraries that pin their own CA bundle slip past, also the streaming uploads and long-lived connections probably have edge cases I haven't hit. It's still alpha, v0.2.1.
Most interested in feedback on the proxy approach itself, that's the part I'm least sure about.
So I sat down one day thinking this sucks, there isn't any platform that solves this problem for lawyers who are not Supreme Court or High Court Related as most companies build for them (more like the middle kid who gets ignored.
So built this, let me know what you guys think.
This covers:
- ITAT (Income Tax Appellate Tribunal)
- CESTAT (Customs, Excise & Service Tax Appellate Tribunal)
- GST AAR (GST Authority for Advance Rulings)
- NCLT (National Company Law Tribunal)
- IBBI (Insolvency & Bankruptcy Board of India)
- DRT (Debt Recovery Tribunal)
- SAT (Securities Appellate Tribunal)
- CCI (Competition Commission of India)
- NGT (National Green Tribunal)
- APTEL (Appellate Tribunal for Electricity)
quick tldr; We are doing a live 60 minutes AMA with folks from Microsoft, Pinecone, Santiago and Alden (CEO CustomGPT.ai) on MCP, sounds interesting? Register.
The goal is to educate about MCP, answer questions, and cover use cases: RAG + MCP, IDEs + MCP, etc. We’ll have live demos, Pinecone folks talking about what they are up to, and much more fun!
If you have been early in the MCP race, this would surely be worth your time.
Why might this interest you?
Model Context Protocol (MCP) is a low-level JSON-RPC protocol for passing structured context and tools to an LLM. Instead of gluing prompts together, you expose one JSON endpoint for a tool (and it takes care of tons of API endpoints for that tool).
MCP is just REST for LLMs! It really is that simple!
We plan to show a live demo of a working MCP, preferably hosted one, setting up configs, with Claude.
We will also answer any questions!
Featured Speakers:
1. Michael Kistler - Principal Program Manager at Microsoft
2. Arjun Patel - Senior Developer Advocate at Pinecone
3. Santiago (https://www.linkedin.com/in/svpino/) - Computer scientist and teaches hard-core Machine Learning; will walk you through Why do we need MCP?, Before MCP vs. After MCP, Architecture, Primitives, and Advantages.
4. Alden Do Rosario - will dissect the RAG + MCP pipeline we run in prod, live demo.
Format: - 3×10 min tech talks (protocol, integration, case study) - 10 min panel on lessons learned - 20 min open Q&A - bring tough questions
We've been using CustomGPT.ai's RAG API for client projects but kept rebuilding the same UI components. Today we're open-sourcing our complete implementation (think ChatGPT interface open sourced).
No vendor lock-in. No telemetry. No "premium" features.
We built this because we needed it. We're sharing it because you need it too.
What does it mean to you?
1. Free and Ready to use UI like ChatGPT with Voice
2. 100% Customizable for you to build on top of it
3. Dev community to fix bugs and feature requests
4. Why create RAG from scratch when you can use free templates?
Technical details:
1. Next.js 14 + TypeScript + Zustand for state
2. Proxy architecture keeps API keys server-side
3. Proper SSE streaming with cleanup and error boundaries
4. Voice: OpenAI Whisper STT + TTS (6 voices)
5. Three deployment modes: widget.js bundle, iframe, or standalone
6. PWA support with service worker
7. Dark mode + full mobile responsiveness.
Interesting challenges solved:
1. Concurrent message streams without memory leaks
2. Widget state isolation when multiple instances on the same page, 100% customizable.
3. CORS handling for cross-domain embedding
4. Citation preview just like ChatGPT
Deployment options:
1. Vercel/Netlify (one-click)
2. Railway/Render
3. Docker
4. Google Apps Script (single file, 20k req/day free for select social RAG AI bots)
Also includes 9 social platform bots (Slack, Discord, Telegram, etc.) that connect to the same CustomGPT.ai backend.
quick tldr; We are doing a live 60 minutes AMA with 3 industry experts on MCP, sounds interesting? Register.
The goal is to educate about MCP, answer questions, and cover use cases: RAG + MCP, IDEs + MCP, etc. We’ll have live demos, Pinecone folks talking about what they are up to, and much more fun!
If you have been early in the MCP race, this would surely be worth your time.
Why might this interest you?
Model Context Protocol (MCP) is a low-level JSON-RPC protocol for passing structured context and tools to an LLM. Instead of gluing prompts together, you expose one JSON endpoint for a tool (and it takes care of tons of API endpoints for that tool).
MCP is just REST for LLMs! It really is that simple!
We plan to show a live demo of a working MCP, preferably hosted one, setting up configs, with Claude.
We will also answer any questions!
Featured Speakers:
1. Santiago (https://www.linkedin.com/in/svpino/) - Computer scientist and teaches hard-core Machine Learning; will walk you through Why do we need MCP?, Before MCP vs. After MCP, Architecture, Primitives, and Advantages.
2. Alden Do Rosario (CustomGPT.ai CEO) - will dissect the RAG + MCP pipeline we run in prod, live demo.
hey folks, Priyansh this side, I just put together a list of all the tools I could find that are openai endpoint compatible, and so now this gives you the power use your openai based project and add RAG functionality.
if I missed some tools, feel free to jot them below
