Jordan Rejaud

Issue tracking has been upgraded 🤖 Before: 3 people on my team, every task was a card with all convo on that specific card: No emails, Slack only for time sensitive stuff. Now: 10 AI agents running in parallel (no more people), worktrees from linear cards, automatic PR creation, following up with the AI agents directly from Github. This is what Kanban looks like in a post-wagie world.

@beeper Yes, the above explainer video was entirely made by AI too (using code, not a generative model)

Claude Code can talk to WhatsApp, Telegram, Slack, etc through @beeper but it can't send files: The official Beeper MCP has no file sending support. I forked it and added file transfers across all chat networks. Send a PDF to WhatsApp, a screenshot to Telegram, a voice memo to Slack: all from your terminal (link in reply)

Whats the use case for @DeadSimpleEmail ? I have my agents send emails using my personal accounts (always ask, ofc). And if I need a dedicated email for them, it costs me 2 bucks a month on zoho

@batuhan When I need an API method that's not in the MCP/ integration yet, I spawn a new CC session to implement it. My goal is to use the AI/ terminal as the interface into everything. I feel my brain slow to a crawl when I have to use boomer GUIs or web br*wsers again

i know i am doing something right when people fix the issues i cause i gotta ship file uploads to beeper mcp now i guess!

github.com/jrejaud/beeper… Fork of Beeper's MCP server with the file sending tool

Thank you @superwhisper for saving me so much time every single week. I don't talk to my GF or my dog anymore, I only ramble to the machine spirit. Imagine choosing to be a typecuck, when you can be a voicechad

@aiedwardyi I use it as a long running memory graph rather than shared session state. Creating permanent memories is very useful. I think it's probably the most useful MCP I've installed.

@JordanRejaud fair point on the edge case, but graffiti memory as a shared state layer is interesting - that's a pattern worth exploring more as agent orchestration gets more complex

I run 10 Claude Code instance at once. Each one spawns 51 MCP server processes. That's 510 processes eating ~25GB of RAM. I containerized all 51 into shared singletons. 510 processes down to 51. 25GB down to 5GB. I asked Claude to make this video to explain it (inspired by @Emarky's Remotion pipeline).

@aiedwardyi Because nobody's running 10 agents at once with 50 MCPs, it's an edge case. They can share state through singletons, ex: writing to grafitti memory.

@JordanRejaud The 10x RAM reduction is impressive. Containerizing MCP servers into shared singletons is such an elegant fix — makes you wonder why this isn't just the default behavior. Are the instances fully isolated from each other or do they share any state through the singletons?

@jason_haugh I haven't had many issues with agents using the wrong tool since I've added skills (aka prompt injection based on keywords).

@JordanRejaud Exactly. My research agent only sees web search and file tools. My engineer only sees git and code tools. No agent touches everything. The memory win matters, but the bigger gain is reliability. When an agent has 5 tools instead of 50, it picks the right one almost every time.

I have 10 claude code instances and 50 MCPs, so was running 500 total, even my 96GB RAM Mac was having trouble lol. So, I turned all my MCP servers into docker container singletons. I updated imap-mcp-server so it can properly download email attachements from the container. PR up on the upstream repo.

@aiedwardyi @AnthropicAI I was hitting CC's $200 plan's session every single day, github.com/rtk-ai/rtk was a massive boon and I haven't hit it since.

Been using @AnthropicAI's Claude Code every day and kept getting blindsided by rate limits. So I built a real-time statusline that shows your 5h and 7d quota remaining, context window usage, and token counts — right in your terminal. Open source. Zero config. Just clone and go. #BuildInPublic #ClaudeCode

@jason_haugh So you reduced the number of MCPs per AI agent? How big is the memory advantage of using that over generic agents that have access to singletons?

@JordanRejaud 500 MCP processes on one machine is bold. I hit similar walls running 8 agents on a Mac Mini. The move that saved me was giving each agent a narrow scope with strict boundaries instead of letting everything talk to everything. Less is more at scale.

If this gets 50 RTs I'll write up exactly how to do it lol

My laptop crashed last night and I just woke up, *cracks neck* @andxdy

I told Claude: I use LMs and the terminal as the interface to everything on my computer. Not just coding. Reading emails, sending messages, browsing, managing servers, making images, checking calendar. Make a video representing this.

@coatol5 There's no way this is accurate

This is what software engineering looks like in 2026. I'm managing 4 concurrent agents at the same time, going back and forth like I'm controlling units in an RTS game. In a single day I'm able to burn through $73 worth of Anthropic tokens. I've even heard of people doing 12 concurrent agents at the same time! It's scary how fast things are changing.

Just broke 1.5 million USD all-time revenue last night, see the evidence for yourself at @marclou 's TrustMRR trustmrr.com/startup/parake…