Hey @thsottiaux - you guys have over-egged it with this. I'm on a Pro plan, and I've been burning up tokens trying to explain to Codex (gpt-5.5-medium) that I'm not trying to hack my own app. This needs to be dialled down ASAP.
I've built an MCP Tray for y'all to toggle project-scoped mcps on and off for Codex, Claude Code and Opencode CLIs.
Opencode already has a toggle that seems to mix global and project-scoped mcps together.
Codex doesn't offer you a toggle for project-scoped mcps, and neither does Claude Code (for now).
Until these tools all give us toggles for project-scoped mcps, this tool will help. I've built it for myself and using it already.
All yours under MIT license to build and run from source (Swift).
Should save you 200k tokens or so to build it yourself. The link to the repo is in the reply.
Hey @thsottiaux - I think there's a bug in codex cli. No matter what I do, Stripe and GitHub plugins re-load despite me disabling them. They are fully uninstalled and disabled through /plugins config dialogue, they aren't in the config files, etc. And yet, they are loading each time Codex cli loads. This consumes a lot of context upfront.
Codex plugins turns out to be a messy business. For some reason, Codex CLI loads all github and Stripe mcp tools by default on startup and it's impossible to remove them manually. Because when Codex reloads, it reloads that marketplace manifest with both these plugins enabled. @thsottiaux
That feeling of a vibe coder's progress when you migrated a long-running function orchestration from Inngest to a VPS running Redis + BullMQ Worker, guided by Claude Code.
And it works in dev and in prod.
Who-hoo!
(sorry)
Totally the right move to stem the influx of OpenClaw spammers.
To un-nuke scheduling tools, I'd enable replies via the API for Premium accounts, and severely rate-limit it (e.g. 5 per day).
The discourse around OpenAI's hire of @steipete has produced the predictable wave of declarations: SaaS is dead, agents will replace everything, the old ways are finished. I want to offer a counterpoint drawn from something that happened to me this week.
I was building an app with a multistep text transformation pipeline. My instinct, shaped by the current moment, was to go fully agentic - set up an orchestrator LLM with tools, some of which would themselves be agents, hand it the task and the user prompt, then let it loose to figure out the path to completion through tool calling, state management and quality evaluation loops.
The reasoning seemed sound. Even today, LLMs lose attention as prompts grow longer and more complex. Prompt chaining with programmatic cleanup still produces outputs that drift from the original requirements. Why not delegate the entire responsibility for end quality to an orchestrating agent that could decide how many iterations the job actually needed and what tools to call when?
So I fed Claude Opus 4.6 (operating in the highest setting) everything I could find - documentation for agent-building frameworks like LangGraph and Vercel AI, my use case, my desired outcome. And then something unexpected happened.
The model pushed back. After several exchanges, it kept insisting that I did not need a fully agentic loop for this use case. What I needed was a largely programmatic flow with specific agentic components for specific tasks within it. It explained, in some detail, why converting the entire workflow into a multi-tool agent loop would be ineffective and inefficient - slower, more expensive, requiring far more LLM calls than necessary.
I took its recommendation and built the workflow exactly as it suggested. The results are excellent. The final text ticks all the boxes that a single turn, or even a conventional chain with multiple steps, would still miss. Parts of the original content that typically get lost in longer pipelines are preserved. And the architecture is only partly agentic. It limits calls for intelligence to those text transformation exercises where nothing can substitute for real reasoning. It minimises my exposure to LLM costs. It runs faster. The user experience is better as a result.
You might argue that costs will eventually become negligible as open-weight models compete with expensive frontier offerings. Perhaps. But speed will remain a consideration. There is inefficiency baked into calling on LLMs that no pricing model eliminates. If logic can deliver exactly the same result faster - if orchestrating small agentic loops inside a harness built as software gets you there quicker and cheaper - then I do not see why you would choose otherwise. Especially for workflows with many repeatable parts.
There is also the user perspective to consider. Agentic loops and chatbot interfaces are extraordinarily powerful. OpenClaw, code agents repurposed for non-coding tasks - these are genuinely impressive tools. But there are times and needs and use cases where you do not want a chatbot. Where you do not want to type anything. Where you do not want to interact at all. Where you just want to press a button, forget about it, and receive an output at the end. Classic software experiences, made more effective by LLMs working under the hood, serve those needs perfectly well.
I remind myself of Ethereum names. For a while they seemed like the future. During COVID I bought one. It proved completely useless because that bubble remained a bubble. The agent-on-everything moment feels similar. This is not to diminish the genius of Peter, who will now share his brilliance at OpenAI. But I do not believe agents killed software, any more than television killed radio. There is still a clear purpose for classic software from both the consumer and the developer standpoint. The interesting work lies in knowing when to reach for which tool.
I'm joining @OpenAI to bring agents to everyone. @OpenClaw is becoming a foundation: open, independent, and just getting started.🦞
https://t.co/XOc7X4jOxq
To be fair to my former colleagues at Amazon, @kirodotdev team is still plugging away.
But launching startups, even inside Amazon, is hard when you're a part of a Fortune-500 behemoth. Not impossible (and Amazon launched a ton of them in addition to core businesses), but hard.