This week's Swipe · September 9, 2026 · 5 min read
Agents that work while you sleep
Tools to keep coding agents running after you close the laptop, keep API keys out of the chat, and let parallel sessions hand off work.
For years I have used caffeinate on my MacBook Pro to keep it working. Now, I just want to close the lid, relax, and know that my agent are still cooking in the background.
There are many cloud agents now, from Cursor to Devin to Amp, to the Claude and ChatGPT apps. But I like to work in the terminal and recently discovered Herdr to have my CLI agents working round the clock on my Mac Mini.
RIP my rate limits.
Most of the tools and skills this week are about how you can move your agents to a remote server that keeps running when you switch off, without your API keys in the chat, and without two of them fighting over the same file.
Let’s get to it!
In this issue
- ▪
Belt, One CLI for everything your agent does [sponsor]
- ▪
keyclasp, Keep API keys out of agent context
- ▪
crew, Parallel agents in one checkout, no worktrees
- ▪
Bounce Router, Fail over when Claude hits its limit
- ▪
Herdr, Agents keep running when you close the lid
- ▪
consensus, Critics attack the fix before it ships
- ▪
herdr-ohmyzsh, Oh My Zsh shortcuts inside Herdr
- ▪
Firstmate, Talk to one agent that runs the crew
Give every agent the same skills, tools, and context. [sponsor]
Why: Every agent you run has its own pile of skills, and none of them share. Belt is one CLI that puts the same skills, tools, and connectors in front of Claude Code, Cursor, and Codex, so a workflow you fixed once turns up everywhere.
Try: Install Belt with curl -fsSL https://cli.inference.sh | sh, then run belt skill use inferencesh/web-search and ask your agent to research something. Open Claude Code, Cursor, or Codex next and the same skill is already there.
Tools
Your agent asks for a secret by name, never the value.
Why: Hand an agent an API key and it ends up in a prompt, a .env, or the transcript. Andrea Catalucci built keyclasp as a local encrypted vault: the agent says API_KEY, and the real value goes only into the one command you trust.
Try: Run npm install -g keyclasp@beta, store a key with keyclasp set, then let the agent run keyclasp run --env API_KEY -- npm test. The tests get the key, the chat never does.
Claude Code, Codex, and OpenCode sessions see each other’s live work.
Why: Two weeks ago Zuse gave every agent its own worktree. Mo Moustafa built crew on the opposite bet: sessions in one checkout see each other’s status and last few transcript lines, so Codex knows Claude is mid-refactor and steers around it.
Try: Run npm install -g @0xmmo/crew, open Claude Code and Codex in the same repo, then type crew send --all "deploying to staging now" from a plain terminal. Both agents get it within seconds, even mid-turn.
One terminal that switches between Claude, Codex, and Muse.
Why: You hit the Claude usage cap mid-task and lose the thread while you paste everything into Codex. Rich Chetwynd built Bounce Router to run all three from one prompt, hand the conversation and git state to the next signed-in provider, and carry on.
Try: Run npm install -g bouncerouter, then bounce inside a repo you’re happy to let it change, because the default mode skips approvals. It reuses the logins already on your machine, and /quota shows how much of each window is left.
Close the laptop and your coding agents carry on working.
Why: I run my agents inside Herdr. Can Celik built it as a background server that owns the terminals, so closing the lid or dropping SSH kills nothing, and the sidebar marks each agent working, blocked, or idle.
Try: Install with curl -fsSL https://herdr.dev/install.sh | sh, start Claude Code and Codex in two panes, then press ctrl+b q to detach and shut the lid. Run herdr later and both are exactly where you left them.
Skills & workflows
Independent reviewers tear the agent’s plan apart before it codes.
Why: The same model double-checking its own idea has the same blind spots twice. consensus sends the lead agent’s written proposal to up to three read-only reviewers, then makes it check every finding against the code and reject the wrong ones with a reason.
Try: Run claude plugin marketplace add alpbahadur/interns-review-plugin, then claude plugin install interns-review@interns-review --scope user. It fires on its own once the agent has a plan, and the receipt tells you what the critics found and what got thrown out.
Long shell commands get the same working and done status as agents.
Why: Robby Russell created Oh My Zsh and is, in his words, “still wrapping my head around” Herdr. His plugin treats any command over ten seconds like an agent: working in the sidebar, then a toast when it’s done.
Try: Run herdr plugin install robbyrussell/herdr-ohmyzsh, add herdr to the plugins list in .zshrc, then start your test suite and switch panes. The toast finds you with the exit code and how long it took.
One first mate spawns the workers and hands you finished PRs.
Why: It’s the agent I run inside Herdr. Kun Chen built Firstmate as a repo of instructions and skills, not an app: you talk to the first mate, it gives each job its own Herdr tab and worktree, and only interrupts you for real decisions.
Try: Run gh auth login, clone the repo, and start claude from a Herdr pane so it uses Herdr tabs automatically. Type “ahoy! fix the flaky login test and add dark mode” and wait for the PR link with its risk note.
Reach out
- Found a tool we should see? Send it over.
- Launching something? Sponsor an issue.
- Built something useful with AI? Tell us about it.
Plain .mdyou can paste straight into Claude, ChatGPT, or any agent.
Disclosure: Our picks are editorially independent. Each one has to give you something useful to swipe. Sponsors are labelled and never reviewed for pay.