Overboard
Multiple Claude Code Agents on Your Mac: You're On Your Own
Using /schedule to run agents in Claude is manageable. Running multiple Claude Code agents unattended on your local Mac is where it gets complicated, and where Anthropic doesn't seem interested in helping.
Every service is trying to put your data in the cloud. That works for some cases but not all workflows.
I run agents that write blog posts, browse social feeds in my own logged-in browser, and cut demo videos from files that live on my disk. None of that fits in a fresh cloud container. So this is a tour of what the cloud path does well, what Anthropic ships for local runs, and the gaps you have to fill yourself once there's more than one agent on the schedule.
Why /schedule is the easy part
Type /schedule daily PR review at 9am and Claude walks you through the rest. The result is a routine: a saved prompt, one or more GitHub repos, and a set of connectors, running on Anthropic's infrastructure. Your laptop can be closed. Nothing prompts for permission. Each run shows up as a session you can open and read.
For work that lives entirely in a GitHub repo, that's a good deal. The limits only show up when the work lives somewhere else:
Every run is a fresh clone. No access to local files. The 40GB of screen recordings on your drive don't exist as far as the routine knows.
No logged-in browser. An agent that reads your social feeds or posts to a forum needs your real browser profile, because that's where the login lives.
Your local MCP servers don't come along. Anything you added with
claude mcp addis stored on your machine, so it has to be set up again as a connector.The minimum interval is one hour, and there's a daily cap on how many runs can start per account.
Routines are also still labeled a research preview. None of this is a complaint. It's a description of the shape: a routine is a repo job. Plenty of useful agent work isn't one.
What Anthropic gives you for local runs
There are three local options, and each one is fine on its own.
Desktop scheduled tasks run on your machine with full access to your files and tools. They only fire while the desktop app is open and the computer is awake. If the Mac sleeps through a scheduled time, the run is skipped, and you get one catch-up run when it wakes.
/loop repeats a prompt inside a session. It's built for polling a deploy, not for a standing schedule: recurring tasks expire after seven days.
Agent view gives you one screen for background sessions that keep running after you close the terminal. It's a way to watch several agents. It is not a scheduler, and a restart stops whatever was running.
Now read the docs' own advice on choosing between them: "Use cloud tasks for work that should run reliably without your machine." The reliable option is the cloud one. The local options are described by what they need from you.
That's what I mean by not interested in helping. The pieces for running one local agent exist. The piece that coordinates several of them doesn't.
Where multiple Claude Code agents collide
One scheduled agent rarely has a problem. The trouble starts with the second one, because a Mac has exactly one of several things.
The clock slot
Desktop's run history lists the reasons a run got skipped. Two of them are "the previous run was still in progress" and "other scheduled tasks were already running." So when two tasks want the same moment, one of them doesn't wait its turn. It just doesn't happen.
Skipping is a safe default. It's also how your nightly blog post quietly fails to exist because the security review ran long. And if you schedule with launchd or cron instead, you get the opposite problem: nothing stops a slow run from overlapping its own next firing.
The browser profile
Chrome won't open the same profile twice. Two agents that both need your logged-in browser will fight over tabs or fail outright. This is the resource the cloud can't give you at all, and on your Mac there's one of it.
The state file
Agents that run repeatedly keep notes: a JSON file of what's done, a queue of what's next. Two agents reading the same keyword queue will both take the top item, and you wake up to two articles on the same keyword. Last writer wins, and the other one's notes are gone.
The usage pool
Every agent draws from the same Claude plan. Anthropic's docs are blunt about it: running ten agents in parallel uses quota roughly ten times as fast as running one. Eight agents at 2am is a decision about how much of your week you want left on Tuesday morning.
The machine itself
Port 3000, the package cache, CPU, RAM. And the Mac has to be awake and logged in, with the lid open. That's the trade you make for keeping your files where they are.
What fixes most of it for free
You can get a long way with habits before you need a tool.
Stagger the start times. Stop scheduling everything at 2am. Put the code review at 2:00, the blog post at 2:30, the social browsing at 3:15. Most overnight collisions are four jobs that got the same start time because 2am was the obvious number.
Give every agent a lane. Write down what it owns: this repo, this folder, the browser, this state file. Two agents that own the same thing don't run at the same time.
Make runs safe to repeat. "Is there already a draft with this slug?" is one check, and it turns a duplicate into a no-op. It's the same "is it safe to run twice?" test I use to decide what gets scheduled at all.
End in something reviewable. A draft, a branch, a rendered video in a folder. Never a publish or a push nobody watched.
Staggering has a ceiling, though. Run times drift, you add a fifth agent, and you're maintaining a timetable by hand.
One queue, on the Mac you already own
This is the gap I built Overboard for Mac to fill. No cloud: it turns your Mac into a 24/7 agent orchestrator, running on the Claude Max or Pro subscription you already pay for. No API key, no token bill that grows with the number of agents you started.
Every scheduled task and every one-off you dispatch from your phone goes into one global queue, priority first and then first in, first out. You pick the concurrency cap, from one at a time up to eight. A task that finds the slot busy waits instead of being skipped, and a schedule never overlaps its own next firing.
Each run is a real Claude Code session in its own Herdr tab, which the scheduler requires. You can scroll back through the night's work or take over a session mid-run. The Mac has to be logged in, and there's a keep-awake menu to stop it sleeping while agents run.
It's $49 once, with a full 14-day trial. The free MIT plugin does the same scheduling and dispatching manually from your browser.
If you haven't decided where your agents should live yet, I compared the options in cloud, desktop, or your Mac and in cloud VMs vs your own Mac.
The short version
/scheduleis great for repo work. It can't see your files, your browser, or your local tools.Anthropic's local options each handle one agent well. None of them queues several.
Multiple Claude Code agents on one Mac share five things: the clock slot, the browser profile, the state file, the usage pool, and the machine.
Stagger start times, give each agent a lane, make runs safe to repeat, and end every run in a draft.
When the timetable gets painful, put everything behind one queue with a cap you chose.