Agent Manager

A macOS app for juggling multiple Claude Code and Codex accounts on the same machine.

It lets you run each account as a separate agent session, keep an eye on how much usage each one has left, and warm up an account's usage window to increase the number of tokens available.

Agent Manager only touches accounts you own and pay for. It drives the official claude / codex CLIs and never proxies or stores your OAuth tokens — it only reads the credentials those tools already wrote.

What it does

1. Run any account independently

am run <id> [<args forwarded to claude/codex>]

This starts a claude or codex session under that account's own isolated config home and hands your terminal to the CLI. Your normal ~/.claude / ~/.codex login is never touched.

Each account's home symlinks back to your real ~/.claude / ~/.codex for everything except the per-account identity file, so accounts share the same settings and the same session history. That means you can pick a session back up under a different account — handy when one account's window runs out mid-task:

am run claude-ms55
# runs out of tokens
am run claude-ms18 --resume <session-id>

The Source home effectively groups accounts: everything pointing at the same folder shares settings and history, so you can run a few work accounts off one source home and your personal ones off another, fully apart.

2. Track usage at a glance

Track each account's usage from the menu bar ( individual menu bar entries or one collapsed) or from the CLI (am usage).

Menu bar — individual	Menu bar — merged
CLI — am usage

3. Warm up token windows

Instead of starting your subscription's 5-hour usage window on your first request, start it at a fixed time beforehand, to maximize the number of tokens available when working.

Simply paint your working hours in the app, flip the Scheduler active switch, and Agent Manager fires a small ping to open each account's window just before you start — so you begin the day with a fresh window instead of starting the clock the moment you sit down.

The pings come from one resident background agent (installed once — macOS shows its "background items" notification a single time, ever). While the switch is on, calendar repaints and account changes apply live; there's no re-apply step.

am ping <id>             # open <id>'s 5h window now (also what the scheduler runs)

Each ping is a real interactive turn in the official CLI over a terminal — never an SDK or programmatic call — to stay future-proof and not lose access through policy changes like this one.

Pings stay minimal — a couple around the edges of your day, not all night. A ping the Mac slept through is skipped, not fired late.

Lid closed? Optionally flip Wake Mac for pings in Preferences: a tiny root helper (approved once in System Settings → Login Items) arms a hardware wake ~45 seconds before each ping, the ping runs, and the Mac goes back to sleep if nobody's around. Firmware rule: a closed lid only wakes on AC power.

Lid closed on battery? That's the one case no software can wake the Mac, so for Claude Code accounts we use a cloud routine's compute to start the token window. Just flip the Claude Routine fallback toggle in Preferences. It keeps a tiny one-shot routine — "AgentManager Routine", visible at claude.ai/code/routines — armed 5 minutes after each scheduled ping. A ping that runs locally disarms the pending routine, so it never fires; but a ping the sleeping Mac misses lets Anthropic's cloud run it instead (one minimal Haiku turn that opens the window), and the redundant local ping is skipped on wake.

CLI reference

am run <id> [<args>]      launch a session as <id>; remaining args go to claude/codex
am list                   list accounts with status + provider
am usage [<id>]           capacity for connected accounts (--week, --provider, --sort, --no-color)
am ping <id>              fire one ping now — opens <id>'s 5h window

Adding and connecting accounts, and painting your work hours, happen in the app (the Agents, Planner, and Monitoring screens). The CLI handles only the running-related actions.

How it works

• Isolated homes. Each account is its own CLAUDE_CONFIG_DIR / CODEX_HOME under the app's folder. Only the identity file (.claude.json / auth.json) is real and per-account; the rest is symlinked from your real config, so accounts share settings and history without stepping on each other's login.
• Official CLI only. Logins, pings, and launches all run the real claude / codex binary. Agent Manager only reads the credentials those tools write — it never relays or stores a token.
• Local only. Network calls go only to the official provider endpoints (api.anthropic.com, chatgpt.com): the usage reads the real CLI already makes, plus — only if you turn the experimental Claude Routine fallback on — managing the anchor routine in your own claude.ai account. No backend, no analytics.
• One quiet background agent. Scheduled pings come from a single resident launchd agent with an in-process queue — flipping the scheduler on and off never churns launchd (and never re-triggers macOS's background-items notifications). The optional wake helper is the one privileged piece: ~200 lines that only read two workspace files and arm wake timers — it links none of the account/keychain/network code.
• Inspectable. Reads, pings, launches, and HTTP calls go to local log files you can read, with auth headers redacted.

Requirements

• macOS 14 (Sonoma) or later
• The claude and/or codex CLI, installed and logged in at least once
• Swift 6 toolchain (Xcode 16+) to build from source

Build & run

git clone https://github.com/marinsokol5/agent-manager.git agent-manager
cd agent-manager

swift build            # library, CLI, and app
swift test             # test suite
.build/debug/am help   # the CLI

# `make build` assembles and codesigns .build/AgentManager.app (the bundle is
# what makes the wake helper's System Settings approval flow possible), and
# `make run` opens it. Signing uses a local dev identity so Keychain grants
# survive rebuilds: create a self-signed "AgentManager Dev" certificate in
# Keychain Access first (or pass CODESIGN_ID="Developer ID Application: …").
make run

Where your data lives

Everything is under ~/Library/Application Support/AgentManager/:

File	Contents
accounts.json	account metadata (label, color, email, keychain service name) — no secrets
schedule.json	your work hours and window length
scheduler.json / scheduler-status.json	the scheduler switch + the background agent's heartbeat and upcoming pings
wake.json	the "Wake Mac for pings" opt-in
cloud-fallback.json / cloud-fallback-state.json	the cloud-fallback opt-in + which claude.ai routine is armed per account
usage.json	last-known usage reading per account
preferences.json	display preferences (e.g. clock style)
audit.log.jsonl / activity.jsonl / network.jsonl	local logs (auth headers redacted)
homes/<id>/	per-account config home (created 0700)

Credentials are not in any of these. Claude's token stays in the macOS login Keychain; Codex's stays in the per-account auth.json the official CLI wrote.

Responsible use

This is a convenience tool for accounts you already pay for. It uses the official CLI, keeps everything on your machine, and keeps scheduled pings minimal. Window warming depends on provider behavior that could change at any time, so treat it as a personal optimization, not a guarantee.

Contributing

Architecture, commands, conventions, and the security rules the code assumes are in AGENTS.md. The logic lives in AgentManagerCore (unit-tested); the app and CLI are thin layers over it.

License

MIT © 2026 Marin Sokol

---

Agent Manager is an independent project — not affiliated with, endorsed by, or sponsored by Anthropic or OpenAI. "Claude", "Claude Code", and "Codex" are trademarks of their respective owners. It manages only accounts you own and pay for, and drives each provider's official CLI.