Free · no trial, no paywall
Every agent, every allowance, one calm place.
Overtem is a native Mac menu-bar companion for people who run AI coding sessions. See what’s running, what it costs and what’s safe to close, then share your building if you want.
Free to use. Native for macOS 14 or later. Bring your own AI accounts.
Monitor
Your Mac, read honestly.
CPU across every core. Memory by what’s actually holding it. Swap in use, kept separate from swap activity.
CPU
32%
all 16 cores combined
Per-core bars are percent of one core.
Memory
22.8 / 48 GiB
used by apps, wired and compressed
- App memory 13.4
- Wired 4.6
- Compressed 4.8
- Cached files 9.6 reclaimable
- Free 15.6
- Pressure
- Normal
- Swap in use
- 0.4 GiB
- Swap activity
- 0 B/s
- Network
- 1.2 MB/s down
- Disk
- 412 GB free
- Battery
- 86%, on power
Cached isn’t pressure. Files macOS can drop instantly are shown, never counted against you.
Swap size isn’t swapping. What’s parked on disk and what’s moving right now are two readings.
Unknown stays unknown. If a reading isn’t valid, Overtem says so instead of showing zero.
Agents
Every agent. Every child process. One tree.
Claude Code, Codex and your own CLIs, grouped by project, with whatever is waiting on you at the top.
Needs you Codex in bai8-api is waiting for input · 4 min
Open sessionReported, not guessed. Task text comes from the agent’s own session files. If it isn’t reported, it says so.
Ownership is proven. Processes join a tree only through verified session registration.
Pins stick. Pin anything essential and nothing in Overtem will offer to stop it.
Safe cleanup
Heavy isn’t stale.
Overtem only offers to stop what it can prove your agents left behind, and shows its evidence first. Everything else is protected.
Worth a look
storybook pid 51022
Evidence
- OwnedStarted by Claude Code in marketing-site, verified at launch
- Session ended42 min ago, exit reported by the agent
- No clientsNo established connections in 3 snapshots over 6 min
Counter-evidence
- Listening on :6006A browser tab could reconnect later
Candidate. Nothing happens until you ask.
Protected by construction
- next devHeavy, 9.4 GiB, 2 active clientsIn use
- nodeOwner unknownUnknown
- postgresUsed by two projectsShared
- mds_storesmacOS SpotlightSystem
- redis-serverPinned by youPinned
Diagnostics only. No stop button unless you explicitly own it, and even then a preview comes first.
Auto-clean is opt-in. Narrow rules start as dry runs and show everything they would have touched.
No stale signals. Process identity is revalidated by start time, never by PID alone.
Graceful first. Force is a separate, explicit decision.
AI efficiency
Watching is free. Thinking is on request.
Readings are local and deterministic. A model runs only when you ask, on the account you choose, within a budget you set.
You asked “Why is my Mac slow?” · one model call · 2,140 of 4,000 tokens
Ask Overtem
Sample workspace- Account
- Claude · Max (personal)
- Model
- Your selected model · source labelled
- Effort
- Medium
- Project
- gradiently-web
- Access
- Read-only
- Budget
- 4,000 tokens
Context is redacted metadata: process names, sizes, states. No prompts, no file contents, unless you approve them. Edits need a separate approval, and Overtem never runs shell commands a model writes.
Bring your own
Use the AI you already pay for.
Connect several Claude, Codex, API or custom accounts. Overtem shows the quota each provider actually reports, and says plainly when it doesn’t.
- Codex · PlusSubscription · Codex CLI41%
5-hour window · resets in 2h 13mProvider reported
- Claude · Max (work)Subscription · Claude Code CLI
Tokens counted from local session logsQuota not reported
- Claude · Max (personal)Subscription · Claude Code CLI
Separate account · local token historyQuota not reported
- API · Team keyAPI key · Keychain
Provider billing source not connectedCost unavailable
- Custom · Local CLIYour executable · argument template
Runs only with your templateNo usage source
Tokens by day · all accounts · 26 weeks
Sample workspace- Model sources stay clear. Codex models come from its local catalogue. Claude and custom model names you enter are labelled unverified.
- Estimates are labelled. Counted tokens, reported cost and missing quota each say where they came from.
- Analysis has a budget. Input estimates and provider-reported tokens are labelled separately. Routine monitoring makes no model calls.
Privacy
It stays on your Mac.
Monitoring runs on your Mac. Usage metadata is read only after you allow it, and observation pauses in one click while your jobs keep running.
On your Mac
- Process list, sizes and states
- Project paths and session metadata
- Usage history and heatmaps
- Rules, pins and audit history
- API keys, in your Keychain
Never read by default
- Your prompts and transcripts
- Your code and file contents
- Anything from excluded paths
Leaves only when you choose
- A profile you made public
- A share card you previewed and published
- Optional account sign-in for your Mac
Pause observation stops collection and new analysis. It never stops your agents. Cancelling a model run is its own, separate button.
Your Mac. Your AI. Your control.
All current features are free to use, without a purchase or expiring trial. Your AI accounts and usage stay with your provider.
Optional social
Show your stack, if you want to.
A profile, a usage heatmap and weekly recaps you can share with friends. Private by default. You preview every card before it’s shared.
Follow, befriend, react. A quiet feed of milestones from people you choose.
No DMs. Nothing to moderate in your inbox, because there isn’t one.
No prompts, no code. Cards carry selected totals, never what you worked on.
See the communitySample Builder
@sample
Week 39 recap
22 – 28 Sep 2026
Claude Code · Codex · Next.js · Swift
Before you share
- Selected totals only
- Project names hidden
- Audience: Friends
Share