Glide path
Your allowance across the week against the even-pace diagonal, with each day named and today lit. Above the line, you're ahead. Pro scrolls it back through past weeks.
For Claude and OpenAI subscribers
Agent Runway lives in your Mac's menu bar and shows how much of your weekly AI allowance is left — paced against how long until it renews. One glance tells you whether to push hard or ease off.
Pacing, not prediction
Runway compares the allowance you have left with the time left in your weekly window. Burn faster than the clock and the cells warm up; bank a surplus and they cool to blue.
⏳ appears when you're blue and renewal is under 48 hours away: a nudge to use the surplus before it resets.
What's inside
Your allowance across the week against the even-pace diagonal, with each day named and today lit. Above the line, you're ahead. Pro scrolls it back through past weeks.
Surplus about to reset gets an hourglass. Reset credits are called out a week before they lapse, and hovering the resets line lists them all, soonest first — Runway never spends one. An early reset gets a line on the card.
24 hours of bars for this Mac, split into cached history, new input and output. Pro adds 7 and 30 days.
The week, the 5-hour session, and a model's own weekly cap, like Fable's: shown when it's tighter than the week, with a banner once it's used up.
Drag across the chart for any span, then copy a share image or Wordle-style text, logo and all.
Offline, signed out, CLI missing, enterprise plan — it says exactly that, keeps the last good reading, and retries. When Codex has no allowance to show, OpenAI’s card says so in one calm line, with help a click away.
Notifications Pro
Runway stays out of the way until something changes what you'd do next: you're back in, your week reset early, you're heading for empty, or you're about to leave a lot on the table.
Each one comes once for each thing that happens and never repeats what another has just said. Every one has its own switch in Settings → Tune interruptions.
The examples are the app's own words, for the same Tuesday morning as the panel at the top.
You're back Once, when a session limit or an outage that stopped you ends.
Early reset or new credit As soon as your week resets early or a reset credit arrives.
Running out When your pace would run out at least 12 hours before renewal. Once a week at most.
Plenty to spare When ⏳ shows and 30% or more would go unused. Once a week at most, in the daytime.
Morning report Each provider in a line at 6 AM.
Activity
Most of what an agent "uses" is its own history, re-read from the prompt cache on every turn. The Activity tab draws that as frosted glass, with the new input and output rising inside it — then flips to New only to show what you really added. On smaller displays, the cards scroll while the header, range controls and footer stay in reach.
Pro adds a list under each chart, By project to start: the same tokens by repo, following whatever you select, with worktrees folded into the repo they came from. Click its title to switch to By agent, main sessions against subagents, or By model, split by model and effort. Claude's card splits the week too, between Claude Code, chats and Cowork.
Want room to look closer? Click Details on a Runway or Activity card for Activity details: your allowance and your tokens on one chart, beside the complete breakdown.
Read from the session logs Claude Code and Codex already keep on your Mac; after the first scan, Runway reads only appended data. Copy Image in the Activity window explicitly includes the visible private breakdown and historical chart, with Runway branding and dates. Check what's visible before sharing. Copy Card, Copy Snapshot, the other image and text shares, CLI and MCP still exclude private breakdowns and past-week curves.
Monitor
Open Monitor from a card in the menu-bar panel. Keep a separate window for each Claude account, with its own view and size. In Current Burn, your last 20 minutes of activity choose the aircraft, while the selected period controls the totals. Pick Activity or Allowance when you want another view.
See what’s driving the work in your project or model mix. When your plan includes breakdowns, expand a project into its models or a model into its projects. Keep it small as a widget, or make it taller for the chart and detail.
Curious what those tokens would look like at API rates? Pro adds an estimated API-equivalent cost, including cached tokens. It’s a comparison, not your subscription bill or your allowance. Some models may not have a known price, and you can turn estimates off in Settings.
Start a one-time 14-day preview of the estimate when you choose, even if your original Pro trial has ended. The preview is for this feature alone.
Activity details Pro
Your allowance line runs over this Mac's token bars, each with its own axis and its own label, because they're different things: the line is your whole account, chats and other Macs included; the bars are only the session logs on this Mac. Rest the pointer anywhere and a small glass readout shows both at that moment, side by side.
Drag to zoom. Drag across the chart and it zooms to exactly that stretch; the total and the Projects, Agents and Models lists always cover what you can see, nothing more. Esc (or Back) returns to where you were, and double-click steps out. Pinch (or ⌘-scroll) zooms where your fingers are, down to 20 minutes of recorded 10-minute bars, and a swipe slides earlier or later.
The whole story, in one strip. Under the chart, a slim strip holds everything Runway keeps, with a box around what you're looking at: drag it anywhere, or drag its edges. 20m, 1h, 3h, 6h, 1d, 7d and 30d are one click away.
Watch what's burning right now. Drag past the right edge and the view stays live from where you started: it reads Since 7:29 AM · live and widens with every refresh, totals and lists with it, so you can see which project is eating tokens while it happens. Live keeps Last 6 hours · live sliding along; turn it off to hold still and look around.
A view you can leave open. Your current allowance and the tokens in view sit side by side. When the provider reports it, a purple session marker at now shows its percentage; hover for the reset time. New window (⌘N) keeps another provider or range open beside it, and each window remembers its range, grouping, zoom and placement.
Try a sample on Free. The Details button opens a clearly marked Sample workspace with working ranges and breakdowns. It uses made-up data, separate from your own activity. Personal Details remains part of Pro and the trial. One quiet Details introduction helps eligible users discover it; Help → Feature tips provides deliberate replay and the automatic-tips preference.
Tokens aren't allowance: Runway never claims a share of your week from them. A dip in the line with no bars under it is real use, just somewhere other than this Mac. Copy Image copies the current viewport, including its visible breakdown and dates; nothing else shares it.
Claude accounts
Use more than one Claude account on this Mac? Runway tells them apart by organization and gives each its own card, with its own activity: Claude A, Claude B and so on, until you rename them.
Live where you’re signed in. The account Claude Code is signed in to shows its allowance live, paced as always, on every tier; with two or more accounts a thin ring on its mark says it’s the one Claude Code is using. Any other keeps its card at its last reading, or, if Runway has never read it, says “Sign in to Claude Code for its allowance”. With Pro, every account stays live.
The Runway chart, on every card. This Mac’s tokens stand as bars behind each allowance line, on one scale across every card, OpenAI’s included, so a busy account looks busy beside a quiet one — on every tier. Open Monitor on a Claude card to keep that account in view. With Pro, Activity details can follow the Claude account too.
Last seen, not forgotten. Switch Claude Code to another account and the first one’s card keeps its week up to its last reading, shown as at most what was left then, and remembers when that allowance renews. It’s drawn in gray, never the pace colors, so a remembered number never passes for a live one.
Cowork counts. Sessions in Cowork, the Claude desktop app’s agent mode, now count toward your activity.
Account names stay on your Mac: shares say Claude, never which account.
Your accounts, your way. In Settings → Arrange your accounts, drag to reorder, click a name to rename it (emoji welcome), give each a short name for the menu bar, and choose whether and how it shows there.
Forget one you don’t need. Undo, or ⌘Z, brings it right back; later it waits under Recently forgotten with a Restore button. If a forgotten account turns up again within 30 days, it comes back as itself, name and place included.
Runway never uses, stores, shows or sends your email.
Several accounts Pro
A work Claude and a personal one, or two ChatGPT accounts? Runway reads each one’s allowance live, side by side, read-only, right on your Mac.
One folder per account. Sign each extra Claude account in once, in a folder of its own, or each ChatGPT account in a Codex home of its own. Runway reads it the way it reads your usual sign-in, and nothing comes out of any allowance.
Three steps, then it’s live. Settings → Arrange your accounts → ⋯ → Live reading… walks you through it: copy the command, sign in, and Runway notices by itself.
About once a month, sign in again. A Claude folder’s sign-in lasts about a month; its card and Settings row say the date it needs signing in again by. Codex renews its own sign-in while you use it.
Read-only, and it stays here. Runway never signs you in or out and never runs those commands for you. Your numbers never leave this Mac.
Live reading of more than one account per provider comes with Pro and the 14-day trial. Your usual Claude Code and Codex sign-ins stay live on every tier.
Past weeks Pro
One bar per finished week, oldest on the left, all on the same 0–100% scale: how much of the allowance you'd used by renewal. A triangle marks a week that ran out, a pause mark one where you hit the session limit, and a sparkle one the provider reset early. A dotted cap means “at least”: Runway's last reading came well before the week ended, so the real number can only be higher.
Codex's history arrives all at once, read from its logs. Claude keeps none, so Runway builds its row from its own readings, one week at a time from the day it starts watching. The morning report adds a line about the week that just finished.
Then go back and look. The end of last week peeks in at the left edge of the glide path, dimmed. Drag the chart, or swipe sideways on a trackpad, and each past week settles into place with a line saying how it ended, in the same words as the bars. Use the arrows below the chart to move between weeks too. The card shows up to eight finished windows; Runway saves up to 52 summaries per provider and 63 days of readings for the curves, so Claude's chart fills in as the weeks go by. Separately, it keeps a daily summary of this Mac's activity for good, on every tier.
A week Runway never saw shows as a gap, never an invented bar. History stays on your Mac unless you choose Copy Image in the Activity window to include the historical chart you're inspecting. Other shares leave it out.
Terminal and AI agents
Install the runway command from Settings, and your allowance shows up in Terminal, in your status line and in your AI agents. Now an agent can check there's room before it starts something big.
runway — the morning report's lines. --json gives scripts every number, and --refresh asks the app for fresh ones first.runway statusline — one short line for Claude Code's status line, tmux or SketchyBar.runway wait — waits out a session limit or an outage, then exits, so the next step in a script runs once there's room. --json says what's blocking and when it's back.runway mcp — a read-only get_runway tool for Claude Code, Claude Desktop, Codex and Cursor.Free; Pro adds the run-out forecast and longer token totals. It reads only what the running app already knows, never contacts Anthropic, OpenAI or anyone else, and sends nothing. man runway has the rest.
runway
🟦🟦🟦🟦⬛ Claude · 82% · 20h ⏳ ↳ 🛬 At this pace it lasts to renewal 🟩🟩🟩⬛⬛ OpenAI/Codex · 59% · 4d ↳ ⚠️ 1 reset expires in 3d (3 total) ↳ 📉 At this pace you'll run out on Friday, about 21 hours before renewal
runway statusline
Claude 🟦🟦🟦🟦⬛ 82% · 20h ⏳ Codex 🟩🟩🟩⬛⬛ 59% · 4d
claude mcp add --scope user agent-runway -- ~/.local/bin/runway mcp
The same usage data behind Claude Code's /usage screen. When its sign-in needs refreshing, Runway runs claude -p /usage — a local command that costs zero tokens.
Runway asks the Codex CLI’s app-server for account limits. Codex setup helps find the local setup you use; recorded activity can remain available separately.
The CLIs' own session logs, and Cowork's, read in place on your Mac.
Read-only by design
Free and Pro
14 days of Pro to start, then the free version keeps going for as long as you like. Pro is $2.99 once: no subscription.
$0
runway command, for Terminal, status bars and AI agents$2.99 once
Everything in Free, plus:
Get started
Unzip AgentRunway.zip and drag Agent Runway to Applications.
Claude Code on a Claude plan, and/or Codex (npm i -g @openai/codex). For ChatGPT subscription limits, use your ChatGPT sign-in. Codex setup can help find your local installation and read saved activity separately.
A short guide checks your tools on first launch. After that Runway checks on its schedule and when you open it, and keeps itself up to date.
No. Every check is a read. Claude's fallback, claude -p /usage, is a local command with no model call: zero tokens, zero cost. Codex is asked only for account and rate-limit data. Extra accounts in their own folder or Codex home are read the same way. The one exception is optional and off until you switch it on: an extra Claude account read with its own token sends Anthropic one tiny request per check, which uses a sliver of that allowance.
App Store apps must run in Apple's sandbox, which forbids exactly what Runway needs: running the claude and codex command-line tools and reading their session logs. Instead it's signed with a Developer ID and notarized by Apple, so it opens without warnings and never expires.
Your numbers never leave it: usage, readings and token history stay in ~/Library/Application Support/AgentRunway. Runway reads your allowance only from Anthropic and OpenAI, using their CLIs’ sign-ins, and checks their public status pages for outages (you can turn that off). A daily update check shows our server the app version and your IP address, like any download. A Pro key is checked on your Mac, offline. There are no analytics and no account, and we have your email only if you buy Pro. The runway command sends nothing either, but an AI agent you let run it passes what it reads to its model provider.
No. Each notification comes once for each thing that happens: one when you're back, one for an early reset or a new credit, at most one run-out warning and one heads-up about spare allowance a week, and the morning report. Quick allowance drop starts off; turn it on and set a rule for each provider to get one alert per continuous drop. Two never say the same thing. Each has its own switch in Settings → Tune interruptions, which also tells you if macOS has them turned off for Agent Runway.
No. The pace is always the whole week's. When a model like Fable has its own weekly cap and it's tighter than the week, the card shows it in a quiet line; once it's used up, a banner says when it's back and that other models still work.
Yes. Install the runway command in Settings → Connect an AI agent, then give Claude Code a read-only get_runway tool with claude mcp add --scope user agent-runway -- ~/.local/bin/runway mcp (Settings has a Copy Command button). Claude Desktop, Codex and Cursor can run runway mcp too. It reads only what the running app already knows, so keep Agent Runway open. Whatever an agent reads goes to its model provider, like anything else you show it: Claude reading your Codex numbers sends them to Anthropic.
There's no personal weekly allowance to pace, so Runway says so instead of guessing — and the Activity tab still shows your token burn.
Agents read far more than they write. Every turn re-sends the conversation so far, served from the prompt cache, so cached input dominates. Output is usually well under 1% of everything processed, and around a tenth of the new tokens.
The percentage comes straight from the provider and is authoritative. The cells round to the nearest fifth, with any allowance at all keeping one cell lit and all five reserved for exactly 100%.
No. Agent Runway is an independent utility and isn't affiliated with or endorsed by either company.