Get more work from your AI coding plan.

Your limits in the status line Free

Every Claude Code prompt shows how much of each window you've used, when it runs out if you keep this pace, and today's even share of the week. Your own status line stays exactly as it was; limitcast adds to it.

Also shows your other Claude account when this one is nearly out, and a /compact nudge once a session re-sends more than ~150k tokens per call.

Runout forecast Free

Burn rate and the clock time each 5-hour and weekly window runs out, for Claude Code and Codex, and for Grok once you set a dollar cap or turn on its weekly usage. The weekly window is paced over the last full day, nights included, so one busy hour doesn't cry wolf. NEAR at 95%, even before there's a rate.

New in 0.14: when a window is on pace to run out, status ends with one line that names the one that runs out first and where else there is room.

Subagent guard

A burst of subagents is the most common way to lose the end of a window. When a window is on pace to run out within 30 minutes, the guard stops new subagents in Claude Code and Codex (and in Grok once you set a dollar cap). While a Claude window is merely tight, new subagents run on Sonnet instead of your main model. Once you're past 100% it steps aside, so paid extra usage keeps working. New in 0.18: a second subagent of the same type and the same prompt is held while the first is still running.

Available to everyone. Install with limitcast install --guard. The guard acts on new spawns; LIMITCAST_ALLOW=1 overrides it.

Your week, paced by the day New in 0.14

The today figure in your status line is your share of what is left of the week, split evenly over the days until it resets. Now it acts: with notifications on, you get one message a day when you have used 80% of today's share. Turn on limitcast budget today on and the guard also holds new subagents once today's share is used up, until local midnight.

The hold is off by default. A window already at 100% is left to the tool, and LIMITCAST_ALLOW=1 overrides it. results shows applied holds separately from historical previews.

A model for each kind of subagent New in 0.13

Most subagents just read and search, and they inherit your main model. Pin a kind to a cheaper one and every new subagent of that kind runs on it, for example Explore on Haiku. A spawn that names its own model is left alone, and while a window is tight a kind pinned to Opus still moves to Sonnet. results counts the subagents it moved.

Needs limitcast install --guard. All features are free; results records the actions taken. Whether a cheaper model costs less of your limit per token isn't something limitcast can tell, so check report --why after a week.

Subagent receipts and budgets New in 0.11

Each time a subagent finishes, you see what it cost against your 5-hour window, priced from your own readings (the model never sees it). Set a budget and the guard holds new subagents once they've used that share of the current window, until it resets.

New in 0.12: before a new subagent starts, you see what that kind has cost you so far. It appears once you have 3 priced runs, and only when a typical run is over 1% of a window or the window is near its limit. It never blocks.

Cold-cache warning New in 0.11

Come back to a big session after a long break and the prompt cache has expired: your next message re-sends the whole context as new tokens. limitcast holds that first message once and tells you what it will cost, so you can /compact first. Send it again and it goes through. The cache lifetime is 60 minutes by default (LIMITCAST_CACHE_MINUTES).

New in 0.18: after 5 idle minutes the same large session's next message is announced and still sent. The hold stays at 60 minutes. Codex and Grok do not get this warning.

Notifications: desktop, phone or Slack

One notification when a Claude window turns NEAR, AT_RISK or HIT, and one when it resets after a HIT, so you can step away and come back at the right time. On macOS they come from limitcast's own notifier app, listed under its name in System Settings. Linux uses notify-send.

New in 0.11: the same messages on your phone through ntfy (limitcast notify ntfy <topic>) or in a Slack or Discord channel (limitcast notify webhook <url>). Opt-in; only the notification text is sent, to the service you chose.

Trim advisor, with apply and undo

Measures what your latest session loaded before you typed (hook output, MCP servers, tool names, skills, agents, instruction files), names the ones you haven't used, and what each costs. One command hides the skills from the model and turns the servers off for that project; another puts everything back. If Claude Code overwrites the settings, the cuts come back at the next session. Codex has trim --codex, and new in 0.12 Grok has trim --grok: its system prompt and the MCP servers in ~/.grok/config.toml you haven't used.

New in 0.14: limitcast trim keep ego-browser puts a server or skill on a keep-list that no --apply ever turns off or hides, for Claude Code, Codex and Grok. It is free to set, and trim shows what it would keep.

Free shows the totals: your starting context and how much of it you never use.

Memory check Free New in 0.16

Claude Code loads only the first 200 lines or 25 KB of MEMORY.md into every session, so a note past that is never read. limitcast memory shows how much of the index fits, which links are broken, which topic files nothing links to and which are over 150 lines. doctor flags an index that is over the cap.

In Claude Code, /limitcast:memory-audit takes the sweep further (stale, contradictory and unlinked notes, with fixes proposed and nothing changed until you agree), and /limitcast:memory-split splits an oversized file into a short hub plus topic files after you approve the boundary.

Grok Free New in 0.12

limitcast install wraps Grok's status line: session model, context % and cost on one line, with the limitcast forecast under it. --guard adds a guard for spawn_subagent, and trim --grok measures what Grok loads before you type.

Grok stores what you spent, not a percentage, so the status line shows your 5-hour and 7-day spend until you set a dollar cap. With a cap, limitcast forecasts that window and the guard blocks new Grok subagents when it is on pace to run out.

New in 0.17: limitcast grok-usage on shows Grok's real weekly percentage. About every 5 minutes limitcast asks grok.com for your billing cycle with the login Grok already keeps in ~/.grok/auth.json. It is off by default, because it is the one place limitcast uses a login: the login goes only to grok.com and is never stored or logged. status, Grok's status line and the guard then forecast the weekly window like the Claude and Codex ones.

Results and the dashboard Free

Evidence from your own data: every limit you hit and how far ahead limitcast warned, the tokens your smaller starting context saved, priced against your own 5-hour limit, and what the guard stopped (free: would have stopped) and whether that window then ran out.

Put results, the current windows and the burn report on one page: limitcast results --html.

Burn report

Claude Code, Codex and Grok tokens by tool, day, project, model and session, read from each tool's own logs, plus the share that went to subagents. See which project or habit is eating your plan.

New in 0.13: the API-equivalent cost: what the same tokens would cost at the API's list prices, per tool, day, project, model and session. It's a yardstick for how much your plan covers, not a bill. Claude models are priced; any other model is left out and counted separately, never as $0. The dollar figure in the example is illustrative. Context reduction is estimated from characters and a recorded baseline; it is not measured cash saved. Blocks and cache notices are events, not proven prevented interruptions.

Does your plan fit? New in 0.13

limitcast keeps the peak of every 5-hour and weekly window you have. Tell it your plan and it says whether you reached your limits often enough to look at the next one, or used so little that a smaller one would have fit, with the monthly price difference. It also shows what your tokens would cost at API list prices, so you can see what your plan covers.

It needs a week or two of history before it recommends anything, and it cannot see extra-usage spend. Anthropic states the 5-hour multiplier for Max (5x or 20x Pro); the advisor assumes weekly limits scale the same way. Example output below; your numbers will differ.

Where your tokens went New in 0.12

The burn report says how much you used; this says what it was spent on. limitcast sorts the new tokens, the ones that move your limit (cached reads barely count), into subagents, session starts, cache rebuilds after an idle pause, tool results by tool and model output. It splits subagents by type and model, lists the single steps that cost the most, and says what to ask for instead.

Claude Code only. The first run reads your transcripts once and caches them. Models may count differently against your limit, so the model tables show token counts, not window shares.

Heaviest scheduled-job starts New in 0.12

Every run of a script or scheduled job that calls claude -p pays your whole starting context again. limitcast groups those runs by the first words of their prompt and shows, per job, how many runs, how many new tokens each start cost, how big the starting prompt was, how many starts were cold (the prompt cache didn't have them) and whether starting a job's runs closer together would help. It ends with how to launch a job leaner. New in 0.18: when two runs of a job both recorded a startup breakdown, the report names the piece that changed between them.

The example's job names are changed. The report shows the first 45 characters of each prompt, on your machine only: don't paste it publicly unredacted.

Alerts and a weekly summary

At the start of a session, a message to you (never the model) when your starting context grew by ~1k tokens or more, naming what grew: a new plugin, a longer CLAUDE.md, a chatty hook. Once a week, one line on recorded controls and estimated context reduction. `results` adds per-tool consumption and next steps; estimates are separate from actual usage and previews.

Large tool-output notice New in 0.12

A tool result enters your context once as new tokens and is then re-read on every later call, so one very large result is worth knowing about. When a single result is about 20k tokens or more, limitcast tells you (never the model), once per tool per session, with what to ask for next time. Turn it on with limitcast install --alerts; LIMITCAST_TOOL_TOKENS changes the threshold.

It runs on Bash, Read, Grep, WebFetch and MCP tools, and adds about 30 ms per call.

New in 0.18: the same install --alerts also checks before the call. After report --why or report --jobs, a Read, Grep, or MCP tool whose results averaged at least 5k tokens is held when the call sets no limit, or its limit is lowered when it is too high. Bash and WebFetch are left alone. Re-run install --alerts to add the check.

Limit-change alert New in 0.11

Every day limitcast re-measures how many tokens move your 5-hour meter by 1%, from your own readings. When that shifts by 15% or more against the weeks before, you get a notification, so you know when the limit itself changed rather than your habits.

Run it after the reset

Out of room with a long job still to run? Queue it: limitcast waits for the window to reset and starts it in the background.

Set up once, checked every day Free

install backs up your settings and uninstall restores them exactly. A daily background check tells you in the status line if anything breaks, including a Claude Code, Codex or Grok log-format change. Hooks survive Node upgrades. In Claude Code, the plugin adds /limitcast:status, /limitcast:results, /limitcast:trim, /limitcast:memory-audit and /limitcast:memory-split.