Codex & Claude usage limit reset API
A free, no-key JSON endpoint and an MCP server, both built for agents and bots — not just the web pages.
Endpoints
GET https://agentresets.com/api/v1/resets.json— this site's own public schema (documented below). Use this one for new integrations.POST https://agentresets.com/mcp— MCP server (JSON-RPC 2.0, Streamable HTTP), documented further down.GET https://agentresets.com/openapi.json— OpenAPI 3.1 description of both the/api/v1/resets.jsonand/mcpendpoints above, for anyone who wants a machine-readable spec instead of reading this page.https://agentresets.com/feed.xml— Atom feed, combined Codex + Claude; per-product feeds at/codex/feed.xmland/claude/feed.xml.GET https://agentresets.com/api/resets.json— legacy endpoint, a direct copy of the upstream field layout. Kept for existing integrations; new ones should use/api/v1/resets.jsoninstead.https://agentresets.com/reset.ics//codex/reset.ics//claude/reset.ics— ICS calendar subscription (Apple/Google Calendar can subscribe directly); lists already-recorded resets only, never a predicted or estimated one. Add combined calendar or download the .ics.
/api/v1/resets.json — schema
Top-level shape:
{
"schema_version": 1,
"generated_at": "2026-11-18T00:25:01Z",
"source": "https://agentresets.com",
"attribution": {
"note": "Reset history recorded before 2026-09-10 includes records seed-imported from public archives at Codex Resets (codex-resets.com) and Claude Resets (claude-resets.com).",
"cutoff_date": "2026-09-10",
"sources": [
{
"name": "Codex Resets",
"url": "https://codex-resets.com"
},
{
"name": "Claude Resets",
"url": "https://claude-resets.com"
}
]
},
"products": {
"codex": {
"name": "Codex",
"vendor": "OpenAI",
"events": [
{
"date": "2026-11-12",
"time": "2026-11-12T21:59:39Z",
"kind": "banked",
"summary": "We shipped some good updates to dots today and the banked reset has been activated and should land in everyones accounts sometime soon!"
}
],
"stats": {
"total": 62,
"last_date": "2026-10-28",
"last30d": 9,
"median_gap": 2,
"max_gap": 68,
"banked_total": 12,
"last_banked_date": "2026-11-12",
"banked_last30d": 4,
"unavailable": false
},
"forecast": {
"engine": "memoryless",
"brier": {
"memoryless": {
"brier_24h": 0.172,
"brier_48h": 0.2503,
"n": 245
},
"conditional": {
"brier_24h": 0.18,
"brier_48h": 0.264,
"n": 245
},
"climatology": {
"brier_24h": 0.1742,
"brier_48h": 0.2596,
"n": 245
}
},
"n": 245,
"source": "https://whenreset.app/api/forecast",
"model": "climatology (fraction of historical gaps <= horizon, unwindowed)",
"probabilities": {
"raw_24h": 0.1639,
"raw_48h": 0.4262,
"raw_7d": 0.7541
},
"hours_since_last": 501.7,
"sample_n": 61,
"method_url": "https://whenreset.app/tools/backtest/",
"whenreset_backtest": {
"brier_24h": 0.1362,
"brier_48h": 0.2019,
"cases": 1304
},
"unavailable": false
},
"announced_schedule": null
},
"claude": {
"name": "Claude",
"vendor": "Anthropic",
"events": [
{
"date": "2026-11-10",
"time": "2026-11-10T21:59:58Z",
"kind": "regular",
"summary": "We've given all Pro, Max, Ultra, and Team users one free full reset, across all models and limits. Use it anytime before December 10 at 2:00PM PT. We recently fixed a few small issues that used more limits than intended, and the Fable resets didn't do much for those of you who'd"
}
],
"stats": {
"total": 14,
"last_date": "2026-11-10",
"last30d": 1,
"median_gap": 8,
"max_gap": 67,
"banked_total": 2,
"last_banked_date": "2026-11-09",
"banked_last30d": 1,
"unavailable": false
},
"forecast": {
"engine": "climatology",
"brier": {
"memoryless": {
"brier_24h": 0.022,
"brier_48h": 0.0503,
"n": 56
},
"conditional": {
"brier_24h": 0.0463,
"brier_48h": 0.0969,
"n": 56
},
"climatology": {
"brier_24h": 0.021,
"brier_48h": 0.0482,
"n": 56
}
},
"n": 56,
"source": "https://whenreset.app/api/forecast",
"model": "climatology (fraction of historical gaps <= horizon, unwindowed)",
"probabilities": {
"raw_24h": 0.0769,
"raw_48h": 0.0769,
"raw_7d": 0.4615
},
"hours_since_last": 170.42,
"sample_n": 13,
"method_url": "https://whenreset.app/tools/backtest/",
"whenreset_backtest": null,
"unavailable": false
},
"announced_schedule": null
}
}
}
| Field | Type | Meaning |
|---|---|---|
schema_version | integer | Starts at 1. Only increases on a breaking change to this shape — see Versioning below. |
generated_at | string | ISO 8601 UTC timestamp of when this specific file was built. |
source | string | Always "https://agentresets.com" — where this data came from, for anything that re-shares it downstream. |
attribution | object | Data-provenance credit for the seed-imported history (see the Usage & attribution section below): note (string), cutoff_date (the seed-import date), sources (array of {name, url}). |
products | object | Keyed by product id: currently codex and claude. |
products.*.name | string | Display name (e.g. "Codex"). |
products.*.vendor | string | e.g. "OpenAI", "Anthropic". |
products.*.events | array | Reset events, newest first. |
products.*.events[].date | string | Announcement date, UTC, YYYY-MM-DD. |
products.*.events[].time | string or null | Exact ISO 8601 UTC timestamp of the announcement when known; null for older entries imported from public archives. |
products.*.events[].kind | string | "regular" (usage cleared immediately) or "banked" (credited, applies later). |
products.*.events[].summary | string | English summary of the announcement. |
products.*.events[].delivered_at | string, optional | ISO 8601 UTC. Present only when this specific reset's actual delivery/landing time was confirmed separately from the announcement in time above. Omitted, not null, when unknown — omission never implies the reset delivered immediately. |
products.*.events[].delivered_source | string, optional | A plain URL to the evidence for delivered_at. Omitted when no URL could be resolved. |
products.*.events[].delivery_note | string, optional | A short English note about the delivery. Omitted when there's no English note on record for this event; this field is never a machine translation of a non-English note. |
products.*.stats | object | Summary stats for the product (same numbers shown on the site's own pages). As of 2026-09-24, total/last30d/median_gap/max_gap are mirrored live from whenreset.app's own /api/v1/latest.json — the single forecast/rhythm source for this whole tracking matrix (this site, codexreset.ai, claudereset.org, whenreset.app, howtok.net) — and are null with unavailable: true when whenreset.app couldn't be reached this build; this site never substitutes a number it computed itself. last_date/last_banked_date/banked_* are unchanged: this site's own resets.json facts. See events[].kind above to tell regular and banked events apart in the raw event list. |
products.*.stats.total | integer or null | Total recorded global reset events for this product, via whenreset.app (excludes banked credits — see banked_total for those). null when unavailable this build. |
products.*.stats.last_date | string or null | Date of the most recent global reset. This site's own fact, unaffected by whenreset.app availability. |
products.*.stats.last30d | integer or null | Global resets in the trailing 30 days, via whenreset.app. null when unavailable this build. |
products.*.stats.median_gap | integer or null | Median number of days between consecutive global resets ("typical gap"), via whenreset.app. null when unavailable this build or on too little history. |
products.*.stats.max_gap | integer or null | Longest gap in days between consecutive global resets on record, via whenreset.app. null when unavailable this build. |
products.*.stats.banked_total | integer | Total recorded banked reset credits for this product (not counted in total above). This site's own fact. |
products.*.stats.last_banked_date | string or null | Date of the most recent banked reset credit, independent of last_date. This site's own fact. |
products.*.stats.banked_last30d | integer | Banked reset credits in the trailing 30 days. This site's own fact. |
products.*.stats.unavailable | boolean | true when whenreset.app's /api/v1/latest.json couldn't be reached this build (total/last30d/median_gap/max_gap are all null in that case). |
products.*.forecast | object | This product's live reset-probability estimate, via whenreset.app, plus this site's own independent research backtest. See /tools/reset-odds/ for the full method. |
products.*.forecast.source | string | Always "https://whenreset.app/api/forecast" — where model/probabilities/hours_since_last/sample_n/method_url/whenreset_backtest below come from. |
products.*.forecast.model | string or null | whenreset.app's own model name for this product (e.g. a climatology base-rate description). null when unavailable. |
products.*.forecast.probabilities | object or null | {raw_24h, raw_48h, raw_7d}, each a 0–1 fraction — whenreset.app's live reset-probability estimate. null when unavailable this build. |
products.*.forecast.hours_since_last | number or null | Hours since the last global reset, per whenreset.app. |
products.*.forecast.sample_n | integer or null | Historical interval sample size whenreset.app's estimate is based on. |
products.*.forecast.method_url | string or null | Link to whenreset.app's own method/backtest page for this estimate. |
products.*.forecast.whenreset_backtest | object or null | whenreset.app's own published {brier_24h, brier_48h, cases} for this product (currently published for Codex only); null when not published or unavailable. |
products.*.forecast.unavailable | boolean | true when whenreset.app's /api/forecast couldn't be reached this build (the fields above are all null in that case). |
products.*.forecast.engine | string | Research only, not used for display since 2026-09-24: "memoryless", "conditional", or "climatology" — whichever scored the lowest combined 24h/48h Brier score on this site's own independent walk-forward backtest of this product's history. See /tools/reset-odds/'s "Independent backtest" section. |
products.*.forecast.brier | object | Research only. Keyed by algorithm name (memoryless/conditional/climatology), each {brier_24h, brier_48h, n} from this site's own backtest — lower Brier is better. Empty object {} when there isn't enough history yet. |
products.*.forecast.n | integer | Research only. Number of historical days this site's own walk-forward backtest actually scored (0 when unavailable). |
products.*.announced_schedule | object or null | OPTIONAL (added 2026-10-02). An officially announced, not-yet-recorded reset time for this product — null in the normal case. Only ever non-null when the vendor explicitly announced an exact UTC time (a merely hinted signal, or an announced schedule with only a day and no exact time, is never represented here — the whole field stays null for either). Also clears back to null once a matching regular reset event has actually been recorded, or more than 24 hours after the scheduled time passes with nothing recorded. |
products.*.announced_schedule.scheduled_for | string | ISO 8601 UTC timestamp of the officially announced planned reset time. |
products.*.announced_schedule.announced_at | string or null | ISO 8601 UTC timestamp of when the vendor posted the announcement, when known. |
products.*.announced_schedule.source_url | string or null | Link to the vendor's own announcement post. null when no usable link was given. |
products.*.announced_schedule.certainty | string | Always "announced" when this object is present at all. |
This is a custom schema maintained by this site, not a passthrough of howtok's internal data format — fields not listed here (including any Chinese-language field) are never included.
Changelog
- 2026-09-23: stats count global (regular) resets only; banked credits are reported separately.
- 2026-09-23: added the top-level
attributionfield (additive, no field removed or renamed). - 2026-09-24: added the per-product
forecastfield (additive) — the walk-forward-backtest-selected probability engine and its Brier scores. - 2026-09-24 (single forecast source):
stats.total/.last30d/.median_gap/.max_gapchanged meaning — they now mirror whenreset.app's own numbers instead of this site's own computation (field names unchanged; a newstats.unavailablefield was added).forecast.engine/.brier/.nare unchanged in name and meaning but are now research-only and no longer used to pick any displayed number; new, additiveforecast.source/.model/.probabilities/.hours_since_last/.sample_n/.method_url/.whenreset_backtest/.unavailablefields carry whenreset.app's own live forecast, the numbers this site actually displays. See /tools/reset-odds/. - 2026-09-27: added optional per-event
delivered_at/delivered_source/delivery_notefields (additive, omitted rather than null when unknown — an event with no delivery data on record has the exact same key set as before this change). These describe when/how a specific reset was confirmed actually delivered, separately from the announcementtime; their absence does not mean the reset delivered immediately. - 2026-10-02: added the per-product
announced_schedulefield (additive,nullin the normal case) — an officially announced, exact, not-yet-recorded reset time, when one exists. A merely hinted signal, or an announced schedule with only a day and no exact time, is never represented here.
Example: curl
curl -s https://agentresets.com/api/v1/resets.json | jq '.products.codex.events[0]'
Example: JavaScript fetch
const r = await fetch("https://agentresets.com/api/v1/resets.json").then(r => r.json());
const last = r.products.codex.events[0];
console.log(`Last Codex reset: ${last.date} (${last.kind})`);
console.log(`Codex median gap: ${r.products.codex.stats.median_gap} days`);
MCP server
A read-only Model Context Protocol server at https://agentresets.com/mcp, so an agent can query reset history directly instead of scraping a page. It's stateless JSON-RPC 2.0 over plain HTTP POST (Streamable HTTP without the SSE stream) — no API key, no session, no write operations.
Add it to Claude Code:
claude mcp add --transport http agentresets https://agentresets.com/mcp
Or a manual client config:
{
"mcpServers": {
"agentresets": {
"url": "https://agentresets.com/mcp"
}
}
}
Tools
| Tool | Arguments | Returns |
|---|---|---|
list_reset_events | product (optional: "codex"/"claude"), limit (optional, default 20) | Reset events, newest first, each optionally carrying delivered_at/delivered_source/delivery_note (see the field dictionary above; omitted, not null, when there's no delivery data on record for that event). |
get_reset_stats | product (required) | The same stats object documented above for /api/v1/resets.json, plus announced_schedule (null unless the vendor has officially announced an exact, not-yet-recorded reset time — see the field dictionary above). |
get_reset_odds | product (required) | Probability of a reset within 24h/48h/7d. As of 2026-09-24 this is a live, server-side passthrough of https://whenreset.app/api/forecast — the single forecast source for this whole tracking matrix, same numbers as the "Current numbers, via whenreset.app" section of /tools/reset-odds/. The response's source field names the upstream endpoint. Returns an MCP tool-level error (isError: true) if whenreset.app can't be reached, never a locally computed fallback. |
The MCP server has no accounts, no auth, and no way to write anything — every tool call only reads this site's own /api/v1/resets.json.
OpenAPI spec
A machine-readable OpenAPI 3.1 description covering both GET /api/v1/resets.json and POST /mcp, including the full stats shape above and the three MCP tool names, is published at:
GET https://agentresets.com/openapi.json
Plain static JSON, no interactive explorer bundled on this page — feed the URL into your own tooling (codegen, an API client, an OpenAPI-aware agent) instead.
Update frequency
The upstream tracker checks its sources roughly every 20 minutes, but events only grows when a new official announcement is actually posted — long stretches with no new event are normal and don't mean the feed stalled. generated_at tells you exactly when a given file was produced; it moves even between reset announcements, since the site rebuilds on every check.
Versioning
Current version: v1. A breaking change to this shape (removed/renamed field, changed type, changed meaning) will ship as a new /api/v2/resets.json path, with /api/v1/resets.json left running unchanged — existing integrations are never broken in place. Adding a new field, or a new product id, is not considered breaking.
Usage & attribution
Free to use, no key. If you publish something built on this data, please link back to agentresets.com. Cached 5 minutes at the edge; please cache client-side too rather than polling in a tight loop.
Reset history before 2026-09-10 includes records from Codex Resets and Claude Resets. That's the same data behind the attribution object documented above. If you reuse or republish this data, please keep that attribution alongside it, in addition to linking back to agentresets.com.