Sipcode is a command-line tool for Claude Code and Codex, which run on your computer. This site is built for the same screen.
Open this page on your laptop or desktop for the full experience.
A free, open-source (MIT) command-line tool that reads your local Claude Code and OpenAI Codex logs and shows where your tokens and usage limits go.
Zero network calls. Nothing leaves your laptop.
Launched on Product Hunt on Tuesday, June 23, 2026. Finished around #12 of 1,086 products that day (134 points, 28 comments).
62.6% is the Claude Code proxy on a locked, synthetic 20-task corpus. Your own numbers come from your own logs.
Every request re-sends the conversation so far. Files get read again. Idle context rides along. Usage limits change quietly, and an update can make things worse. Sipcode makes all of it visible from your own logs, with numbers checked against an independent counter that shares no code with it.
Bloated context costs quality too. Same task, same model:
Add a sign-out button to the navbar.
Adding a SignOutButton to src/NavBar.tsx, wired to the existing useSession() hook. Diff below.
Add a sign-out button to the navbar.
There are a few places this could live. Earlier you mentioned a sidebar, but the project also has a top nav. Should I add it to both? I also want to double-check the session hook still exists.
sharper outcomes when context stays lean. About 40% fewer agent errors under the same conditions.
Cited from published research. Not measured by Sipcode.
Answer quality drops measurably as input tokens climb. Even on tasks the same model aced at small context sizes.
A hook rewrites bulky commands to their compact form before Claude ever sees them. Claude Code only: Codex setup (rules, proxy) is not supported yet.
sipcode proxy --install claude code · zero config
Reads your local Claude Code and Codex logs. stats, today, forecast, trend and impact show one section per tool; why and receipt take your newest session from either. --agent codex or --agent claude-code picks one tool.
sipcode stats · why · today claude code + codex · zero network calls Silent unless your latest Claude Code or Codex session drifts off your baseline. Names what moved and how to fix it.
sipcode drift silent unless it matters Codex writes your plan's usage-limit readings into its own logs. Sipcode reads them, and the rest of your Codex history, on your machine. For Codex it only reads: it does not change what Codex does.
On the maintainer's machine, 1% of the weekly limit bought about 425K tokens (week of Mar 28), 357K (Apr 1), then 77K (week of Apr 13): about 5x fewer. Sipcode flags that as a possible limit cut, not a proven one.
Checks your own Claude Code and Codex logs and pins each finding to the version that brought it, comparing every version with your other versions: prompt-cache misses, cache written with a 5-minute lifetime instead of 1 hour, large uncached input after tool results, the whole cache rebuilt about every hour, and a Codex version using more of your limit per token.
--bug-report adds a block to paste into a GitHub issue: token counts, versions, dates and short session ids only. No message text.
Works in the ChatGPT desktop app and the Codex CLI. It runs the sipcode command on your computer: no hooks, no MCP server, no network calls. Setup below.
npx sipcode benchmark What this means: your newest session is behaving differently from your recent norm, in ways that waste tokens and can make Claude less reliable.
(This is "context rot": answer quality drops as context gets bloated.)
Signals that regressed (2):
Each step is sending far more context than your norm. Bloated context costs more tokens and can bury the detail Claude needs.
→ Fix: Start a fresh chat for your next task to reset the context.
Much less of your context is being reused from cache (about 10x cheaper).
→ Fix: Avoid changing MCP servers or config mid-task.
Install the package, then read your logs.
$ npm i -g sipcode
$ sipcode stats
$ sipcode limits
$ sipcode why
Node.js 20 or newer. Reading the logs Codex compresses (.jsonl.zst) needs Node.js 22.15 or newer. sipcode limits is Codex only.
After npm i -g sipcode, add the plugin. npx runs the Codex CLI for you, so you do not need it installed.
$ npx -y @openai/codex plugin marketplace add Anuj7411/sipcode
$ npx -y @openai/codex plugin add sipcode@sipcodeRestart the ChatGPT desktop app, then ask in a new chat: "how close am I to my Codex limit?"
Turn on the hook that trims bulky commands before Claude sees them.
$ sipcode proxy --installRestart Claude Code. This part is Claude Code only: Codex setup (rules, proxy) is not supported yet.
Add the MCP server to your config, then restart Desktop.
{
"mcpServers": {
"sipcode": { "command": "npx", "args": ["-y", "-p", "sipcode", "sipcode-mcp"] }
}
}After restart, ask Claude "what sipcode tools do you have?" to confirm.
The npm route is what we ship. No curl | sh, no one-click installer that does nothing.
The Desktop pipeline is closed. The proxy only works in Claude Code. Desktop gets the analytics tools.
For Codex, Sipcode reads the logs: spend, usage limits, regressions. Codex setup (rules, proxy) is not supported yet.
Sipcode once over-counted Claude Code tokens by 11.4% on one machine: resumed sessions copy earlier requests, and both copies were summed. Fixed in 1.7.0. An independent counter that shares no code with Sipcode cross-checks every period command against the same logs.
npm install -g sipcode puts the package on your machine. Tokens drop after you run sipcode proxy --install.
The 29% and 40% reliability numbers come from published research, cited not measured. We measure the tokens we save.
No telemetry. No network calls in any core path. A CI guard fails the build if that ever changes.
A free, open-source (MIT) command-line tool that reads your local Claude Code and OpenAI Codex logs and shows where your tokens and usage limits go: spend per day and month, the largest requests in a session, drift from your own baseline, Codex usage limits, and regressions after an update. For Claude Code it can also act: a proxy hook trims bulky tool output before it reaches the model, re-read dedup refuses to reload files Claude already has, and AST-aware reads return only the symbol Claude searched for. Zero network calls.
Yes, for reading. stats, today, forecast, why, drift, trend, impact and receipt read Codex logs next to Claude Code's; add --agent codex or --agent claude-code for one tool. sipcode limits shows your Codex usage limits, sipcode regress checks whether an update made things worse, and the Codex plugin lets you ask Codex about all of it in plain words. Codex setup (rules, proxy) is not supported yet, so Sipcode does not change what Codex does.
No. Sipcode makes zero network calls during normal use and reads only files on your machine. A privacy test fails the build if network code is imported into the source. To make repeat runs fast it keeps small caches under ~/.sipcode/ (token counts, models, timestamps). They never hold message text and are never uploaded.
sipcode limits counts how many tokens 1% of a Codex limit bought in each finished period, comparing periods on the same model. When a period bought at most half the median of at least two earlier periods, it is flagged as a possible limit cut. It always says "possible": limits are account-wide, and Codex use on your other devices or in the cloud raises the percent without leaving tokens in your local logs, which looks the same. On the maintainer's machine, 1% of the weekly limit bought about 425K tokens (week of Mar 28), 357K (Apr 1), then 77K (week of Apr 13): about 5x fewer, flagged as possible.
No. Install Sipcode once with npm i -g sipcode, then run npx -y @openai/codex plugin marketplace add Anuj7411/sipcode and npx -y @openai/codex plugin add sipcode@sipcode. npx runs OpenAI's Codex CLI without installing it. Restart the ChatGPT desktop app and ask in a new chat. If you already have the Codex CLI, run the same two commands with codex in place of npx -y @openai/codex.
Yes. Sipcode ships 16 MCP tools. Ask "how am I doing today?" and Claude routes to get_today_summary: spend so far plus comparison against your adaptive 30/14/7/3 day median. Ask "how much will I spend this month?" and it routes to forecast_monthly_spend: projected month-end with a confidence band. No SDK, no account, no telemetry; Sipcode reads the local .jsonl files Claude Code already writes.
Sipcode covers terminal use through Claude Code and chat use through Claude Desktop via 16 MCP tools. RTK is terminal only. Beyond coverage, Sipcode adds reliability infrastructure RTK does not have: a drift detector that catches context rot live, re-read dedup that refuses to reload files already in context, integrity scoring that tells you how much signal a rewrite kept, and AST-aware reads that return only the symbol the model searched for. You can use both. They do not conflict.
For the Claude Code proxy, the reproducible benchmark shows a 62.6 percent median. Its 20 tasks are synthetic, scripted transcripts rather than live sessions, so your number will depend on your workload. After install, sipcode proxy --stats shows the exact tokens rewritten on your machine.
Research shows leaner context lifts agent reliability. We cite that research. Sipcode measures the tokens it saves, not the quality lift, so we do not put a number on the quality side. You will judge for yourself.
Yes. MIT licensed. No account. No telemetry. No paid tier.
A gulp is a context dumped into Claude without restraint. It costs more and makes the model worse. A sip is the same information, measured. Sipcode keeps the intake at a sip.
Reads your Claude Code and Codex logs on your machine. No account. No telemetry. About 60 seconds to set up.