Pasting a whole repo into ChatGPT or Claude from VS Code: Send to AI vs Repomix vs gitingest
Every AI chat has a bigger context window than it did a year ago, and every developer I know still copies files into the chat one at a time. The tools that fix this fall into two groups: command-line bundlers that turn a repository into one text file, and editor extensions that do the same from inside VS Code. I maintain one of the latter, so take the comparison with that in mind, and check the numbers yourself; every claim below has a link.
What changed in 2026, and why it matters for all three tools
The limits moved, and the chat apps changed how they handle big pastes.
Context windows, per plan. These are total windows, input plus output. ChatGPT gives Free users 27K tokens on instant models, Go/Plus/Business 54K on instant and 256K on thinking models, and Pro 128K/400K (chatgpt.com, tabulated by ToolColumn). In the claude.ai chat app, paid plans give 1M to Fable 5.1, Opus 5/5.5, Sonnet 5/5.5 and Haiku 5.5, 500K to Fable 5 and the 4.6 to 4.8 generation, and 200K to older models (Claude help; Claude Code gets 1M on all of them). Gemini without a plan is 32K, AI Plus 128K, AI Pro and Ultra 1M (Google help). A tool that still assumes "8K for ChatGPT Free" trims far more than it needs to.
Big pastes become attachments. ChatGPT started converting pastes over 5,000 characters into a file attachment for Plus, Pro and Business in March 2026 (release note coverage); since June it does it on every plan, with the threshold raised to 10,000 characters (ChatGPT release notes). Claude's web app does the same for large pastes (Simon Willison); Anthropic does not publish the cutoff, and user tests put it a little under 4,000 characters. So whichever tool you use, a bundle arrives as a file. That is fine, the model reads all of it, but it means "save as .txt and attach" and "paste" are now the same thing.
Code tokenises differently on every model. The 4-characters-per-token rule of thumb is roughly right for OpenAI's tokenizer on code, and badly wrong for Claude. One measurement of the same TypeScript file gives about 4.2 characters per token on GPT, 3.7 on Gemini, and 2.4 on the tokenizer that Claude 4.7 and later use (playcode.io). Anthropic's own docs say that newer tokenizer produces roughly 30 percent more tokens than earlier Claude models for the same text (Anthropic token counting). So a tool that estimates at chars/4 is close for ChatGPT and about 40 percent low for Claude, and a bundle that "fits" by that estimate may not.
The three tools
| | Repomix | gitingest | Send to AI |
|---|---|---|---|
| Runs | CLI, MCP server, browser extension, web app, GitHub Action | web app, CLI and Python package, browser extension | VS Code sidebar |
| Output | XML, Markdown, JSON, plain | plain text digest | Standard, Claude XML, Compact, Minimal |
| Picks files by | include/exclude globs, .gitignore | include/exclude patterns, file-size cap, .gitignore | tick boxes, Open Tabs, Git Changes, relevance to your prompt |
| Token count | yes, o200k | yes, o200k | yes, 3.3 chars/token, between GPT and Claude |
| Bundles the Problems panel | no | no | yes, "Bundle Errors" |
| Applies the AI's reply back | no | no | yes, with diff preview and undo |
| Price | free, MIT | free, MIT | free tier; Pro $14.99 once, Team $39 |
| Stars/installs (Nov 2026) | 29k stars | 16k stars | about 200 installs, new |
Repomix (yamadashy/repomix) is the most complete bundler: a mature CLI with an MCP server so agents can call it, a browser extension, a web app at repomix.com, a GitHub Action, and a long list of output options. If you want bundles in CI, or you drive an agent framework, it is the right tool. gitingest (coderamp-labs/gitingest) is the fastest way to get a digest of a public repository without installing anything: swap hub for ingest in any github.com URL. Neither touches your editor, and neither knows what you are working on right now.
Send to AI (Marketplace, Open VSX) is built for the daily loop inside the editor. You tick files, type what you want, and the bundle lands on the clipboard with a tree and a token count. Two things the CLIs cannot do because they do not run inside VS Code:
- Bundle Errors reads the Problems panel, collects every file with an error or warning plus the exact messages, and writes a prompt that says "fix these". One paste instead of an afternoon.
- Apply AI Reply parses the chat's answer, whether it came back as full files in fenced blocks or as SEARCH/REPLACE blocks, shows a diff preview, and writes the edits as one undoable change.
!Send to AI demo
The half nobody automates: getting the answer back into the files
Bundling is the easy half. The hard half is applying what the model says. The agent tools solved this with their own edit formats: aider uses SEARCH/REPLACE blocks, and on its polyglot leaderboard GPT-5 at high reasoning produced a well-formed edit 91.6% of the time, Gemini 2.5 Pro 99.6 to 100% using the variant with the path inside the fence, and Claude Opus 4 97.3 to 98.7% (aider leaderboards). Line-numbered diffs are the least reliable because models miscount; OpenAI's own guide says its apply_patch format "does not use line numbers, as the context is enough to uniquely identify code" (GPT-4.1 prompting guide, explained in depth in this V4A write-up).
Send to AI accepts both shapes. Full-file blocks are parsed from the fence line, the line above it, or a first-line comment. SEARCH/REPLACE blocks are matched exactly first, then with whitespace and line-ending differences ignored, and anything ambiguous is reported rather than guessed. The bundle itself ends with a short instruction asking the model to reply in one of those two shapes, which in practice is what makes the apply step reliable. If a match fails, you see it in the file picker before anything is written.
When to use which
- Repomix when you want bundles in scripts, CI or an agent pipeline, or you need XML for a specific prompt format.
- gitingest when the code is public and you just need a quick digest link.
- Send to AI when you are in VS Code, on a chat plan with no API access, and you want the round trip: errors in, fixes out, applied with undo.
Honest limits of Send to AI
The free tier bundles the 50 most relevant files for your prompt rather than the whole project; Pro ($14.99 once, five machines, 30-day refund) removes the cap and adds git-diff mode, presets and fit-to-window trimming. A Team pack ($39 once) covers five developers on one key. There is no agent mode and no API integration by design: it is for people who paste into a chat. And it only runs in VS Code and its forks.
If you try it and something breaks, the issues page is where I look first.
.
How to give ChatGPT your whole codebase in one file with Repomix
ChatGPT is good at reviewing code, explaining an unfamiliar project and writing tests. It is bad at guessing what is in the 40 files you did not paste. Copying files one by one is slow, you lose the folder structure, and you never know how close you are to the context limit.
Repomix packs an entire repository into a single AI-friendly file with one command. It respects your .gitignore, counts tokens per file and for the whole repo, scans for secrets with Secretlint so credentials stay out of the output, and can compress code down to its structure with Tree-sitter. You then upload that one file to ChatGPT, Claude, Gemini or any other model.
This guide walks through the workflow end to end.
Step 1: pack your repo
You do not even need to install it. In your project folder:
npx repomix@latest
If you will use it often, install it once. On a Mac, Homebrew works:
brew install repomix
npm (npm install -g repomix), yarn and bun work too. Then just run:
repomix
Repomix writes repomix-output.xml to the current folder: a file summary, the directory structure and every file's contents. The terminal output shows the token count and your five largest files, so you know before uploading whether it will fit.
Step 2: hand it to ChatGPT
Upload repomix-output.xml to a ChatGPT chat (or paste it if it is small) and start with a prompt like the one the README suggests:
This file contains all the files in the repository combined into one.
I want to refactor the code, so please review it first.
Want to skip the upload? --copy puts the output straight on your clipboard:
repomix --copy
Step 3: keep it inside the context window
Big repos are where most people give up. Repomix has four levers for this.
Pack only what matters. Use glob patterns to include or ignore paths:
repomix --include "src/**/*.ts,**/*.md"
repomix --ignore "**/*.log,tmp/"
Compress. --compress uses Tree-sitter to keep the essential structure (classes, functions, interfaces) and drop the rest. Great for "explain the architecture" questions where the model does not need every line:
repomix --compress
Strip noise. --remove-comments and --remove-empty-lines cut tokens further.
Split it. If a pack is still too big for one upload, --split-output writes numbered files of a size you choose:
repomix --split-output 2mb
To find the heavy folders first, --token-count-tree prints your file tree with token counts. The default counting uses the o200k_base tokenizer, which the README lists as GPT-4o's, and you can change it with --token-count-encoding.
Step 4: pick a format the model reads well
XML is the default. You can also ask for Markdown, JSON or plain text:
repomix --style markdown
Markdown is easy to skim yourself before you upload. JSON is useful if you want to query the pack with jq.
Pack a GitHub repo you have not cloned
This is the underrated feature. Want ChatGPT to explain a library before you adopt it? Point Repomix at the repo:
repomix --remote yamadashy/repomix
It accepts full GitHub URLs, the user/repo shorthand, and a branch, tag or commit with --remote-branch. Combine it with compression for a quick architecture overview:
repomix --remote yamadashy/repomix --compress
Then use the README's library overview prompt:
This file contains the entire codebase of library. Please provide a comprehensive overview of the library, including its main purpose, key features, and overall architecture.
For security, Repomix does not load config files from remote repos unless you pass --remote-trust-config.
Give the model history, not just code
When you are asking "why did this break?", recent changes matter more than the full source. Repomix can add git context:
repomix --include-logs --include-diffs
--include-logs adds the last 50 commits by default, with dates, messages and changed files. --include-diffs adds your working tree and staged changes.
Prompts worth saving
The README ships a few prompts you can paste after uploading:
- Code review: "This file contains my entire codebase. Please review the overall structure and suggest any improvements or refactoring opportunities, focusing on maintainability and scalability."
- Docs: "Based on the codebase in this file, please generate a detailed README.md that includes an overview of the project, its main features, setup instructions, and usage examples."
- Tests: "Analyze the code in this file and suggest a comprehensive set of unit tests for the main functions and classes. Include edge cases and potential error scenarios."
Save your settings
Once you find flags you like, write them to a config file so every run uses them:
repomix --init
That creates repomix.config.json in the project. Add --global to make it your default everywhere.
No terminal? Three other ways in
- Website: repomix.com packs a public repo in the browser, with format options and a token estimate.
- Browser extension: Chrome and Firefox add-ons put a Repomix button on GitHub repo pages.
- VS Code: the community extension Repomix Runner runs it on any folder from the editor.
Going further
We wrote up this workflow in a shorter form, with who it suits and where it falls down: Repomix tutorial: pack your codebase into one file for an LLM. The Repomix repo page on Deduped collects other ways people use it.
If your real goal is to understand a codebase you just inherited rather than chat about it, compare this with how to map an unfamiliar codebase in one run with Graphify.
Deduped tracks the AI repos people are actually building with and writes up how to use them. Repo: yamadashy/repomix.
---
Get one email every Monday with the week's hottest AI repos and exactly how to use them: subscribe free at dedupedai.com
Mentions
要让 AI 读懂 GitHub 项目，不用逐个复制代码。 Repomix 把仓库打成一份文件：打开官网，贴公开仓库链接，点 Pack，下载后交给 AI 解释目录、定位入口。 CLI 免费开源，有中文文档；AI 服务另算。私有代码上传前，先检查有没有密钥。 🔗 repomix.com GitHub：yamadashy/repomix
Pasting a whole repo into ChatGPT or Claude from VS Code: Send to AI vs Repomix vs gitingest Every AI chat has a bigger context window than it did a year ago, and every developer I know still copies files into the chat one at a time. The tools that fix this fall into two groups: command-line bundlers that turn a repository into one text file, and editor extensions that do the same from inside VS Code. I maintain one of the latter, so take the comparison with that in mind, and check the numbers yourself; every claim below has a link. What changed in 2026, and why it matters for all three tools The limits moved, and the chat apps changed how they handle big pastes. Context windows, per plan. These are total windows, input plus output. ChatGPT gives Free users 27K tokens on instant models, Go/Plus/Business 54K on instant and 256K on thinking models, and Pro 128K/400K (chatgpt.com, tabulated by ToolColumn). In the claude.ai chat app, paid plans give 1M to Fable 5.1, Opus 5/5.5, Sonnet 5/5.5 and Haiku 5.5, 500K to Fable 5 and the 4.6 to 4.8 generation, and 200K to older models (Claude help; Claude Code gets 1M on all of them). Gemini without a plan is 32K, AI Plus 128K, AI Pro and Ultra 1M (Google help). A tool that still assumes "8K for ChatGPT Free" trims far more than it needs to. Big pastes become attachments. ChatGPT started converting pastes over 5,000 characters into a file attachment for Plus, Pro and Business in March 2026 (release note coverage); since June it does it on every plan, with the threshold raised to 10,000 characters (ChatGPT release notes). Claude's web app does the same for large pastes (Simon Willison); Anthropic does not publish the cutoff, and user tests put it a little under 4,000 characters. So whichever tool you use, a bundle arrives as a file. That is fine, the model reads all of it, but it means "save as .txt and attach" and "paste" are now the same thing. Code tokenises differently on every model. The 4-characters-per-token rule of thumb is roughly right for OpenAI's tokenizer on code, and badly wrong for Claude. One measurement of the same TypeScript file gives about 4.2 characters per token on GPT, 3.7 on Gemini, and 2.4 on the tokenizer that Claude 4.7 and later use (playcode.io). Anthropic's own docs say that newer tokenizer produces roughly 30 percent more tokens than earlier Claude models for the same text (Anthropic token counting). So a tool that estimates at chars/4 is close for ChatGPT and about 40 percent low for Claude, and a bundle that "fits" by that estimate may not. The three tools | | Repomix | gitingest | Send to AI | |---|---|---|---| | Runs | CLI, MCP server, browser extension, web app, GitHub Action | web app, CLI and Python package, browser extension | VS Code sidebar | | Output | XML, Markdown, JSON, plain | plain text digest | Standard, Claude XML, Compact, Minimal | | Picks files by | include/exclude globs, .gitignore | include/exclude patterns, file-size cap, .gitignore | tick boxes, Open Tabs, Git Changes, relevance to your prompt | | Token count | yes, o200k | yes, o200k | yes, 3.3 chars/token, between GPT and Claude | | Bundles the Problems panel | no | no | yes, "Bundle Errors" | | Applies the AI's reply back | no | no | yes, with diff preview and undo | | Price | free, MIT | free, MIT | free tier; Pro $14.99 once, Team $39 | | Stars/installs (Nov 2026) | 29k stars | 16k stars | about 200 installs, new | Repomix (yamadashy/repomix) is the most complete bundler: a mature CLI with an MCP server so agents can call it, a browser extension, a web app at repomix.com, a GitHub Action, and a long list of output options. If you want bundles in CI, or you drive an agent framework, it is the right tool. gitingest (coderamp-labs/gitingest) is the fastest way to get a digest of a public repository without installing anything: swap hub for ingest in any github.com URL. Neither touches your editor, and neither knows what you are working on right now. Send to AI (Marketplace, Open VSX) is built for the daily loop inside the editor. You tick files, type what you want, and the bundle lands on the clipboard with a tree and a token count. Two things the CLIs cannot do because they do not run inside VS Code: - Bundle Errors reads the Problems panel, collects every file with an error or warning plus the exact messages, and writes a prompt that says "fix these". One paste instead of an afternoon. - Apply AI Reply parses the chat's answer, whether it came back as full files in fenced blocks or as SEARCH/REPLACE blocks, shows a diff preview, and writes the edits as one undoable change. !Send to AI demo The half nobody automates: getting the answer back into the files Bundling is the easy half. The hard half is applying what the model says. The agent tools solved this with their own edit formats: aider uses SEARCH/REPLACE blocks, and on its polyglot leaderboard GPT-5 at high reasoning produced a well-formed edit 91.6% of the time, Gemini 2.5 Pro 99.6 to 100% using the variant with the path inside the fence, and Claude Opus 4 97.3 to 98.7% (aider leaderboards). Line-numbered diffs are the least reliable because models miscount; OpenAI's own guide says its apply_patch format "does not use line numbers, as the context is enough to uniquely identify code" (GPT-4.1 prompting guide, explained in depth in this V4A write-up). Send to AI accepts both shapes. Full-file blocks are parsed from the fence line, the line above it, or a first-line comment. SEARCH/REPLACE blocks are matched exactly first, then with whitespace and line-ending differences ignored, and anything ambiguous is reported rather than guessed. The bundle itself ends with a short instruction asking the model to reply in one of those two shapes, which in practice is what makes the apply step reliable. If a match fails, you see it in the file picker before anything is written. When to use which - Repomix when you want bundles in scripts, CI or an agent pipeline, or you need XML for a specific prompt format. - gitingest when the code is public and you just need a quick digest link. - Send to AI when you are in VS Code, on a chat plan with no API access, and you want the round trip: errors in, fixes out, applied with undo. Honest limits of Send to AI The free tier bundles the 50 most relevant files for your prompt rather than the whole project; Pro ($14.99 once, five machines, 30-day refund) removes the cap and adds git-diff mode, presets and fit-to-window trimming. A Team pack ($39 once) covers five developers on one key. There is no agent mode and no API integration by design: it is for people who paste into a chat. And it only runs in VS Code and its forks. If you try it and something breaks, the issues page is where I look first. .
Works with Claude, ChatGPT, Gemini, Grok and more, plus a built-in MCP server (repomix --mcp) for MCP clients. MIT license, 81 contributors. Helps with code reviews, refactors and onboarding. Limit: huge repos still overflow context, so filter files. yamadashy/repomix
How to give ChatGPT your whole codebase in one file with Repomix ChatGPT is good at reviewing code, explaining an unfamiliar project and writing tests. It is bad at guessing what is in the 40 files you did not paste. Copying files one by one is slow, you lose the folder structure, and you never know how close you are to the context limit. Repomix packs an entire repository into a single AI-friendly file with one command. It respects your .gitignore, counts tokens per file and for the whole repo, scans for secrets with Secretlint so credentials stay out of the output, and can compress code down to its structure with Tree-sitter. You then upload that one file to ChatGPT, Claude, Gemini or any other model. This guide walks through the workflow end to end. Step 1: pack your repo You do not even need to install it. In your project folder: npx repomix@latest If you will use it often, install it once. On a Mac, Homebrew works: brew install repomix npm (npm install -g repomix), yarn and bun work too. Then just run: repomix Repomix writes repomix-output.xml to the current folder: a file summary, the directory structure and every file's contents. The terminal output shows the token count and your five largest files, so you know before uploading whether it will fit. Step 2: hand it to ChatGPT Upload repomix-output.xml to a ChatGPT chat (or paste it if it is small) and start with a prompt like the one the README suggests: This file contains all the files in the repository combined into one. I want to refactor the code, so please review it first. Want to skip the upload? --copy puts the output straight on your clipboard: repomix --copy Step 3: keep it inside the context window Big repos are where most people give up. Repomix has four levers for this. Pack only what matters. Use glob patterns to include or ignore paths: repomix --include "src/**/*.ts,**/*.md" repomix --ignore "**/*.log,tmp/" Compress. --compress uses Tree-sitter to keep the essential structure (classes, functions, interfaces) and drop the rest. Great for "explain the architecture" questions where the model does not need every line: repomix --compress Strip noise. --remove-comments and --remove-empty-lines cut tokens further. Split it. If a pack is still too big for one upload, --split-output writes numbered files of a size you choose: repomix --split-output 2mb To find the heavy folders first, --token-count-tree prints your file tree with token counts. The default counting uses the o200k_base tokenizer, which the README lists as GPT-4o's, and you can change it with --token-count-encoding. Step 4: pick a format the model reads well XML is the default. You can also ask for Markdown, JSON or plain text: repomix --style markdown Markdown is easy to skim yourself before you upload. JSON is useful if you want to query the pack with jq. Pack a GitHub repo you have not cloned This is the underrated feature. Want ChatGPT to explain a library before you adopt it? Point Repomix at the repo: repomix --remote yamadashy/repomix It accepts full GitHub URLs, the user/repo shorthand, and a branch, tag or commit with --remote-branch. Combine it with compression for a quick architecture overview: repomix --remote yamadashy/repomix --compress Then use the README's library overview prompt: This file contains the entire codebase of library. Please provide a comprehensive overview of the library, including its main purpose, key features, and overall architecture. For security, Repomix does not load config files from remote repos unless you pass --remote-trust-config. Give the model history, not just code When you are asking "why did this break?", recent changes matter more than the full source. Repomix can add git context: repomix --include-logs --include-diffs --include-logs adds the last 50 commits by default, with dates, messages and changed files. --include-diffs adds your working tree and staged changes. Prompts worth saving The README ships a few prompts you can paste after uploading: - Code review: "This file contains my entire codebase. Please review the overall structure and suggest any improvements or refactoring opportunities, focusing on maintainability and scalability." - Docs: "Based on the codebase in this file, please generate a detailed README.md that includes an overview of the project, its main features, setup instructions, and usage examples." - Tests: "Analyze the code in this file and suggest a comprehensive set of unit tests for the main functions and classes. Include edge cases and potential error scenarios." Save your settings Once you find flags you like, write them to a config file so every run uses them: repomix --init That creates repomix.config.json in the project. Add --global to make it your default everywhere. No terminal? Three other ways in - Website: repomix.com packs a public repo in the browser, with format options and a token estimate. - Browser extension: Chrome and Firefox add-ons put a Repomix button on GitHub repo pages. - VS Code: the community extension Repomix Runner runs it on any folder from the editor. Going further We wrote up this workflow in a shorter form, with who it suits and where it falls down: Repomix tutorial: pack your codebase into one file for an LLM. The Repomix repo page on Deduped collects other ways people use it. If your real goal is to understand a codebase you just inherited rather than chat about it, compare this with how to map an unfamiliar codebase in one run with Graphify. Deduped tracks the AI repos people are actually building with and writes up how to use them. Repo: yamadashy/repomix. --- Get one email every Monday with the week's hottest AI repos and exactly how to use them: subscribe free at dedupedai.com
stop manually copying files into LLMs. repomix packs your repo into clean XML with file tree, line numbers and token counts. tree-sitter --compress cuts ~70% tokens on big repos. yamadashy/repomix what do you use to feed context to agents? #AIcoding t.co
Cut Claude Code & Cursor Token Bills by 72% (4 Viral GitHub Tools Tested) #shorts Why does Claude Code or Cursor burn 100,000 tokens ($4.20) in 10 minutes? We tested the 4 most viral open-source GitHub token-saving tools (402,000+ combined stars) head-to-head — and uncovered why using Caveman solo only saves 8.5% in JetBrains' empirical benchmark, while stacking 3 orthogonal layers cuts your real bill by 72%! 🔥 The 4 Open-Source Token Savers Tested: 1. RTK (rtk-ai/rtk · 84.4k★): Single Rust CLI proxy that compresses noisy git, pytest, and build logs by 60–90% before they hit your LLM context. → GitHub: rtk-ai/rtk 2. Caveman (JuliusBrussee/caveman · 114.6k★): Telegraphic output skill ("why use many token when few do trick") that cuts chat prose by 65%. → GitHub: JuliusBrussee/caveman 3. Ponytail (DietrichGebert/ponytail · 174.8k★): Lazy senior dev YAGNI skill that stops AI over-engineering and cuts code bloat by 80%. → GitHub: DietrichGebert/ponytail 4. Repomix (yamadashy/repomix · 29.0k★): Tree-sitter AST repo map that slashes blind file reads by 70% on large monorepos. → GitHub: yamadashy/repomix ⚡ The 3-Layer Golden Stack (-72% Real Session Cost): • Input Layer: brew install rtk && rtk init -g • Code Layer: npx skills add DietrichGebert/ponytail • Output Layer: npx skills add JuliusBrussee/caveman #ClaudeCode #CursorAI #GitHub #OpenSource #AICoding #CodingTools #SoftwareEngineering #LLM #DevTools #Shorts
Your repo is too big for a single AI prompt? 📦 Repomix packs source files into an AI-friendly output and lets you exclude what you don't need. Handy before an LLM code review. 🤖 yamadashy/repomix #DevTools #浇朋友 #蓝朋友 #互关 #fo
代码库太大，AI 一次读不完？📦 Repomix 将项目文件打包成适合模型阅读的文本，还能排除无关文件。做 AI 代码审查前先整理上下文。🤖 yamadashy/repomix #DevTools #浇朋友 #蓝朋友 #互关 #fo
Repomix: mete TODO tu repositorio en un solo archivo para la IA Repomix empaqueta un repositorio entero en un solo archivo para pasárselo a una IA, con recuento de tokens por archivo y revisión de secretos. Probado en una API de tareas de ejemplo. Repositorio: yamadashy/repomix yamadashy/repomix Licencia: MIT Estrellas: 28.876 (datos de GitHub consultados el 17/10/2026 19:50 (hora de Madrid)) Las capturas son de mi propia terminal, con un proyecto de prueba. Este vídeo no está patrocinado ni afiliado al proyecto. Curso gratis «Programar con IA desde cero» en el canal Control Z. #Shorts #GitHub #Programación #HerramientasParaProgramadores