Analysis updated 2026-05-18
Cut down how many tokens a coding agent spends reading large files.
Automatically configure Claude Code, Cursor, or Codex to use compressed file reads.
Retrieve a specific elided code range from a compressed file when needed.
Verify that code shown to an agent exactly matches what is on disk before editing.
| tokz-dev/compress | 0xkinno/neuralvault | 0xmayurrr/ai-contractauditor | |
|---|---|---|---|
| Stars | 1 | 1 | 1 |
| Language | TypeScript | TypeScript | TypeScript |
| Setup difficulty | easy | hard | easy |
| Complexity | 2/5 | 4/5 | 2/5 |
| Audience | developer | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Codex CLI currently hides stdio MCP tools from the model, so Codex may not use it automatically yet.
@tokz/compress is a local tool that helps AI coding agents avoid wasting effort reading entire files when they only need a few specific functions from them. It runs as an MCP server, a kind of local plugin that AI coding tools like Claude Code, Cursor, or Codex can call, and instead of handing the agent a whole file, it shows a compressed version containing just the parts relevant to what the agent is actually looking for. Everything else in the file is reduced down to a short signature showing that more code exists there, and it can be pulled back in full almost instantly if it turns out to matter after all. The key guarantee is that any code shown to the agent is an exact, byte-for-byte copy of what is really in the file on disk, including correct line numbers, never a summary or a paraphrased version. This is checked automatically every time, and any piece of code that fails that check is simply dropped rather than shown incorrectly. This matters because if an agent tries to edit code based on a paraphrased version that does not exactly match the real file, the edit will fail or, worse, the agent may reason about code that was never actually there. Setup is done through a single command line tool that automatically detects which AI coding tools are installed, such as Claude Code, Cursor, Antigravity, Windsurf, and Codex, and configures each one correctly for its own operating system and expected file paths. It handles some tricky details on its own, like using the correct command form on Windows so it does not fail to launch, and giving Claude Code a special plugin installation that also adds automatic hooks so compression happens without the agent needing to choose to use it. Every configuration change it makes backs up the original file first, and it will not touch a file that is not valid JSON. The project runs entirely on the user's own machine, requires no account, no API key, and no network connection, and is released under the MIT license.
A local MCP server that shows AI coding agents a byte-exact compressed slice of a file, relevant to their query, instead of the whole file.
Mainly TypeScript. The stack also includes TypeScript, Node.js, MCP.
MIT license: use it, modify it, and share it freely, including commercially.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.