Extract, version, move and reuse the code, documents and files your AI chats produce.
What an AI chat produces stays inside the chat. You copy a code block, paste it into an editor, guess the extension. If you want an earlier version of a document there is no way back. If a tool returned 200 rows, the interface shows you ten.
Offcut turns these into files — on Claude, ChatGPT, Gemini, Perplexity and the other chat interfaces in the registry.
| Kind | Note |
|---|---|
| Code blocks | Correct extension, name derived from the code itself (backfill.py) |
| Documents / canvases | Every version separately, with the line difference between them (+12 −3) |
| Tool / MCP outputs | The whole result rather than the truncated view, with the call parameters |
| Citations | Source list as its own file where the chat searched the web |
| Generated images | Named from the prompt that produced them |
| Attachments | What you uploaded and what was produced, in separate groups |
| The conversation | As Markdown; optionally continued on another provider |
You can drag a file straight into your editor, save to a fixed folder, or take a selection as one zip.
Anything an MCP server produces — Linear, Sentry, GA4, FindAgent agents, one you wrote yourself — lands inside the conversation, which makes it an ordinary item to Offcut. You can download an agent's output in full, with the call parameters that produced it. No integration required.
Not in the store yet. Development install:
chrome://extensions → Developer mode → Load unpacked → this folder
No external requests. No telemetry. No account. Everything it reads stays in your browser. Nothing is stored beyond your settings, and download history is off by default.
This is not a promise but a constraint verified on every release: CI checks that the source contains no address outside the permitted chat sites and refuses to build the package otherwise. → store/privacy.en.md
node selftest.js # pure-function tests + adapter conformance suite
node tools/check-spec.mjs # specification consistency gates
node tools/pack.mjs # store package
No dependencies, no build step, no bundler.
| Design document | Decisions and their reasoning. Where to start is at the top |
| Adding a provider | Usually one line |
| Known limitations | What it cannot do, and why |
| Breakage runbook | Symptom → diagnosis → fix |
| Provider discovery | Step one: replacing assumptions with measurements |
| Design screens | Visual evidence for the decisions |
MIT — Ömer Faruk Ceylandağ (Kata Technology)
Not affiliated with Anthropic, OpenAI, Google or Perplexity; their names are used only to identify the supported services.
Report an issue: team@katatechnology.co