CodeBridge by taughtful.ai

CodeBridge is a free command-line tool by taughtful.ai that turns the last answer from a coding agent, Claude Code or OpenAI Codex, into a live voice lesson. You type ! codebridge inside your agent session, and the answer opens at taughtful.ai/codebridge as an interactive lesson: a voice tutor explains it out loud while a visual board renders diagrams, code blocks, and charts in sync with the speech. You can interrupt the tutor mid-sentence by voice, ask why, and the lesson rebuilds around your question. At the end, the tutor can write a ready-to-run prompt for your coding agent.

How do I install CodeBridge?

The CLI installs from npm and works in any terminal:

npm i -g @taughtful/codebridge

That puts the codebridge command on your PATH, which is what the ! codebridge bang inside a Claude Code or Codex session runs. To try it once without installing: npx @taughtful/codebridge.

CodeBridge also ships as a plugin for both agents, which adds a native command with your agent's own fuzzy file picker on @: /codebridge in Claude Code (via /plugin marketplace add taughtful-ai/codebridge then /plugin install codebridge@taughtful) and $codebridge in Codex (via codex plugin marketplace add taughtful-ai/codebridge then codex plugin add codebridge@taughtful). Restart the agent once after installing a plugin.

How does it decide what to teach?

CodeBridge reads your agent session's local transcript and selects the answer to your last real message. Output produced by CodeBridge runs themselves, such as the lesson URL a previous run printed, is never selected, so running it twice re-teaches the same real answer. Your prompts and tool output are never selected, only the assistant's answer text.

Can I include files in the lesson?

Yes. Tag them with @ paths: ! codebridge @src/auth.js @spec.pdf. Code and text files show a character meter as they resolve. PDFs, Office documents, ebooks, and images (photos of notes, whiteboards, screenshots) ride to the same ingestion pipeline taughtful.ai's Docs mode uses: structure detection for PDFs and vision transcription for images and scans. Each file becomes its own section of the lesson. Files that look like secrets (.env, keys, credentials) are refused on principle, and unsupported binaries are skipped.

What leaves my machine?

Only what you choose: the one answer being taught, plus any files you explicitly tag. It travels as an unlisted lesson and saves to your library when you are signed in. Everything else in the session, your prompts, tool output, other answers, stays local. The CLI contains no telemetry, and it is MIT-licensed with every line readable in the npm package.

What happens with very short answers?

CodeBridge declines to teach anything under 200 words when no files are tagged. Instead of minting a lesson from a two-word reply, it prints a line asking your agent to check what you actually want taught and to tag that content as a file. Nothing is uploaded and no model call is spent on a refused run.

Which agents does it work with?

Claude Code and OpenAI Codex, with the same commands. The bang lane (! codebridge) works in both TUIs; the session picker (codebridge browse) lists recent sessions from both agents side by side. CodeBridge is an independent tool by taughtful.ai and is not affiliated with or endorsed by Anthropic or OpenAI.

Frequently asked questions

Does my code get sent somewhere?

Only what you choose: the one answer you teach, plus files you tag. Your prompts, tool output, and the rest of the session never leave the machine. The CLI is MIT-licensed and every line ships readable in the npm package, if trust-but-verify is your love language.

What does it cost?

The command is free and burns zero model tokens. Lessons run on taughtful.ai: your first minutes are free, and plans exist for when one lesson turns into a habit.

Claude Code or Codex?

Both, same command. We are not making you pick a side in that particular rivalry.

Can it teach my PDFs, or a photo of my notes?

Yes. Tag them with @ and vision does the reading: whiteboards, screenshots, scanned pages. Each one becomes its own section of the lesson.

What happens if I run it on a two-word answer?

It politely declines. Under 200 words it suggests tagging the part actually worth teaching, because nobody needs a voice lecture on the word "done".

Can I interrupt the tutor mid-sentence?

That is the whole point. Barge in, ask why, and the lesson rebuilds around your question. It is the rare teacher that likes being interrupted.

How does it visualize so fast?

Secret sauce: a rendering engine we built for exactly this, tuned until the board keeps up with the voice. And we keep folding more into it: richer diagrams, charts, motion. The sauce thickens ;)

Published 2026-07-23 · Last updated 2026-08-08 · Written and maintained by the taughtful.ai team.