Aftercode

Turn your daily AI coding sessions into personalized learning podcasts.

Aftercode connects to your coding-agent workflow (Claude Code, Codex, Cursor, …), figures out the deeper technical topics behind what you built and debugged that day, and generates a short two-speaker podcast episode — in Hebrew or English — that teaches you the concepts behind the code you shipped. Listen and browse them in a clean web playlist.

Turn vibe coding into real learning.

Open source (MIT OR Apache-2.0), CLI-first, self-hostable.

Demo

▶ Watch with sound (full-quality MP4)

How it works

aftercode episode
   │  auto-detect your coding-agent session (Claude Code / Codex / Cursor) →
   │    your prompts + what the agent did/explained
   │  + the real git diff (the Redis client, RabbitMQ config, new deps it added)
   │  scan for secrets, apply ignore rules, redact, cap size
   ▼
backend pipeline
   normalize → extract topics → detect knowledge gaps → rank →
   write two-speaker script (Hebrew or English) → TTS per segment
   (ElevenLabs or OpenAI) → concat + silence gaps → MP3 → store → episode
   ▼
web UI
   playlist · filter by topic · play through your episodes · transcript + quiz

Repository layout

crates/aftercode-core    shared types (session context, episode, audio)
crates/aftercode-server  Axum API backend + generation pipeline + serves the web UI
crates/aftercode-cli     the `aftercode` CLI
web/                     React + Vite web UI (playlist, filters, player)
migrations/              SQLite schema (auto-applied on startup)
docs/                    PRD, self-hosting, architecture, design specs

Install

You need: Docker (runs the backend + web UI) and
Rust (for the aftercode CLI — it must run natively to read your local
coding-agent session files). That's it — no database, no Node.

Copy-paste these four commands:

# 1. Get the code
git clone https://github.com/ron3899/aftercode && cd aftercode

# 2. Start the backend + web UI  →  http://localhost:8080
#    (first run builds the image once, ~5 min; after that it's instant)
docker compose up -d

# 3. Install the CLI
cargo install --path crates/aftercode-cli

# 4. Log in (opens your browser → click Approve)
aftercode login

Now open http://localhost:8080 to browse episodes, and run aftercode episode in any
project. Step 2 works with no API keys (mock mode) so you can try it immediately.

For real episodes, create a .env in the repo with your keys, then docker compose up -d
again:

LLM_PROVIDER=openai
OPENAI_API_KEY=sk-...
TTS_PROVIDER=openai        # voice via OpenAI tts-1; or set TTS_PROVIDER=elevenlabs + its keys

Port 8080 taken? Run PORT=9000 docker compose up -d and use that port everywhere.

Use it inside your coding agent

The CLI auto-detects your current session from whatever agent you used, so the simplest
workflow is: finish a chunk of work, then make the episode.

• In the terminal (any agent): aftercode episode --language en (or he). It reads the
  latest Cursor, Claude Code, or Codex session for that repo automatically.
• Ask the agent to do it for you. Since the agent has a terminal, just tell it:

  "Run aftercode episode --language en to turn what we just did into a podcast."

• Claude Code / Codex (CLI agents): run it right after your session in the same folder —
  detection finds the transcript. If you ran from a different folder, hand the transcript in
  directly: aftercode episode --transcript - (pipe a transcript) — never fails detection.
• Cursor / Windsurf (editor agents): open the editor's terminal in your project and run
  aftercode episode; it reads Cursor's live session for that workspace.

Run aftercode preview first to see which agent + session was detected and exactly what will
be sent.

Requires Rust + Node.

cd web && npm install && npm run build && cd ..   # build the UI (served at /)
cp .env.example .env                              # add keys, or leave mock to try
cargo run -p aftercode-server                     # first run prints a login token + URL
cargo install --path crates/aftercode-cli
aftercode login && cd your-project && aftercode episode --language en

Web UI

The backend serves a React app at / (built from web/dist):

• Playlist — your episodes with a persistent bottom player that plays through the
  current filtered list (prev/next = your queue).
• Filter by topic (chips) + language + search — instant.
• Episode detail — player, topics, key takeaways, quiz, full host/expert transcript.
• Responsive (mobile + desktop). Sign in reuses the CLI approve flow — no token paste.

Dev with hot reload: cd web && npm run dev (Vite on :5173, talks to the backend; set
VITE_API_BASE in web/.env). Rebuild for production with npm run build.

Auth

aftercode login (no arg) opens <backend>/cli/authorize; click Approve and the token
is captured automatically — for the CLI via a localhost loopback, for the web UI via a
browser redirect (no copy-paste). aftercode login <token> and seed-user remain for
scripts/CI; aftercode status validates the token against the backend.

Local-approval performs no identity check — it's for trusted localhost/self-host only.
A public/multi-user deployment needs a real OAuth provider (not yet built). This build is
effectively single-user: any signed-in user sees all episodes.

Reliable session capture

Episodes are built from your agent session + the git diff. Aftercode auto-detects the
session from Claude Code / Codex / Cursor on-disk storage (Cursor reads its live SQLite WAL),
but detection can miss (e.g. a run from a subfolder). Two safeguards:

• Hand the transcript in directly: your-transcript | aftercode episode --transcript -
  (accepts a Claude Code .jsonl, simple {"role","text"} JSONL, or plain text) — skips
  disk auto-detection.
• Thin-episode guardrail: if no session conversation is captured, episode refuses (so
  you never get an episode about a lone package-lock.json). Override with --allow-thin.
  Run aftercode preview first to see what was collected and which agent was detected.

Configuration

All secrets/config via env (see .env.example and docs/SELF_HOSTING.md):

• DATABASE_URL — SQLite, default sqlite://aftercode.db?mode=rwc (auto-created + migrated)
• LLM_PROVIDER — anthropic (default, claude-opus-4-8) / openai / mock; ANTHROPIC_API_KEY / OPENAI_API_KEY
• TTS_PROVIDER — elevenlabs (default) / openai / mock
  • ElevenLabs: ELEVENLABS_API_KEY, ELEVENLABS_HOST_VOICE_ID, ELEVENLABS_EXPERT_VOICE_ID
  • OpenAI: OPENAI_TTS_MODEL (default gpt-4o-mini-tts), OPENAI_TTS_VOICE_HOST, OPENAI_TTS_VOICE_EXPERT
• BLOB_STORE — localfs (default) / s3 (+ S3_BUCKET); audio served at /static/episodes/<id>.mp3
• APP_PUBLIC_URL, BIND_ADDR, WEB_DIST (path to the built web app)

LLM providers are pluggable behind a trait; topic + script stages use structured JSON output.
The script is written in the episode's language (Hebrew keeps tech terms in English, the way
Israeli devs speak).

Privacy

Aftercode sends your agent-session transcript (prompts + agent messages + the commands/edits
it ran) and the git diff hunks — never full source files, .env, secrets, or ignored
paths. A regex secret scanner redacts anything that looks like a key/token before upload,
payload size is capped, and aftercode preview shows exactly what would be sent.

Status

• Phase 1 (done): Rust CLI + API backend — session capture, topic/script generation,
  TTS, audio assembly, episode storage. Browser login.
• Phase 2 (done): React + Vite web UI — playlist, topic filter, player, episode detail.
• Phase 3 (planned): real OAuth/multi-user, auto-daily generation, download/share,
  private RSS, editor extensions. See docs/superpowers/specs.

License

Licensed under either of MIT (LICENSE-MIT) or Apache-2.0 (LICENSE-APACHE) at your option.