Version control for AI conversations. Save, rewind, branch, and merge your chats like code, with voice.
- Every conversation turn becomes a Git commit
- Branch conversations to explore "what if" scenarios
- Rewind to any point in a conversation
- Fork, merge, and diff conversation threads
- Push conversation history to GitHub
- Talk to it -- voice-first interface with a floating orb UI
+-----------------------------------------------------------+
| BROWSER |
| |
| VoiceOrb ThreadSidebar GitGraph ChatPanel |
| (mic/speaker) (branches) (commits) (text) |
| | | | | |
| +--------+-------+-------+-------+-----------+ |
| | | |
| voiceSocket.js api.js |
| (WebSocket) (REST) |
+--------------|----------------|---------------------------+
| |
wss://ws/voice https://api/*
| |
+--------------|----------------|---------------------------+
| v v FASTAPI SERVER |
| +-------------------+ +------------------+ |
| | Voice Pipeline | | REST Handlers | |
| | | | | |
| | WebM -> WAV | | /api/chat | |
| | Pulse STT | | /api/checkpoint | |
| | Sentence Buffer | | /api/fork | |
| | Waves TTS | | /api/merge | |
| +--------+----------+ | /api/push ... | |
| | +--------+---------+ |
| | | |
| +-----------+-----------+ |
| | |
| v |
| +-----------------------+ |
| | LangGraph Supervisor | |
| | (Claude Sonnet) | |
| +-----------+-----------+ |
| | |
| +-------------+-------------+ |
| | | | |
| v v v |
| conversation git_ops github_ops |
| _agent _agent _agent |
| | | | |
| | +----+----+ +----+----+ |
| | | Tools | | Tools | |
| | +---------+ +---------+ |
| | checkpoint push |
| | rewind gist |
| | fork issues |
| | merge |
| | diff |
| | | | |
| +-------------+-------------+ |
| | |
| +----------+----------+ |
| | | |
| v v |
| +------------------+ +-------------------+ |
| | GitCheckpointer | | GitHub API | |
| | (.conversations/ | | (PyGithub) | |
| | local git repo)| +-------------------+ |
| +------------------+ |
| thread = branch |
| checkpoint = commit |
| state = JSON in tree |
+-----------------------------------------------------------+
+-------+
| START |
+---+---+
|
v
+--------+--------+
| SUPERVISOR |
| (Claude LLM) |
| |
| - classifies |
| user intent |
| - picks agent |
| or FINISH |
+---+----+----+--+
| | |
+-------------+ | +-------------+
| | |
v v v
+---------+------+ +--------+-------+ +--------+--------+
| conversation | | git_ops | | github_ops |
| _agent | | _agent | | _agent |
| | | | | |
| General chat, | | Git tools: | | GitHub tools: |
| planning, | | checkpoint | | push |
| brainstorming, | | time_travel | | share_as_gist |
| questions | | fork | | create_issue |
| | | merge | | create_pr |
| | | diff | | |
| | | list_branches | | |
| | | conv_log | | |
+-------+--------+ +-------+--------+ +--------+--------+
| | |
+-------------------+-------------------+
|
v
+--------+--------+
| SUPERVISOR |
| (routes again |
| or FINISH) |
+--------+--------+
|
v
+---------+---------+
| MAYBE_SUMMARIZE |
| |
| if messages > 20: |
| compress older |
| into summary |
+---------+---------+
|
v
+---+---+
| END |
+-------+
MIC SPEAKER
| ^
v |
[MediaRecorder] [Web Audio API]
| (WebM/Opus, 250ms chunks) ^ (decode + play)
v |
[WebSocket binary frames] [base64 audio_chunk]
| ^
v |
[audio_converter] [Smallest.ai Waves]
| WebM -> WAV (ffmpeg) ^ TTS per sentence
v |
[Pulse STT] [sentence_buffer]
| WAV -> transcript ^ tokens -> sentences
v |
[Supervisor] -----> [Agent] -----> [stream response tokens]
- User speaks or types a message
- Supervisor LLM classifies intent and routes to the appropriate agent
- Agent executes tools (git operations, GitHub API calls, or plain conversation)
- Response streams back as text tokens, buffered into sentences for TTS
- Audio chunks stream to the browser for playback
- After playback finishes, mic auto-resumes for the next turn
Routing rules:
conversation_agent-- general chat, planning, brainstorminggit_ops_agent-- save, checkpoint, rewind, fork, merge, diff, time travelgithub_ops_agent-- push, share, issues, pull requests
| Method | Path | Description |
|---|---|---|
| POST | /api/chat |
Send a message, get agent response |
| POST | /api/checkpoint |
Create a named checkpoint |
| POST | /api/time-travel |
Rewind to a specific checkpoint |
| POST | /api/fork |
Fork a conversation at a checkpoint |
| POST | /api/merge |
Merge two conversation branches |
| GET | /api/threads |
List all conversation threads |
| GET | /api/threads/{id}/log |
Get conversation log for a thread |
| GET | /api/threads/{id}/diff/{a}/{b} |
Diff between two checkpoints |
| POST | /api/github/push |
Push a thread to GitHub |
| GET | /health |
Health check |
| Path | Description |
|---|---|
/ws/voice |
Bidirectional voice -- audio in, audio + text out |
/ws/chat |
Text-based streaming chat |
gitcheckpoint/
├── main.py # Entry point (uvicorn)
├── Dockerfile
├── pyproject.toml
├── requirements.txt
│
├── src/
│ ├── config.py # Pydantic settings (.env)
│ │
│ ├── checkpointer/
│ │ └── git_checkpointer.py # Core: Git-backed LangGraph checkpointer
│ │
│ ├── graph/
│ │ ├── state.py # ConversationState type
│ │ └── supervisor.py # Supervisor graph + routing logic
│ │
│ ├── agents/
│ │ ├── conversation_agent.py # General conversation
│ │ ├── git_ops_agent.py # Git operations
│ │ └── github_ops_agent.py # GitHub operations
│ │
│ ├── tools/
│ │ ├── git_tools.py # @tool: checkpoint, rewind, fork, merge, diff
│ │ ├── github_tools.py # @tool: push, gist
│ │ ├── github_helpers.py # GitHub API utilities
│ │ └── memory_tools.py # Cross-session memory
│ │
│ ├── voice/
│ │ ├── tts_service.py # Smallest.ai Waves TTS
│ │ ├── audio_converter.py # WebM to WAV conversion
│ │ ├── sentence_buffer.py # Sentence-level TTS streaming
│ │ ├── command_parser.py # Voice intent parsing
│ │ ├── atoms_agent.py # Smallest.ai Atoms config
│ │ └── session_manager.py # Voice session management
│ │
│ └── api/
│ └── server.py # FastAPI app, all routes, WS handlers
│
├── frontend/
│ ├── index.html
│ ├── vite.config.js
│ └── src/
│ ├── App.jsx # Main layout
│ ├── components/
│ │ ├── VoiceOrb.jsx # Floating animated orb (voice UI)
│ │ ├── ChatPanel.jsx # Text chat interface
│ │ ├── ThreadSidebar.jsx # Thread list sidebar
│ │ ├── GitGraph.jsx # Commit graph visualization
│ │ ├── ControlsBar.jsx # Action buttons
│ │ └── DiffViewer.jsx # Checkpoint diff display
│ └── lib/
│ ├── api.js # REST client
│ ├── voiceSocket.js # Voice WebSocket client + audio playback
│ └── websocket.js # Chat WebSocket client
│
└── tests/
├── test_git_checkpointer.py
├── test_supervisor.py
├── test_git_tools.py
├── test_github_tools.py
├── test_voice.py
└── test_api.py
# Backend
cp .env.example .env # Fill in API keys
pip install -e ".[voice,dev]"
python main.py
# Frontend
cd frontend
npm install
npm run dev| Variable | Required | Description |
|---|---|---|
ANTHROPIC_API_KEY |
Yes | Claude API key |
SMALLEST_API_KEY |
Yes | Smallest.ai API key |
GITHUB_TOKEN |
No | GitHub fine-grained token (for push/gist) |
GITHUB_OWNER |
No | GitHub username |
GITHUB_CONVERSATIONS_REPO |
No | Repo name for pushed conversations |
VITE_API_URL |
No | Backend URL for frontend (production) |
- Backend: Python, FastAPI, LangGraph, LangChain, GitPython, PyGithub
- LLM: Claude Sonnet (via langchain-anthropic)
- Voice: Smallest.ai Waves (TTS) + Atoms (voice agent)
- Frontend: React, Vite, Framer Motion, Web Audio API
- Storage: Git repository (conversations), MemorySaver / Postgres / Redis (state)
python -m pytest tests/ -vNoah Kostesku. All Rights Reserved.