Say "add a calculator function" in Discord. Get a reviewed PR in 60 seconds.
You: "Add a calculator with add, subtract, multiply, divide"
VoiceCode: Creates PR #9 with calculator.py → CodeRabbit reviews → Ready to merge
- You spoke into Discord
- AI understood your intent and wrote production code
- Code ran in an isolated Daytona sandbox
- A real PR was created on GitHub
- CodeRabbit reviewed it and found 2 issues
- AI fixed the issues automatically
- PR approved and ready to merge
Total time: 47 seconds.
Every developer has been in this meeting:
"Can someone add a simple validation function?" "I'll create a ticket." "It's just 20 lines of code..." "Fine, I'll do it after standup."
Three days later, the ticket is still in the backlog.
VoiceCode eliminates the gap between "someone should build this" and "it's deployed."
┌─────────────────────────────────────────────────────────────────────┐
│ VOICECODE PIPELINE │
├─────────────────────────────────────────────────────────────────────┤
│ │
│ Discord Coding Agent GitHub Review │
│ ┌─────┐ ┌─────────┐ ┌──────┐ ┌────────┐ │
│ │ Voice│ ───► │ Daytona │ ───► │ PR │ ───► │CodeRab-│ │
│ │ /Text│ │ Sandbox │ │ │ │ bit │ │
│ └─────┘ └─────────┘ └──────┘ └────────┘ │
│ │ │ │ │ │
│ │ │ │ │ │
│ ▼ ▼ ▼ ▼ │
│ "add a Writes code Creates PR Finds 2 │
│ calculator" in isolation with changes issues │
│ │ │
│ ┌─────────────────────────────────┘ │
│ │ │
│ ▼ │
│ Agent fixes issues │
│ automatically until │
│ review passes │
│ │
└─────────────────────────────────────────────────────────────────────┘
Most AI coding tools generate code and hope it works. VoiceCode actually runs the code:
async with Sandbox(config) as sandbox:
# AI writes code
await sandbox.write_file("calculator.py", generated_code)
# Actually execute it to verify it works
result = await sandbox.run("python calculator.py")
# Only create PR if code actually runs
if result.success:
create_pull_request()The code runs in a Daytona sandbox—an isolated container that:
- Has no access to your production environment
- Can be destroyed instantly if something goes wrong
- Lets AI experiment freely without risk
# Required API keys
DISCORD_TOKEN=your_discord_bot_token
GITHUB_TOKEN=your_github_pat
DAYTONA_API_KEY=your_daytona_key
NOVITA_API_KEY=your_novita_key # or OPENAI_API_KEY
# Optional (for voice)
ELEVENLABS_API_KEY=your_elevenlabs_keygit clone https://github.com/lil-Zlang/VoiceCode.git
cd Daytona
pip install -r requirements.txt
python -m src.voice_bot- Invite the bot to your Discord server
- Join a voice channel
- Type
/jointo connect the bot - Say or type:
!voicecode add a hello world function - Watch the magic happen
| Agent | Responsibility | Tech |
|---|---|---|
| Discord Bot | Voice/text capture, user interaction | discord.py, ElevenLabs STT/TTS |
| Coding Agent | Understand task, write code, fix issues | DeepSeek/GPT-4, tool-use loop |
| Orchestrator | Coordinate flow, manage state, handle retries | async Python, state machine |
while iterations < MAX_ITERATIONS:
# 1. Ask LLM what to do next
response = llm.chat(messages, tools=AVAILABLE_TOOLS)
# 2. Execute tool calls
for tool_call in response.tool_calls:
if tool_call.name == "write_file":
await sandbox.write_file(tool_call.args)
elif tool_call.name == "run_command":
await sandbox.run(tool_call.args)
elif tool_call.name == "read_file":
content = await sandbox.read_file(tool_call.args)
# 3. Check if done
if response.finish_reason == "stop":
breakThe agent can:
read_file- Read any file in the repowrite_file- Create or modify filesrun_command- Execute shell commandslist_files- Explore directory structuresearch_code- Find relevant code patterns
When CodeRabbit finds issues:
PR Created → CodeRabbit Reviews → Issues Found?
│
┌─────────────────┴─────────────────┐
│ Yes │ No
▼ ▼
Agent reads feedback PR Approved
Agent fixes issues Ready to merge
Push new commit
Loop back to review
Input: "add a calculator function with add subtract multiply divide"
Output: PR #9
# calculator.py (generated)
def add(a, b):
"""Add two numbers."""
return a + b
def subtract(a, b):
"""Subtract b from a."""
return a - b
def multiply(a, b):
"""Multiply two numbers."""
return a * b
def divide(a, b):
"""Divide a by b with zero-division handling."""
if b == 0:
raise ValueError("Cannot divide by zero")
return a / bInput: "create a PR and adding a more complex js version for it"
Output: PR #10 - 237 lines of JavaScript
- Stop context-switching for small tasks
- Keep focus on complex problems
- Let AI handle boilerplate
- Meeting action items become PRs instantly
- No more "I'll create a ticket for that"
- Junior devs can contribute faster
- Ship features during the meeting that requested them
- Reduce time-to-PR from days to seconds
- AI handles code, humans handle decisions
- Discord voice/text integration
- Sandboxed code execution (Daytona)
- Automated PR creation
- CodeRabbit review integration
- Auto-fix review feedback
- Slack integration
- VS Code extension
- Multi-file refactoring
- Test generation
- Deploy preview environments
| Component | Technology |
|---|---|
| Voice Bot | Python, discord.py, discord-ext-voice-recv |
| Speech-to-Text | ElevenLabs |
| Text-to-Speech | ElevenLabs |
| LLM | DeepSeek v3.1 via Novita AI (or OpenAI) |
| Sandbox | Daytona SDK |
| Version Control | PyGithub |
| Code Review | CodeRabbit |
| Orchestration | asyncio, custom state machine |
Q: Is the generated code any good?
A: It's reviewed by CodeRabbit before you see it. If there are issues, the agent fixes them automatically. You only see PRs that pass review.
Q: What if the AI writes something dangerous?
A: Code runs in isolated Daytona containers with no access to your production environment. The sandbox is destroyed after each task.
Q: Can it modify existing code?
A: Yes. The agent can read your entire repo, understand the codebase, and make targeted modifications while respecting existing patterns.
Q: How much does it cost?
A: The LLM calls cost ~$0.01-0.10 per task depending on complexity. Daytona has a free tier. CodeRabbit has a free tier for open source.
Q: Why Discord?
A: Teams already use Discord for communication. VoiceCode meets developers where they are. Slack integration is on the roadmap.
PRs welcome! Areas we need help:
- Slack integration - Port the Discord bot to Slack
- Better prompts - Improve code generation quality
- More tools - Add capabilities like test generation
- Documentation - Tutorials, examples, guides
MIT
Built with:
- Daytona - Sandboxed development environments
- CodeRabbit - AI code review
- Novita AI - LLM API
- ElevenLabs - Voice AI
Stop writing tickets. Start shipping code.
Star this repo if you think meetings should produce PRs, not tickets.