elmkit 0.1.0

No project description provided

elmkit 🌳 Project Plan

Philosophy

• No frameworks, just functions
• Build only what we actually use
• Under 1000 lines total
• Ship in 3 days max

Core Components to Build

1. Client (client.py)

• Basic chat completion wrapper
• String or messages input support
• Sync + async methods
• Streaming support
• Built-in retry for rate limits
• Token counting
• Cost calculation
• Support OpenAI-compatible providers (DeepSeek, Groq)

2. Messages (messages.py)

• Simple Message dataclass
• .to_dict() for API compatibility
• Optional: ToolMessage variant
• Keep under 50 lines

3. Tools (tools.py)

• Function to OpenAI tool schema converter
• Tool execution helper
• Tool result formatting
• Simple decorator if helpful

4. Tracking (tracking.py)

• Simple call tracker/logger
• Track: latency, tokens, cost, errors
• Optional decorator for function tracking
• JSON structured logging

5. Utils (utils.py)

• Token counting helpers
• Cost calculation tables
• Any shared utilities

6. Package Setup

• __init__.py with clean exports
• pyproject.toml (modern Python packaging)
• MIT License
• .gitignore

Documentation

README.md Structure

• One-line description: "No-framework LLM utilities"
• Installation: pip install elmkit
• 3-4 real usage examples
• "No roadmap, no promises" disclaimer
• Anti-framework philosophy statement

Examples Folder

• simple_chat.py - Basic usage
• tool_calling.py - Function calling example
• streaming.py - Streaming responses

What NOT to Build

• ❌ Registry systems
• ❌ Graph/node abstractions
• ❌ Memory management
• ❌ Chain/pipeline classes
• ❌ Callbacks
• ❌ Custom executors
• ❌ CLI tools
• ❌ Provider abstractions (just OpenAI-compatible)

Development Process

Phase 1: Core Development

1. Start with client.py and messages.py
2. Test with real project immediately
3. Add other components as needed
4. Keep testing with actual use cases

Phase 2: Packaging

1. Create proper package structure
2. Write README with examples
3. Add pyproject.toml
4. Test local install

Phase 3: Publishing

1. Create GitHub repo
2. Push code with clear README
3. Optional: Publish to PyPI
4. Optional: Write blog/social post

Quality Checks

• Every file under 200 lines
• Total codebase under 1000 lines
• Can explain any function in one sentence
• No abstract base classes unless absolutely needed
• API feels natural: client.chat("Hello")

Success Metrics

• ✅ Can replace LangChain in our projects
• ✅ Debugging is trivial
• ✅ New developer understands in 5 minutes
• ✅ We actually want to use it

Marketing (Optional)

• GitHub repo with clear anti-framework stance
• X/Twitter post: "Replaced 1000 lines of LangChain..."
• LinkedIn post: Professional version
• Submit to HN: "Show HN: elmkit - LLM tools without the framework"

Remember

• We expect 0 users (besides ourselves)
• No maintenance obligations
• Just solving our own problem
• If it helps others, bonus

Next Steps

1. Create folder structure
2. Start coding client.py
3. Test immediately in real project
4. Iterate based on actual usage
5. Ship when it works for us

---

"The best code is boring code. The best framework is no framework." 🌳