Overview

Elith is a universal repo-aware AI framework that gives any LLM (Claude, GPT, Gemini, local models) the same repository skills that IBM Bob has natively. It's not a chatbot wrapper or copilot plugin—it's a skill layer that makes any AI model repo-aware.

// What Elith Solves

LLMs don't know your codebase. Elith's context engine scans your repository and Obsidian vault, then builds a task packet with the 4-6 most relevant files. The model gets exactly what it needs—no more "I don't have access to that file."

Most AI tools give context-free code snippets. Elith reads your repository structure, understands your patterns, and generates code that fits your existing architecture.

Elith works with any LLM provider through a universal skill layer:

• Claude (Anthropic) — Native tool calling
• LM Studio — Local models with OpenAI-compatible API
• OpenRouter — Access multiple providers through one API

The skill layer normalizes tool calling across all providers, so you can switch models without changing your workflow.

// How It Works

Elith provides 4 core operations:

Analyze and explain your codebase architecture in plain language.

elith explain src/auth/

Propose architectural improvements specific to your codebase.

elith architect --problem "Add rate limiting to API"

Refactor code with focus on specific improvements.

elith refactor src/cache/ --focus "modernize redis client"

Generate comprehensive test suites for your code.

elith test-gen src/auth/login.ts

// The 12 Skills

Every operation has access to 12 repository skills:

• read_file — Read file contents with line numbers
• write_file — Write or overwrite files
• list_files — List directory contents recursively
• search_code — Regex search across all files
• git_diff — Show uncommitted changes
• git_commit — Commit changes with message
• run_tests — Execute test suite
• find_references — Find symbol references
• analyze_dependencies — Parse dependency tree
• explain_function — Extract and explain functions
• install_package — Add dependencies
• read_logs — Parse application logs

// Context Engine

The context engine solves the token limit problem:

Smart file selection based on:

• Key files (README.md, package.json, main.py, etc.)
• Entry points (main.py, index.ts, app.py, etc.)
• Operation-specific heuristics
• Files in target directory

// Supported Providers

| Provider | Status | Model Examples | |----------|--------|----------------| | Claude | [implemented] | claude-sonnet-4-20250514 | | LM Studio | [implemented] | Any local model | | OpenRouter | [implemented] | openai/gpt-4o, meta-llama/llama-3.1-8b |

// Session Logging

Every operation is logged to bob-reports/ for IBM Bob hackathon submission:

bob-reports/
  session_20260517_153549_f27343a8.md
  session_history.json

Each report contains:

• Full operation transcript
• All skill calls and results
• Files changed
• Timestamps

Note: Session data is stored in the sessions database. The bob-reports/ folder is specifically for hackathon judges to verify Bob usage.

// What's Next

• Quickstart — Install and run your first operation
• Installation — All install methods
• Configuration — Set up your LLM provider