Chat
Interactive terminal chat that lets you converse with an LLM about your codebase while it invokes skene tools to gather information.
Prerequisites
- An API key configured (see Configuration) or a local LLM running
- A codebase to analyze
Basic usage
# Chat about the current directory
uvx skene chat
# Chat about a specific codebase
uvx skene chat /path/to/project
# Using the shorthand (defaults to chat)
uvx skene
The skene entry point defaults to the chat command, providing a convenient shorthand for interactive sessions.
Flags
| Flag | Short | Description | Default |
|---|---|---|---|
--api-key | API key for LLM provider | SKENE_API_KEY env var | |
--provider | -p | LLM provider | Config or openai |
--model | -m | LLM model name | Provider default |
--base-url | Base URL for OpenAI-compatible API endpoint. Required when provider is generic. | SKENE_BASE_URL env var | |
--max-steps | Maximum tool calls per user request | 4 | |
--tool-output-limit | Max tool output characters kept in context | 4000 | |
--debug | Log all LLM input/output to .skene/debug/ | Off |
How it works
The chat command starts an interactive terminal session where:
- You type a question or request about your codebase
- The LLM decides which skene tools to call (analyze, search, read files, etc.)
- Tool results are fed back to the LLM within the context window
- The LLM synthesizes a response based on the tool outputs
The --max-steps flag controls how many tool calls the LLM can make per request. Increase this for complex queries that require multiple analysis passes. The --tool-output-limit flag controls how much of each tool's output is kept in context to avoid exceeding token limits.
Tips for effective use
- Be specific — "What growth features does this codebase have?" works better than "Tell me about this code"
- Increase max-steps for deep analysis — Use
--max-steps 8when you want the LLM to do thorough multi-step analysis - Use debug mode to understand behavior —
--debuglogs all LLM interactions so you can see what tools are being called
Next steps
- Analyze command — Run a full codebase analysis
- LLM providers — Configure different providers for chat
- Configuration — Set defaults so you don't need flags every time