ChatterChat

An autonomous software engineer in your terminal, powered by the Claude API.

ChatterChat reads and writes real files, runs real commands, searches the web, and can reach into the apps you already use — Slack, Notion, GitHub, Linear, Sentry, Stripe and more — over the Model Context Protocol. It runs on Claude Opus 5 by default, or Claude Fable 5 for long autonomous runs.

pip install chatterchat

or with Homebrew:

brew install casper-white/tap/chatterchat

Setup

ChatterChat uses your own Claude Platform access. Either export an API key:

export ANTHROPIC_API_KEY=sk-ant-...

…or sign in with the Anthropic CLI, which stores a profile the SDK picks up automatically:

ant auth login

Use it

chatterchat                              # interactive session in the current directory
chatterchat "add tests for auth.py"      # run one task and exit
chatterchat -p "why is CI failing?" -y   # non-interactive, auto-approve tool calls
chatterchat --continue                   # resume the previous conversation
chatterchat -m fable -e max "port this service to async"

Inside a session:

/help                   show all commands
/model fable            switch model
/effort xhigh           low, medium, high, xhigh, max
/thinking off           hide Claude's reasoning
/tools                  what this session can do
/apps                   connectable apps and their status
/connect notion         connect an app
/cost                   token usage and estimated spend
/clear                  start fresh

What it can do

On your machine — read_file, write_file, edit_file, list_dir, glob, grep, run_command. Writes and shell commands ask for approval the first time; answer a to stop being asked for that tool, or start with -y to approve everything up front.

On the web — Claude's server-side web_search and web_fetch, so answers about current library versions and APIs come from the live web rather than memory.

In your apps — any remote MCP server. Built-in registry:

chatterchat apps                  # list them with connection status
chatterchat connect github        # prompts for a token
chatterchat connect slack --token-env SLACK_MCP_TOKEN
chatterchat connect mine --url https://mcp.example.com/mcp --token abc123

Most servers use OAuth. To get an access token for one, run npx @modelcontextprotocol/inspector, point it at the server URL, complete the Quick OAuth Flow and copy access_token. GitHub also accepts a personal access token. Servers that need no token at all (huggingface, context7, deepwiki, exa, cloudflare-docs) connect immediately.

Configuration

Settings live in ~/.chatterchat/config.json; tokens live in ~/.chatterchat/credentials.json (mode 600), or are read from an environment variable if you connected with --token-env.

chatterchat config                          # show everything
chatterchat config --set model=fable
chatterchat config --set effort=xhigh
chatterchat config --set auto_approve=true
chatterchat config --set max_tokens=32000
chatterchat models

Notes

• Conversations are saved to ~/.chatterchat/sessions/ and --continue resumes the latest one.
• Long sessions use server-side compaction, so they keep going past the context window.
• Ctrl-C interrupts the current turn without losing the conversation; Ctrl-D exits.
• Requests enable server-side refusal fallbacks, so a policy decline is retried on a fallback model inside the same call rather than failing the turn.

License

MIT