coco — CLI Coding Agent

A command-line coding agent powered by LangChain, LangGraph, and Anthropic Claude.

Features

• Multi-agent system: specialized agents for code, planning, Q&A, and web search
• Auto-routing: automatically selects the best agent and model for each task
• Codebase indexing: semantic search over your codebase via Chroma + embeddings
• Context compression: auto-summarizes long conversations to save tokens
• Human-in-the-loop: asks for clarification when needed
• LLM-agnostic: built on LangChain abstractions, swap providers easily
• Persistent caching: SQLite cache for LLM responses
• Session history: full JSONL logs of all interactions

Installation

pip install coco

Or from source:

git clone <repo>
cd code-assistant
pip install -e .

Setup

1. Copy .env.example to .env and add your Anthropic API key:

   cp .env.example .env
   # Edit .env and set ANTHROPIC_API_KEY=...

2. Run coco in any directory:

   coco

3. (Optional) Run /setup inside coco to configure per-project settings.

Usage

coco> /help              # Show all commands
coco> /index             # Index current codebase for semantic search
coco> /setup             # Configure models and settings
coco> /status            # Show current config and index stats

# Direct agent commands
coco> /ask how does this codebase work?
coco> /code add a type hint to all functions in utils.py
coco> /plan build a REST API with authentication
coco> /search latest LangGraph documentation

# Auto-routed (coco picks the best agent)
coco> What does the authenticate() function do?
coco> Fix the bug in my parser
coco> /clear             # Reset conversation
coco> /exit              # Exit

Configuration

coco stores project config in a .coco file (TOML) in the current directory. Global settings live in ~/.coco/config.toml.

[model]
complex_model = "claude-sonnet-4-6"
simple_model = "claude-haiku-4-5-20251001"
temperature = 0.0
max_tokens = 4096

[index]
chunk_size = 1000
chunk_overlap = 200

[safety]
confirm_file_writes = true
confirm_shell_commands = true

Architecture

coco/
├── config/      # TOML config management
├── memory/      # Context compression + JSONL history
├── tools/       # LangChain @tool definitions
├── indexer/     # Chroma codebase indexing
├── graph/       # LangGraph StateGraph workflow
├── agents/      # Specialized agent implementations
└── cli/         # Rich terminal UI + prompt_toolkit REPL

Security

• API keys are read from environment variables only, never written to config files
• File writes require confirmation (configurable)
• Shell commands use argument lists (no shell injection)
• Add .env to your .gitignore