opencode — Terminal AI Coding Agent

A CLI-based AI coding agent with a terminal UI (TUI) that can read/write files, run shell commands, search the web, and delegate tasks to sub-agents — all from your terminal.

Features

• Interactive TUI chat interface powered by Ink
• Supports multiple LLM providers: Gemini, OpenAI, and a local model
• Tool-calling loop: the LLM can invoke tools autonomously until a task is complete
• Streaming responses with real-time chunk display
• Conversation memory with automatic summarization when context grows too large
• Agent workspace at .agent_workspace/projects/ for all file operations

Tools Available to the Agent

Tool	Description
WebSearch	Searches Google via SerpAPI
BashTool	Executes shell commands
ReadFileTool	Reads a file from the workspace
WriteFileTool	Writes/creates a file (auto-creates directories)
UpdateFileContentTool	Replaces specific line ranges in a file
runAgentTool	Spawns a sub-agent with a given prompt

Prerequisites

• Bun or Node.js
• API keys for your chosen provider(s)

Installation

npm install
# or
bun install

Configuration

Set your provider and API keys using the CLI:

# Set provider (gemini | openai)
bun cli.ts providers set --provider gemini --model gemini-flash

# Log in with your API key
bun cli.ts providers login --provider gemini --key <your_api_key>

# For web search (optional)
# Add SERP_API_KEY to your .env file

Or edit store/data.json directly:

{
  "defaultProvider": { "name": "gemini", "model": "gemini-flash" },
  "setKeys": { "gemini": "<your_key>", "openai": "<your_key>" }
}

Usage

Interactive TUI Chat

bun cli.ts chat

Type your prompt, press Enter to send. Press Esc to exit.

One-shot Agent

bun cli.ts agent --prompt "Create a React todo app in the workspace"

List Available Models

bun cli.ts models

Provider Management

bun cli.ts providers list
bun cli.ts providers login --provider openai --key <key>
bun cli.ts providers logout --provider openai
bun cli.ts providers set --provider gemini --model gemini-flash

Project Structure

.
├── cli.ts                  # Entry point, registers all commands
├── actions/
│   ├── llm.ts              # LLM call logic (Gemini & OpenAI), tool loop, memory management
│   ├── tools.ts            # Tool definitions and implementations
│   └── memory.ts           # Message history and summary memory
├── commands/
│   ├── chat.tsx            # `chat` command — launches TUI
│   ├── agent.ts            # `agent` command — one-shot prompt
│   ├── models.ts           # `models` command — list models
│   └── providers/          # `providers` subcommands (login/logout/set/list)
├── tui/
│   └── TUIChat.tsx         # Ink-based interactive chat UI
├── store/
│   ├── data.json           # Persisted provider/key config
│   ├── db.ts               # Exports provider config
│   └── db1.ts              # Reads/writes data.json
└── .agent_workspace/
    └── projects/           # Agent's sandboxed file workspace

How It Works

1. User sends a message via TUI or CLI.
2. callLlm builds a system prompt with tool definitions and conversation history, then calls the active provider.
3. The LLM responds with either a { "tools": [...] } JSON to invoke tools, or a { "answer": "..." } to finish.
4. Tools are executed (in parallel where dependsOn allows), results are fed back to the LLM, and the loop continues until a final answer is produced.
5. Conversation summaries are compressed automatically when they exceed a token threshold.