Code Apex

CodeApex

A Claude Code-style coding agent CLI in your terminal, built in pure TypeScript on [LangChain Deep Agents](https://docs.langchain.com/oss/javascript/deepagents/overview): **planning/todo list, subagent delegation, real filesystem + shell, and web search**, driven from an interactive Ink TUI.

Prerequisites

• Node.js 22+
• A running Ollama daemon, signed in to Ollama cloud (the default model is deepseek-v4-flash:0731-cloud) — or any local Ollama model with solid tool-calling.

Run

npm install --legacy-peer-deps   # openwiki pins older ink/react; see dev notes below
npm start              # interactive Ink TUI, real FS + shell rooted at ./workspace
npm start -- --memfs   # in-memory virtual filesystem, nothing touches disk
npm start -- --yolo    # skip shell-command approval prompts (trusted tasks only)
npm run smoke          # cheap end-to-end check (todos + file write + shell)

**Shell approvals (on by default):** every `execute` call pauses the run before it touches your machine (details under the demo below). Disable with `--yolo`; `--memfs` has no shell at all.

Give it a task at the prompt, e.g.:

Research the top 3 TypeScript test runners, write a comparison to notes.md, then have the critic verify the file cites sources.

You'll watch it work like Claude Code in a **React Ink TUI**: a scrolling transcript (streaming answers with 🧠 dim thinking, 🔧 tool calls with results, 🤖 delegations indented per subagent), a live 📋 plan panel, a bordered `❯` input with command history (↑/↓), and a spinner + `esc to interrupt` while a task runs.

Thinking display requires a reasoning-capable model — the default `deepseek-v4-flash:0731-cloud` emits 🧠 reasoning deltas, as do `glm-5.2:cloud`, `kimi-k3:cloud`, `qwen3`, … Models without thinking support stream answers fine but emit no thinking tokens. Set `OLLAMA_THINK=0` if a model errors on think mode.

**Approvals, Claude Code-style:** every gated `execute`/`write_s