GitHub - tallesborges/zdx: An agentic CLI, built in rust.

Zdx is a coding agent built in Rust. I made it to have customization freedom and also for learnig purpose.

Disclaimer

• YOLO mode: the agent can access your files and tools. It can make destructive changes. Use at your own risk.
• Early-stage: expect breaking changes and unstable behavior while the project evolves.

Quickstart

Install

From source:

git clone https://github.com/tallesborges/zdx.git
cd zdx
cargo install --path crates/zdx-cli

Run

Exec mode

Telegram bot

First time? Run zdx config init and set your provider + telegram keys.

Features

Just the ones I use — but that turns out to be a lot.

Interfaces

• Interactive TUI — the main terminal experience (built with Ratatui)
• Exec mode — headless, scriptable one-shot runs (zdx exec -p "...")
• Telegram bot — drive the agent from your phone
• Monitor dashboard — a separate TUI to watch agent runs, services, logs, and config

Providers & models

• API providers — Anthropic, OpenAI, Google Gemini, xAI (Grok), DeepSeek, Moonshot (Kimi), MiniMax, Mistral, StepFun, Z.ai (GLM), Xiaomi MiMo
• Subscription / CLI auth — Claude CLI, Gemini CLI, OpenAI Codex, Google Antigravity, Xiaomi Token Plan
• Aggregators — OpenRouter, opencode-go
• Local — LM Studio
• Custom — any OpenAI-compatible provider defined straight from config
• OpenAI Responses transport (opt-in) + priority fast mode (/fast)
• Model registry with pricing, context limits, and reasoning detection

Built-in tools

• Filesystem & shell — read, write, edit, apply_patch, bash, glob, grep
• Web — web_search, fetch_webpage
• Agent — todo_write (task tracking), invoke_subagent, memory_search, memory_get, thread_search, read_thread

Subagents

Delegate scoped work to isolated child agents:

• task — general-purpose delegated execution (default)
• explorer — read-only, multi-round local/repo/thread/doc exploration (has read-only bash)
• oracle — deep-reasoning advisor for review, debugging, planning, and architecture
• thread-searcher — saved-conversation retrieval and multi-thread synthesis

Skills & commands

• Bundled skills — brainstorm, define-goal, deep-research, imagine (image gen/edit), memory, ship-first-plan, skill-creator
• Your own skills — project and user SKILL.md files, auto-discovered per scope
• Bundled commands — plan, execute-plan, investigate, review-loop, simplify, skillify, commandify, skillpack, self-improve, init-agents-md, onboarding
• Custom Markdown slash commands, fuzzy-matched in the command palette

Memory & threads

• qmd-backed memory — hybrid/vector/keyword search across notes, calendar, and past threads
• Threads — resume, rename, append, export, search, and tool-call inspection; open as tabs and forks

TUI niceties

• Tabbed threads/forks, image attachments via the Kitty graphics protocol, voice dictation
• Overlays for context usage, thread TLDR, and skills; model picker with Tab-cycled favorites
• Inline questions + follow-up suggestions, /prompt-builder, and a btw side chat

Telegram bot

• Interactive questions & follow-up buttons, per-project chat routing
• /model override, thinking controls, and file/document attachments

Automations & MCP

• Automations — scheduled, headless agent jobs (reports, reminders, actions)
• MCP — connect remote MCP servers (with OAuth), plus schema/call helpers

Context & config

• Ancestor-scoped AGENTS.md/CLAUDE.md and .zdx/skills discovery
• TOML config, provider login/logout, git worktree management, and thread-scoped artifact dirs

Inspiration

This project was inspired by several excellent tools in the agentic coding space:

• pi-mono — AI-powered coding agent with terminal UI and SDK for AI-assisted development
• codex — OpenAI's open-source terminal-based agentic coding assistant
• AMP — Great UX inspiration for agentic coding workflows
• opencode — Open-source AI coding agent with TUI, desktop app, and VS Code extension
• oh-my-pi

License

MIT