docs

dex documentation

dex is a terminal coding agent: it talks to OpenAI-compatible Chat Completions / Responses APIs and Anthropic's native Messages wire, calls tools (read, bash, write, edit, grep, find, ls, plus MCP servers) to operate on your local files, and offers an interactive TUI, one-shot prompt mode, and a raw JSON tool mode. All agent work can run in a daemon over HTTP+SSE, and conversations persist as resumable JSONL sessions.

Guides

Getting started Install dex, configure a provider, make your first edit, and learn the TUI basics. Configuration One config file, one precedence order, one selection knob: model: provider/model. Client–daemon mode Headless daemons, remote TUIs, bearer tokens, reconnects, and remote approvals. Sessions Crash-safe JSONL journals, resume and reattach, compaction, steering, undo. Skills SKILL.md frontmatter, discovery directories, and on-demand /skill: loading. Tools Built-in tools, permission modes, then_run, and sub-agents. MCP servers stdio, HTTP, and SSE servers; OAuth login; tool naming and schema caps. Lua extensions Sandboxed extensions: manifest.yaml, the dex.* API, hooks, commands, and consent. Composability How skills, MCP, extensions, sub-agents, and the CLI snap together — with a working pipeline. Experiments Opt-in and kill-switch features: verify hook, audit log, durable sessions, compaction modes, cache warming. Environment variables Every DEX_* knob and related provider variable, with defaults.

Where to next

One-page overview: the README in the dex repo.
Bugs and feature requests: the issue tracker.
Design/development notes live in the repo's docs/ directory (mcp-plan.md, subagent-plan.md, …) — those are development artifacts, not user documentation.