Eue

Eue is a self-evolving AI agent. It runs as a CLI conversational agent and a Telegram autonomous agent, with workspace isolation, strict command boundaries, and dynamic skill loading. Eue does not expose HTTP, does not require a database, and runs locally and deterministically.

Installation

mix escript.install hex eue

If you installed an older escript build, reinstall with --force:

mix escript.install hex eue --force

Before running

Set at least the LLM provider, model, and API token (in your shell or .env):

EUE_PROVIDER — LLM provider. Default openrouter.
EUE_AI_TOKEN — Required. API token for the chosen provider.
EUE_MODEL — Model id (e.g. openai/gpt-4o, z-ai/glm-5).

Example:

export EUE_PROVIDER=openrouter
export EUE_MODEL=openai/gpt-4o
export EUE_AI_TOKEN=sk-...

Running

CLI

eue

Workspace defaults to the current directory. To use another directory: eue --workspace /path/to/workspace.

Telegram

EUE_TELEGRAM_TOKEN=bot_token eue telegram

Or pass token/workspace via options:

eue telegram --token bot_token --workspace /path/to/workspace

Configuration

All parameters are environment variables. Set them in your shell or a .env file.

Variable	Default	Description
Provider & model
`EUE_PROVIDER`	`openrouter`	LLM provider (e.g. openrouter).
`EUE_AI_TOKEN`	—	Required. API token for the provider.
`EUE_MODEL`	provider default	Model id (e.g. `openai/gpt-5.2-chat`, `z-ai/glm-5`).
`EUE_FALLBACK_MODEL`	provider default	Fallback model when primary is unavailable.
Loop & limits
`EUE_MAX_STEPS`	`20`	Max agent loop steps per run (positive integer).
`EUE_HISTORY_MESSAGES`	`40`	Max history messages sent to the model.
`EUE_REQUEST_TIMEOUT_MS`	`60000`	CLI request timeout in milliseconds.
`EUE_EXECUTE_TIMEOUT_MS`	`30000`	Timeout for execute tool in milliseconds.
Workspace & paths
`EUE_WORKSPACE_PATH`	cwd (CLI) / unset	Workspace root (Telegram/CLI).
`EUE_WORKSPACE_ROOT`	—	Backward-compatible alias for `EUE_WORKSPACE_PATH`.
Execute
`EUE_EXECUTE_SHELL`	`$SHELL` or `/bin/sh`	Command runner executable for `,execute` tool.
Telegram
`EUE_TELEGRAM_TOKEN`	—	Required for Telegram. Bot token.
`EUE_TELEGRAM_ALLOW_CHATS`	—	Empty = all chats; non-empty = only these chat_id (use strict allowlist in production).
`EUE_TELEGRAM_ALLOW_FROM`	—	Required for Telegram. Sender allowlist (ids/usernames). Empty = deny all.
`EUE_TELEGRAM_SHOW_STEPS`	`true`	Show step progress lines in Telegram; `false`/`0`/`no`/`off` to hide.
Behavior
`EUE_AUTORUN_PREFLIGHT`	`false`	If `true`, auto run `pwd` and `ls` for directory/project questions.
`EUE_ALLOWED_SKILLS`	—	Restrict to given skills (comma-separated); unset = all allowed.
Output & logging
`EUE_PROGRESS_STYLE`	`verbose`	Progress style: `verbose` or `compact` (summary only).

If your configured model is not found, Eue retries with a fallback model (set EUE_FALLBACK_MODEL; when EUE_PROVIDER=openrouter the default is openrouter:openai/gpt-5.2-chat).

Workspace and commands

Each workspace keeps its own state under <workspace>/.eue/ (skills, tapes, sessions, artifacts, logs). Only lines beginning with , execute; use ,help for commands.

For building from source and contributor workflows, see Development.

Acknowledgments

Eue was inspired by Bub — a project that showed me the beauty of minimal, focused tooling.

Bub's philosophy of "no forced preflight" directly influenced Eue's approach to agent autonomy.

This may be my last project of "朴素编程" (Naive Programming).
朴素编程 can be summed up as: open the editor, type code, run code. The opposite is Vibe Coding: tell your requirements to an AI coding agent; the AI writes the code, runs it, and debugs it.
— TJ (Thaddeus Jiang)