Elanous

Blog · 2026.09.29

The agent drives the agent

Inside Codex and Claude Code — plugins, trust screens and hand-offs

Last time Elanous drove a plain terminal. This time the terminal has another coding agent in it. Elanous opens Codex or Claude Code in a real terminal, works inside the agent’s own screens — its plugin menu, its folder-trust prompt — and hands one mission from one agent to the next.

엘라누스가 터미널 안의 Codex·Claude Code 를 직접 몹니다. 에이전트 자기 화면의 플러그인 메뉴와 신뢰 확인을 넘기고, 한 미션을 에이전트에서 에이전트로 넘깁니다.

Teaser (60 s). Some shots are staged; the terminal and web-app shots are real recordings.

A plugin, installed from inside the agent

Skills come as plugins from a signed marketplace that Codex can read. So instead of preparing the agent beforehand, Elanous can install one in the middle of a mission, in the agent’s own screen:

elanous agent-mission mission --backend codex --plugin elanous-basics@elanous "make this PDF folder searchable"

It opens /plugins, moves to the marketplace entry, installs it, waits until the screen says so, closes the menu and only then types the mission.

Codex in a terminal. Its plugin screen shows 'elanous-basics · Installed · elanous — Official elanous essentials for research, onboarding and everyday productivity', and above it the line 'Installed elanous-basics plugin.'
Codex’s own plugin screen, driven by Elanous: the official elanous-basics plugin, installed.
Codex working on the mission: it read skill files, checked which PDF tools exist (pdftotext, pdfinfo, ocrmypdf, tesseract, mutool), read README.md and index.sh, and shows 'Working'.
Then the mission: it reads the skills and checks which PDF tools the machine has before it changes anything.

The official marketplace is public — ElanvitalAI/elanous-plugins. Its index is signed with a published key (4d809b69). You can also add it to Codex yourself: codex plugin marketplace add ElanvitalAI/elanous-plugins.

One mission, several agents

--chain codex,claude,codex,elanous passes one mission along: Codex builds, Claude Code reviews what changed, Codex applies the review, and Elanous checks the evidence, commits and opens the pull request. The reviewer gets the diff; the next agent gets a short summary — never the raw screen. Claude Code is driven through its own terminal with your own sign-in; Elanous does not switch it to a non-interactive mode.

The first full run went through all four steps today. The last step — opening the pull request — failed as expected, because that demo repository’s remote was a local folder, not GitHub.

What running it for real caught

  • Trust screens. The first time an agent opens a folder it asks whether to trust it. Claude Code’s prompt defaults to No, and our driver spent a whole mission on that one screen. Elanous now answers the trust prompt before it types anything.
  • Menus take a moment. Codex had /plugins typed but not yet open, and the driver judged the screen too early. It now waits for the menu and presses Enter once more if the command is still sitting in the input line.

Neither showed up in tests with a fake screen. Both showed up the first time a real agent was on the other side.

Try it

elanous agent-mission mission --backend codex --plugin elanous-basics@elanous "…"
elanous agent-mission mission --chain codex,claude,codex,elanous "…"

Source: github.com/ElanvitalAI/elanous · Docs: docs.elanous.ai · Plugins: ElanvitalAI/elanous-plugins

Beta — install it today; provided as is (Apache-2.0). Get update news at elanous.ai. Read also It drives the terminal itself.