DEV Community

Cover image for Coding Agent Updater: The Tool You Didn't Know You Needed
xbill
xbill Subscriber

Posted on Edited on

Coding Agent Updater: The Tool You Didn't Know You Needed

#bash #linux #ai #productivity

This article provides a step by step guide to keeping every coding agent on a Linux machine current with one command. A set of Bash scripts updates Claude Code, the OpenAI Codex CLI and the three Antigravity tools, and reports what is installed without changing anything.

https://github.com/xbill9/agent-update

One command, update-agents, updates Claude Code, Codex and Antigravity's CLI, hub and IDE. --check reports the installed versions and changes nothing. A failing installer for one agent leaves the others to finish, and the exit status names the one that failed.

What is this project trying to Do?

A developer machine in 2026 runs several coding agents. Each one has its own installer, its own update command and its own release cadence. Claude Code and Codex update from a vendor install.sh. Antigravity's CLI has agy update, while its hub and IDE ship as tarballs that cannot update themselves when installed that way.

The goal: one command that brings all of them to the latest release, one flag that shows where each one stands, and the same behaviour on every machine.

At This Point You Should Have…

  • A Linux machine with bash, curl, tar, sha256sum and python3
  • Any of Claude Code, Codex or Antigravity installed, or none of them
  • Optionally shellcheck, which ./test uses to lint the scripts
  • A clone: git clone https://github.com/xbill9/agent-update

Step 1 — Install the Scripts

The Makefile tests the scripts, then copies them into ~/bin:

cd agent-update
make install
Enter fullscreen mode Exit fullscreen mode

make install runs ./test first: bash -n on every script, shellcheck at warning level, and --help on each one.

test: ok
Enter fullscreen mode Exit fullscreen mode

make install BINDIR=/usr/local/bin puts them somewhere else, and make uninstall removes them.

Step 2 — See Where Every Agent Stands

--check reports installed versions and changes nothing. It is the first thing to run on a machine:

update-agents --check
Enter fullscreen mode Exit fullscreen mode 
==> claude-update
[claude] installed: 2.1.322 (Claude Code)

==> codex-update
[codex] installed: codex-cli 0.166.0

==> agy-update
[agy] installed 1.5.4 (run without --check to update)
[hub] installed 2.29.0, latest 2.31.0
[ide] up to date (2.6.0)
Enter fullscreen mode Exit fullscreen mode

For the hub and the IDE, --check also reads the release channel, so it shows the version an update would install. On this machine the hub is two minor releases behind and the IDE is current.

🔎 Tip: make check runs the same thing from the repository, so the versions are one command away before and after a change.

Step 3 — Update Everything 

update-agents
Enter fullscreen mode Exit fullscreen mode

It runs claude-update, codex-update and agy-update in that order. Each vendor installer prints its own progress; the lines each script adds start with the agent's name in brackets:

update-agents 2>&1 | grep '^\['
Enter fullscreen mode Exit fullscreen mode 
[claude] 2.1.322 (Claude Code) -> 2.1.322 (Claude Code)
[codex] codex-cli 0.166.0 -> codex-cli 0.167.0
[agy] 1.5.4 -> 1.5.4
[hub] installed 2.29.0, latest 2.31.0
[hub] now 2.31.0  (previous kept at /home/xbill/Antigravity-x64.prev)
[ide] up to date (2.6.0)
Enter fullscreen mode Exit fullscreen mode

One run moved Codex to 0.167.0 and the hub to 2.31.0, and left Claude Code, the agy CLI and the IDE alone, since they were already on the latest release. The exit status was 0.

Names pick a subset:

update-agents claude codex      # just these two
agy-update --check hub ide      # one tool, two of its parts
Enter fullscreen mode Exit fullscreen mode

Each per-agent script also runs on its own. A running agent keeps its old binary until it restarts; the new version starts with the next session.

Step 4 — Claude Code and Codex: The Vendor Installer, Unattended

Both scripts call the vendor's own installer, which installs the agent when it is missing and updates it otherwise:

curl -fsSL https://claude.ai/install.sh | bash
curl -fsSL https://chatgpt.com/codex/install.sh | CODEX_NON_INTERACTIVE=1 sh
Enter fullscreen mode Exit fullscreen mode

CODEX_NON_INTERACTIVE=1 makes the Codex installer run without prompts, so it works from cron or from another script. set -o pipefail makes a failed download fail the update, along with a failed install. Each script prints the version before and after the update.

Step 5 — Antigravity: Three Tools, Three Channels

agy-update handles three targets, each from a different source:

Target What Installed at Update source
agy CLI ~/.local/bin/agy agy update
hub Antigravity ~/Antigravity-x64 hub updater channel, tarball
ide Antigravity IDE ~/Antigravity IDE IDE updater API, sha256-checked

The hub and IDE tarball installs refuse to update themselves with APPIMAGE env is not defined, so the script does the whole update:

  1. Read the channel for the latest version and the download URL.
  2. Read the installed version: ideVersion in the IDE's product.json, and for the hub, the version in the package.json packed inside resources/app.asar.
  3. Compare with sort -V; stop if the installed copy is current.
  4. Download to a .part file and rename it when complete.
  5. Verify: the IDE tarball against the sha256hash its API publishes, the hub tarball with a full tar -tzf listing, since its channel publishes no hash.
  6. Extract to a staging directory, then swap it into place.

The swap keeps the previous copy as <dir>.prev. When the extraction fails, the installed copy stays untouched; when the final move fails, the previous copy goes back.

🔎 Tip: a hub or IDE that is not installed is reported and left alone. --force installs it, or reinstalls one that is already current:

agy-update --force ide
Enter fullscreen mode Exit fullscreen mode

Step 6 — Keep Going Past a Failure

update-agents runs without set -e. A failing child is recorded and the loop moves on, so a broken Codex installer leaves Claude Code and Antigravity to update. At the end it lists what failed on stderr and exits 1:

update-agents: failed: codex
Enter fullscreen mode Exit fullscreen mode

Every script follows the same exit contract: 0 on success, 1 if an update failed, 2 on a bad argument.

update-agents --bogus
Enter fullscreen mode Exit fullscreen mode 
unknown argument: --bogus
Enter fullscreen mode Exit fullscreen mode

That run exits 2. A typo in a cron line stops at the argument parser before any installer runs.

Step 7 — The Help Is the Header

Each script's header comment is its documentation, and --help prints it: usage, options, environment and exit status.

update-agents --help
Enter fullscreen mode Exit fullscreen mode 
update-agents — update every coding agent on this machine.

Runs each <agent>-update script that sits next to this one, in order, and
keeps going when one fails, so a broken installer for one agent does not
block the others.

Usage: update-agents [--check] [agent ...]   (default: all agents)
  --check   report installed versions only; passed through to each script
  agent     claude | codex | agy
Enter fullscreen mode Exit fullscreen mode

One sed line in each script does it, so the help and the code sit in the same file and change in the same commit.

Step 8 — Adding Another Agent

A new agent is three changes:

  • bin/<agent>-update, with --check, --help, and exit 2 on an unknown argument
  • its name in AGENTS in bin/update-agents
  • a row in the README table

The repository ships a Claude Code skill, /add-agent, that makes all three, and a hook that runs shellcheck on every edit to a script.

🔎 Tip: the umbrella is named update-agents, with the verb first. A machine that runs every *-update script in ~/bin from a scheduled job would otherwise run each agent's update twice: once through the umbrella and once directly.

Compare and Contrast

By hand update-agents
Commands per machine one per tool, each different 🥇 one
See versions without changing anything run each --version, check each channel 🟢 --check, latest included for hub and IDE
Hub and IDE tarball installs download, verify, extract, move by hand 🟢 automatic, previous copy kept
IDE download verified only if you compare the hash 🟢 sha256 from the vendor's API
One installer fails the rest wait for you 🟢 the rest finish; exit 1 names it
Unattended Codex prompts 🟢 CODEX_NON_INTERACTIVE=1

So, Which One?

For a machine with more than one coding agent, update-agents. Run update-agents --check to see where each one stands, then update-agents to bring them current. For a single agent, its own script does the same with the same flags.

Summary

The goal of this project was to keep every coding agent on a Linux machine current with one command. The key to the solution was one script per agent with the same flags and exit codes, and an umbrella that runs them all and keeps going past a failure. The results were:

  • 🟢 update-agents updates Claude Code, Codex and Antigravity's CLI, hub and IDE.
  • 🟢 One run on 2026-11-10 updated Codex 0.166.0 to 0.167.0 and the hub 2.29.0 to 2.31.0, and left the three current tools alone.
  • 🟢 --check reports installed versions, and the latest available for the hub and IDE, without changing anything.
  • 🟢 Hub and IDE tarballs are verified and swapped in place, with the previous copy kept as <dir>.prev.
  • 🟢 A failed installer is listed and the others finish; exit status 1 for a failure, 2 for a bad argument.
  • ⚠️ The hub channel publishes no hash, so its tarball is checked for structure only.
  • ⚠️ Linux x64 only; macOS and Windows are not covered.

Scope: one Linux x64 machine, scripts at commit a0f675b, with Claude Code 2.1.322, codex-cli 0.166.0, agy 1.5.4, hub 2.29.0 and IDE 2.6.0 installed as reported by --check on 2026-11-10. The output above comes from ./test, --check, --help, a bad-argument run and one full update-agents run on the same day. Parts of the writing were done with AI assistance (Claude).

The strategy for keeping coding agents current was validated with an incremental step by step approach.

References

Top comments (0)