mexxi

Host details ​

mexxi works with the coding agent you already use, so you need not switch tools to get it. The installer sets up Claude Code, Codex and OpenCode from the same shared files, so a change to the workflow reaches all three. What differs is what each one adds, how it takes its settings, and the step or two you take after installing.

  • OpenCode reads the shared commands, skills and agents. Its config is rendered from yours.
  • Claude Code adds a pane beside the chat that shows where the run stands, and fixes shell commands that would otherwise fail, such as unquoted globs.
  • Codex runs commands as $mexxi-… skills, with a reviewer that can only read.

Claude Code ​

The installer merges only the keys it owns into your settings.json: permissions.deny, env.BASH_MAX_TIMEOUT_MS and the hook entries. It records what it added, so changing config.json later removes only those entries. It also installs the mods that config.json names from this clone. Open sessions pick up changes with /reload-plugins.

Codex ​

Start a new session after installing. Codex skips new hooks until you approve them once in /hooks. A non-empty AGENTS.override.md in Codex's home takes precedence over the installed AGENTS.md; remove it if you want the shared guidance. Run commands as $mexxi-wettaq <request>.

OpenCode ​

Its opencode.json is rendered, including the Playwright server and the guard plugin.

Playwright MCP ​

The pin, browser path and output directory come from the browser group in config.json. Every agent session with the server registered starts it, so it runs with node from a copy the installer keeps in ~/.local/share/mexxi/playwright-mcp/<version>, never through npx. That saves about 100 MiB per session, which npx's npm wrapper would hold. Claude Code and Codex keep MCP servers in files that are yours, so the installer prints the registration lines and never runs them:

bash
claude mcp add --scope user playwright -- node ~/.local/share/mexxi/playwright-mcp/0.0.83/node_modules/@playwright/mcp/cli.js --headless --isolated --browser=chromium --output-dir=<worktree root>
codex mcp add playwright -- node ~/.local/share/mexxi/playwright-mcp/0.0.83/node_modules/@playwright/mcp/cli.js --headless --isolated --browser=chromium --output-dir=<worktree root>

The printed lines spell out your home directory. They use Playwright's own Chromium: install it once with npx -y @playwright/mcp@0.0.83 install-browser chromium. With browser.executablePath set, they pass --executable-path instead. install --check fails when the pinned copy is missing. It also fails when a registered server is missing, runs through npx, or has another version, browser, output directory or --allow-unrestricted-file-access. To change the pin, edit config.json and rerun the installer, which fetches the new version. Then remove the old server and register it again with the new lines; the next installer run removes the old version once no host runs it.

The output directory stays the worktree root, not a folder inside it. The server writes only inside its output directory or its working directory. Rendered QA saves screenshots to each task's evidence directory, which sits beside its worktree.

The guard ​

Claude Code and Codex run scripts/hooks/guard.mjs before every shell command, and OpenCode runs the same code through a plugin. Codex runs it only once you trust it in /hooks. The rules in AGENTS.md are what bind the agent; the guard is a backup for some of them, not a substitute. A command it lets through is not thereby allowed. A refusal reaches the agent as the tool's error, with the reason.

A second hook, the loop breaker, tells the agent to re-diagnose when the same shell command fails a third time.