mexxi

Coding agents write code fast. mexxi makes them prove it works.

A coding agent is an AI tool that works in your repository: it reads the code, runs commands and makes changes. Give the one you already use, Claude Code, Codex or OpenCode, a task in plain words. You get back one pull request that passes your repository's own checks, with screenshots from a real browser and a plain call on whether it is safe to merge. You still decide: mexxi never merges. It runs on your machine, with the agent and subscription you already have, and needs no service or account of its own.

mexxi is Maltese for "to lead, to manage". The names explains the rest of the Maltese.

The problem

Coding agents can now take a whole task: read the code, make the change and open a pull request. That moves the slow part of the job. Writing the code is quick; trusting it is not.

What an agent hands back on its own is hard to trust:

• It skips tests. One run is careful; the next ships without proof, and says "all tests pass" either way.
• It never looks at the UI. The code compiles, and nobody has opened the page.
• It loses time to its environment. It retries the same failing command and hides failed steps.
• Its pull request needs reverse-engineering. You read the whole diff to learn what it did and why.

So you either read every line, which costs the time the agent saved, or merge on trust.

How mexxi solves it

mexxi moves the checking into the run, so the pull request arrives with its proof. The ideas behind mexxi explains each one.

• Each task gets its own copy of the repository, so your checkout stays untouched and tasks can run side by side.
• A pass/fail command defines done. Your repository's checks, npm run verify by default, must pass before work starts and again at the end. ibni, the command-line tool the agent runs, runs those checks itself and records the exact code they passed on, and it refuses to push any other code. A later change has to be checked and recorded again.
• Somebody looks at the page. For UI changes the agent drives the running app in a real browser at phone, tablet and desktop sizes, and the pull request shows before and after.
• A second reviewer judges it. A reviewer agent that cannot edit code judges substantial changes. It is an extra opinion; the checks and the browser are the proof.
• Guardrails in code as well as in the rules. A hook backs up the rules against force pushes, merges and pushes to the default branch, and another stops an agent that keeps retrying a failing command.
• The first screen does the explaining. The pull request opens with the merge call, the decisions only you can make, and the evidence: which checks ran, and which did not.

To see the difference, compare two pull requests for the same request in a small demo app: without mexxi and with mexxi. Reviewing a mexxi pull request explains each part and what to check in a minute.

Getting started

You need one of Claude Code, Codex or OpenCode, Node.js 22.13 or newer, and Git. You also need the forge's CLI, signed in, to open the pull request: gh for GitHub or glab for GitLab.

Three tools are optional. The Playwright MCP server lets the agent drive a browser itself; without it, the pull request's evidence says which checks did not run. The review tool OpenCodeReview (ocr) picks what the reviewer looks at. ffmpeg makes GIFs.

Paste this into your agent to install mexxi. It checks the prerequisites, asks what it cannot detect, and shows what it will change before it changes anything:

Set up mexxi for me. Read https://raw.githubusercontent.com/alexportelli12/mexxi/main/skills/ibda/SKILL.md and follow it.

Then paste this in a repository to prepare it. /lesti opens one pull request that adds what mexxi needs, such as the verify command, and fixes the checks that already fail.

Make this repository ready for mexxi: run /lesti here.

From the repository's default branch, type this in your agent's chat to ask for a change with /wettaq, the command that builds it:

/wettaq show the last sync time on the settings page

The setup guide covers installing by hand and what each prerequisite is for.

What's in the box

Coding agents load their instructions as commands you type, agents they can hand part of a task to, and skills they load only when a task needs them. The glossary explains each term.

• Commands you type in your agent: /wettaq builds a change, /ittestja tests a UI by hand in a browser, and /lesti gets a repository ready. Three more look after the workflow itself: /fassal writes UX principles, /sewwi changes the workflow and /irrifletti looks back at sessions.
• Agents: an independent code reviewer and a UI designer, each with the tools it needs and no more.
• Skills: browser checks, repository mapping, commit and pull request writing, and the review step.
• ibni, which takes each task's worktree from prepare to publish, with step timings.
• Guardrails: rules forbid force pushes and merges, and a guard hook backs them up. When one command fails a third time, the agent must stop and rethink.
• Retrospectives: /irrifletti measures how recent sessions went, privately, on your machine.
• One experiment: Jev triage, which asks an outside model which checks a task needs and only records the answer. It sends metadata, never code, and is off unless you turn it on.

The same files serve Claude Code, Codex and OpenCode; Host details says what each one adds.

The names

The tools are named in Maltese, the author's native language, like Verifika. Each name is an everyday word for the tool's job: /wettaq means "to carry out", /lesti "to get something ready" and ibni "to build". The names is a small dictionary of all of them.

Learn more

The details live in the docs, generated from this repository so they cannot drift from it:

• The ideas behind mexxi: each problem, and what mexxi does about it.
• Reviewing a mexxi pull request: what to check, and what stays your job.
• How a run works, step by step.
• Commands, agents and skills.
• The ibni CLI reference and every configuration key.
• Where the problems come from: what my own sessions measured.
• The principles behind it, and a glossary.

Using it yourself

Go for it. I made this repository public to share the workflow I use day to day. It is opinionated around how I like to work, so take the parts that fit and change the rest.

mexxi is released under the MIT licence. To send a change back, read CONTRIBUTING.md first.

Credits

The frontend-design and skill-creator skills are adapted from Anthropic's skills repository, under the Apache License 2.0. Each folder keeps its licence in LICENSE.txt.