MANİ

Agents do the work.
You make the call.

MANİ is the system I built to run real projects with AI agents. Claude Code, Codex and Cursor work from one set of tools and one shared log. Every output is checked before I see it, and nothing is published, sent or paid for until a person approves it.

Talk to me on LinkedIn Write an email

Watch the 90-second tour

40 tabs. 12 tools. 1 person.

That was my week before MANİ. Most of it was not hard work. It was copying, checking, waiting and remembering. Agents are good at that part. The decisions are still mine.

How it works

From one sentence to a finished, approved result

  1. +Write the product page for the new planner and make a cover image↑
    1

    Say it in plain words

    I describe the job the way I would tell a colleague. No forms, no prompt templates.

  2. Plan4 steps
    Product copy, in English
    Cover image, 1280 × 720
    Checks on both
    Publish? Waits for you
    2

    It becomes a plan

    MANİ turns the sentence into steps, finds the project it belongs to and asks only the questions that matter.

  3. 3

    Agents pick it up

    Coding agents connect through one MCP server, so each one sees the same tasks, rules and past decisions.

  4. Ready for youwaiting
    cover.png · product.md
    ApproveSend back
    4

    You approve

    The result stops at a gate. Approve it, or send it back with a note and the agent tries again.

Gates

Publish. Send. Pay.
A person approves first.

Anything that leaves the system stops at a gate: a post, an email to a customer, a payment, an account change. Agents can prepare all of it. Only a person can let it through. There is no setting that turns this off.

Try the card. This is how a gate looks inside MANİ.

Agent · Inboxwaiting for you

14 supplier emails sorted. 9 reply drafts ready.

  • 3 invoices matched to their orders
  • 2 flagged: the price changed since the last order
Guards

Checked before you ever see it

Every file goes through 16 automatic checks: does it open, is the audio level right, is the text readable, is the page the right size. Then a separate vision model reviews it against a written rubric. Five small robots keep the system itself clean, around the clock.

16automatic checks on every file
1independent model review
205self-tests that run before each release, offline
Who builds it

I'm Tahir Arpacık. I built MANİ and run my own work on it every day.

What I set up for teams

  • A map of your workflow and what is worth automating first
  • One workflow built end to end and tested on your data
  • Claude Code for your repository: instructions, permissions, MCP servers
  • A custom MCP server for one of your internal systems
  • Approval gates that your team controls

How I work

  • Your passwords and keys stay with you. I never log into your accounts.
  • Work happens in your repository or a sandbox you give me.
  • Everything is delivered as files you own.
  • Anything that sends, publishes or spends waits for your approval.

Tell me what takes up your week.

I reply to every message myself.

LinkedIn linkedin.com/in/tahir-arpacik
Email [email protected] Write