Free and open source · macOS

Design the system behind the prompt.

You already write the whole loop by hand — who does what, what gets reviewed, what happens when tests fail. Anthill lets you draw it instead, hands you the prompt, and then shows you what your agent is actually doing.

One email when the first build is ready. Nothing else, ever.

01Today

The whole workflow lives in one paragraph you re-type every time.

Roles, an order, a review step, a loop, a place where you want to be asked first — you write all of it as prose, in one breath. Then you do it again for the next task, because there was nowhere to keep it.

claude

Build the checkout flow. Have one agent design it and a second one attack the design before anything is written. Then implement it, run the suite, and loop back on failures until the tests pass. Ask me before changing any public contract.

Two agents, a review gate, a loop and an approval. Four structural decisions, none of them visible, none of them reusable — and once it is running, no way to tell which one you are in.

02Draw it instead

Draw the workflow the way you would on paper.

Blocks are steps. Connections say what happens next, and they carry a type — so a loop is a loop on the page, not a sentence you hope was read. Save it, open it next week, change one step.

Next — accepted, move on

Rework — send it back

Question — someone must answer

Stop — this path ends

And design the agents, not just the steps.

An agent is a reusable profile: a name, a role, what it is for, which model it runs on. Assign one to several steps and it is the same agent returning to the work, not a stranger each time. Anthill writes the subagent files to match.

DeveloperImplementation · sonnet2 steps

TesterVerification · sonnet1 step

Ready-made profiles to start from, and they are ordinary files you can edit.

03Not an orchestrator

Keep your own tools. Nothing to hand over.

Anthill does not run your agents and does not want your keys. It compiles the diagram into one prompt — you paste it into Claude Code, Codex or Pi exactly as you always do, signed in as yourself, and press enter.

Anthill · promptCopied

GOAL
Land the change with the suite green.

AGENTS
- developer — implements, and fixes what
  the tests find.
- tester — runs the suite and reports the
  real result.

STEP 1 · Implement (developer)
STEP 2 · Run tests (tester)
  rework → Step 3 when tests fail

Copy promptPlus the subagent files

claude — your session

$ claude
> GOAL
> Land the change with the suite green.
> AGENTS — developer, tester▍

Your sign-in, your machine, your usage.

Claude CodeOpenAI Codex CLIPiNo API key to enter. Nothing leaves your machine.

04Watch it happen

Stop re-reading the transcript to find out where you are.

Anthill recognises the session and lights up your workflow as the work moves through it. Which step is running, which are finished, how long each one took, which one is waiting on you — at a glance, not by scrolling back.

Activitynewest first

◌Bash14:26:19

Likely · Run tests

Run the project's test suite

◌Tester14:26:02

Likely · Run tests

Subagent started — mapped by agent name

?WebFetch14:25:58

Not mapped to a step

No step claimed it, and Anthill says so rather than guessing

Claude Code14:25:51

Confirmed · Implement

Starting on the limiter. I will put the counter behind a small interface first.

Anthill reads the records the CLI already writes on your machine. No transcript, no file contents, and none of the model's private reasoning.

Where it pays off

Built for the long jobs — many steps, many agents, hours of work.

A one-line fix needs none of this. A migration, a refactor across packages, a feature that has to be designed, reviewed, built and tested does — and those are exactly the jobs where a paragraph of prose stops being enough and the agent disappears for an hour.

Many steps

Ten or twenty stages with reviews and loops between them stay readable as a diagram, and stay editable next week.

Many agents

Specialists that hand work to each other, argue with each other, and come back to their own work later.

Long runs

When a job takes an hour, the question is not what the agent said — it is which step it is on, and whether anything is waiting on you.

Questions

Does Anthill run my agents?

No. It plans the work and then watches. The prompt it produces is an instruction for Claude Code, Codex or Pi, and the ordering, conditions and limits in it are for that tool to honour — Anthill does not start, stop, pause or answer a session, and turning off observation leaves a running session untouched.

How does it see what is happening?

It reads the records the CLI already writes on your machine, and it says how confident it is about each one. An event it cannot tie to a step is shown as unmapped rather than guessed at. Nothing leaves your computer.

Do I need an API key?

No. Anthill borrows a coding CLI you have already installed and signed into. There is nothing to connect and no key to paste.

Which tools does it work with?

Claude Code, the OpenAI Codex CLI and Pi. One of them installed is enough to start; the workflow you draw is not tied to any of them.

Is it really open source?

Yes, and free. Workflows are ordinary JSON files on disk, so one you build is yours to read, diff and commit next to the code it is about.

What do I need to run it?

macOS, and a coding CLI you already use. The first build is in early access — the waitlist gets it as soon as it is worth your time.

Stop writing loops. Start shaping them.

Anthill is free and open source, and the first macOS build is close. Leave an email and we will write once, when there is something to run.