Claude Code, Codex, Antigravity and Kimi — side by side on your own Mac.
Runs on the AI you already pay for. No new subscription.
Every engine keeps its colour and its shape — in the pipeline, on the timeline, under every reply. You always know which AI did which piece of the work.
Amounts are illustrative. In the app, every reply shows what it would have cost at API prices — and how much of the plan you already pay for is left.
Bring what you already have. One engine is enough to start — four makes it a cockpit. Three of them sign in the way you always do; Antigravity takes a key. Image and video models such as Seedance aren't accounts you sign into — the engines above call them for you.
That's not a tagline. It's the person who made it, and he is the heaviest user of it.
I've spent about 20 years in web and IT — building, shipping and keeping things running. I've tried every tool there was. And still, the more real work I put through AI, the more overhead showed up: conversations scroll away, nothing stays organised per project, and the day turns into moving between Claude, Codex, Gemini and a code editor.
In the end, I had to build it myself. Nothing in here is a feature that just looked good on a list. Every one of them is something I was missing on a working day — threads, child threads, switching engines mid-conversation, knowledge, code, sharing, AI teams, running several jobs at once. I still fix something in it most weeks.
I opened my first thread on 26 June 2026. Sixty-three days later I'm at 216 threads — about 3.4 a day. The copy on this page was written inside WorkPilot too.
I'm not willing to go back to working without it. That isn't an opinion about my own product — it's just what my week looks like.
The script goes in ChatGPT. The thumbnail needs a different tool. The video needs another one. The code lives in an editor. And every single time, you paste your company profile, your product specs and your tone of voice all over again.
You didn't buy four AI subscriptions to become a copy-paste operator.
Pick the engine per thread — Claude, ChatGPT, Kimi, Antigravity — and the model inside it. Switch mid-conversation and the context comes with you.
Then go further: write one instruction, and WorkPilot breaks it into steps and runs each one on the engine you named — in the order you wrote it, handing each result to the next step.
A library of roles is there from the day you install. Add your own, delete the ones you never use, rename any of them.
A commander assigns the work. Or take the wheel yourself with Playbook mode: your flow, your assignees, your engines, executed exactly as written. Save the playbook and the team runs it again next week.
And none of them are fixed pieces. Open one up, teach it, grow it, and trade it.
You wrote the flow. You picked who does what, and which engine each one runs on. Anything that leaves your Mac still stops and asks.
An agent isn't finished the day it ships.
Four sections you can read and edit — and they don't start empty.
Say "from now on, do it this way" — you approve what gets absorbed.
Export as one file, or import one a friend has raised.
Your own double grows the same way — then answers in your place.
Six modes — chat, image, video, article, audio, code — and each one runs to the end, not to the draft.
Video runs on Higgsfield over MCP. Audio uses ElevenLabs. ffmpeg ships inside the app.
Every other AI app hands you a flat list of chats. The moment a new idea shows up, you open an empty window and explain the background all over again. In WorkPilot a thread can split, inherit, and report back.
Hover a message, press branch, and a child thread opens — already knowing everything that was said up to that point. Or split off whatever you were half-way through typing.
The child holds a snapshot of the conversation as it was, so the parent can race ahead without shifting the child's premise. Files and shared notes stay live — those you always want current.
When the side quest is done, the child summarizes its conclusion and posts it into the parent. You branch freely precisely because you can always come back.
One menu item reads the thread and pre-fills an entire team: name, goal, task list, and which roles to hire. Every member starts with the same context. Nothing gets re-briefed.
Threads that grow long can be split into chapters, each carrying a running synopsis forward so the story survives the length. And parent and child run at the same time — a thread that's busy never blocks the others.
Most AI apps let you attach a file to a message. WorkPilot lets you equip a thread. Name a combination, and it's a Gear Set. One tap to equip it again.
The same question, asked twice. The only thing that changed is what the thread was carrying.
Your documents, specs, photos. Photograph a balance sheet and the figures are read out as text.
Say "turn this into a skill" and the workflow you just did becomes a reusable procedure.
mee, your AI double, carrying your phrasing and your judgment — so the first draft already sounds like you.
Gear rots if you let it. Run the same instruction with a skill and without it, side by side.
Correcting it in the conversation helps once. Filing it helps every time after.
How good an answer is depends on which slice of memory gets handed over this time.
Knowledge is kept in OKF, one Markdown file per topic.
The AI gets the table of contents and reads only the section it needs — about 7% of the payload.
A child thread hands back only its conclusion. Configured connections are passed along automatically.
Instead of pushing everything into one long conversation, the work splits into a parent-child graph and only the conclusions flow back.
Not a chat app talking to a cloud. Your phone talks to your Mac, and the AI running on it.
Speak the instruction, say "over" to send, and go. Your Mac does the work and sends back a summary. Shot something on the way? It's already in your library when you sit down.
And the work that comes back every week doesn't need you to ask at all. That's a Loop.
Recurring work needs no cron, no YAML, no settings screen. Talk it into shape — what to do, when, when to stop, what to check with you, what happens when it's done — and it runs on schedule, on its own.
You describe the job in a conversation and it becomes a brief you can edit later.
Mac asleep at 9am? It catches up when you open it.
Plus the files that run actually produced, ready to open.
Each Loop picks its own engine and model.
Servers over SSH, MCP tools, APIs and outgoing mail all live on one Connections screen.
The hosts in ~/.ssh/config can be pulled in with one click.
Chrome control, Suno and Obsidian are added from presets, not hand-written config.
Passwords and keys go into the macOS Keychain, not into a conversation.
Write a rule down once, and every answer in that thread is built on it.
Rules are injected whole into every linked thread.
Proposals carry the rule they came from — or admit there isn't one.
See exactly what it will stop and ask about.
Developer ID signed. Installs and opens without a warning.
Conversations, knowledge and output are plain files in your Documents folder. Sharing runs on your own folder or NAS. Use Obsidian? Keep them inside your vault instead.
Uploads, posts, deploys and emails always stop for approval.
No API metering. It runs inside the subscriptions you already have.
WorkPilot doesn't resell AI. It signs in to the services you already pay for — Claude, ChatGPT, Kimi — through each vendor's official CLI, so nothing is metered on top. Antigravity runs on the Gemini API free tier.
WorkPilot is free and will stay free — no paid tier, no upsell, no data business. It's funded out of pocket. If it saved you an afternoon, you can pay it forward.
Free. Apple Silicon. Notarized. Updated most weeks — and it updates itself.
