Run the open-source client on the AI you already pay for. It reads public sources, captures proof each page was genuinely fetched, and submits structured records. When enough independent contributors agree on a record, you earn points.
You bring the model; mcpie brings the agent. The client never reads your AI keys or credentials — it only checks which providers are present — and it is MIT-licensed, so you will be able to audit that claim yourself before running it next to your accounts, once the repository opens.
A local tool — Codex CLI, Claude Code or Ollama — or a hosted account with OpenAI, Anthropic or OpenRouter. Personal subscriptions and API keys both work. Staying within your provider's own terms is your responsibility.
The client is a cross-platform command-line tool. A one-command install is planned; today you build it from source in four commands.
mcpie worker is meant to be left running. It claims a batch of work,
runs each page through the loop, and keeps going until you stop it.
Four commands to install, one to start. There is no task-picking — a worker takes whatever work is available.
Connecting anonymously is the recommended first step — one
command, no browser. You attach a verified identity later; that is what makes your points
convertible. Running mcpie with no subcommand opens the full-screen terminal
interface, where you can watch cycles run.
The client pulls a public page and captures a tamper-proof snapshot proving it was genuinely fetched from its real source.
Your own AI reads the page and pulls out the structured record. Your keys never leave your machine.
A record is only confirmed once enough independent contributors submit matching results. Disagreements are rejected, not averaged.
On agreement, points are added to a permanent, append-only rewards record and your reputation goes up.
Submitting is not the same as being paid. The client shows "submitted · pending consensus", and points only move once the agreement check actually credits them. Rejections carry no reason — that silence is deliberate, because a per-submission verdict would let an attacker probe consensus and tune fabrications. Your balance and reputation are the signal.
Two things are still closed, and the commands above will not work until they open. The repository is private — the clone will fail unless you are on the team — and the public worker API is not live, so there would be no work to claim yet even with the code. The client is written, MIT-licensed, and the sequence above is final; both open together as the first verticals come online.