MACHINARCLIBRARY

10 pages in the library

docsThe MachinArc handbookUnderstand Machines, connect them into Arcs, and follow work from its first instruction to its final result.docsYour first Machine runExplore the demo, configure a focused worker, and learn to read its execution history.docsDesigning a MachineConfigure the objective, tools, memory, output format, and limits of a single worker.docsConnecting work with ArcsCompose Machines into a visible workflow with decisions, operator review, delays, and outputs.docsPermissions and approvalsSet capability boundaries and review proposed actions before they execute.docsSchedules, webhooks, and eventsStart work from time, an incoming request, or a completed workflow while keeping execution visible.docsConnecting providers and toolsUnderstand mock execution, provider selection, credentials, search, and custom APIs.docsOperating MachinArcThe deployment, persistence, recovery, and verification model behind the workspace.whitepaperFrom prompts to inspectable workThe architecture and thesis behind MachinArc: bounded Machines, visible workflows, and durable human oversight.For investorsBuilding the operating layer for AI workA candid view of the product thesis, intended users, commercial hypotheses, and validation milestones.
Open app
Browse the library
Product guide

Your first Machine run

Explore the demo, configure a focused worker, and learn to read its execution history.

Enter a workspace

Choose Enter Demo to explore without an account, or register an operator account to create a workspace you can return to. Demo access creates an isolated workspace containing example Machines and an Arc with seeded history.

Treat seeded history as sample material. A fresh run is the useful next step: it shows how your instruction passes through the current runtime and appears in the workspace.

Give the Machine a clear job

Open a Machine or create one from a starter. Write an objective that names the input, the expected result, and the boundaries of the task. Add success criteria so you have a concrete way to judge the output.

Keep the initial configuration small. Use Mock Runtime, a short runtime, a modest turn limit, and only the tools the task needs. The example below is an instruction you can adapt, rather than a claim about an existing result.

Example
Objective: Summarize the product notes supplied by the operator.
Success criteria: Identify the main changes, separate facts from questions, and produce a concise briefing.
Boundary: Use supplied material only. Do not contact external systems.

Submit and inspect

Submit a short task, then open its run. Follow the status, tool requests, tool results, and final output. If the task pauses for approval, read the proposed action before deciding whether to continue.

A completed status means the runtime reached a completed outcome. Review the substance of the result separately against your success criteria. Check whether the run used the requested provider or a mock fallback before interpreting any usage or cost figures.

  • Confirm the input matches the task you intended to submit.
  • Inspect tools and approvals rather than judging only the final text.
  • Revise one part of the configuration at a time, then run another small task.

Move to real execution

In a registered workspace, open Settings → Integrations and connect an OpenAI or Anthropic key. Choose that provider and an available model on the Machine. Start with a short task without tools that can change external systems.

Verify the provider, output, usage, and configured cost calculation in the run. Next, enable a calculator and inspect its request and result. Evaluate any approval-controlled external action separately before relying on an automated workflow.

Living documentation · October 2026Product status & limits ↗