Build tools
Create and inspect recipes and builds through authenticated MCP operations tied to the OpenFactory workflow.
Tool access with a machine boundary
Agents can request builds, start VMs, inspect screenshots, interact with a desktop, and run test scenarios through explicit tools. The Linux image and VM provide a concrete environment boundary; they do not make model behavior deterministic.
Desktop interaction
Capture screenshots and drive a full VM desktop for applications that cannot be tested through an HTTP response alone.
Repeatable scenarios
Define and rerun functional or visual checks, retaining the result and supporting captures.
Isolation boundaries
Use VM and organization boundaries to keep agent work separate from the host and from unrelated customer resources.
From agent instruction to inspected result
The useful loop ends with a running system and captured evidence, not with generated files.
- 01
Describe
Give the agent the target machine or application behavior and the checks that define success.
- 02
Build and operate
Use MCP tools to build the image, start the isolated VM, and exercise the real desktop, terminal, or service.
- 03
Record and review
Preserve test outcomes and screenshots so a human or downstream workflow can inspect the result.
Agent infrastructure questions
Which agent clients can use OpenFactory?
Any client that can connect to the authenticated OpenFactory MCP endpoint and invoke its exposed tools can participate in the workflow.
Does isolation make AI output deterministic?
No. VM isolation and repeatable tests constrain and observe execution. They do not guarantee deterministic model decisions or generated code.
Can agents test graphical applications?
Yes. OpenFactory can expose screenshots and desktop interaction for full Linux VMs, allowing functional and visual checks beyond browser-only applications.
Give one agent a testable definition of done
Pilot a real workload whose success requires a build, a running machine, and evidence a reviewer can inspect.