Testers.ai CARBON Test Harness Lite

What Testers.ai CARBON Test Harness Lite MCP server does

Testers.ai CARBON Test Harness Lite MCP server connects an MCP client to the Lite evidence engine used for local QA investigations. The standalone MCP bundle adds stdio tools and prompts around that engine; it is separate from the serverless Codex plugin and does not expose Pro functionality.

The workflows are designed for bounded, risk-based assessment. An agent can be directed at a project, URL, folder, repository, API, application, feature, or requirements. When the target is unclear, CARBON asks for more detail before proceeding. Results can include findings, recorded evidence, screenshots, coverage gaps, prioritized next steps, and confidence scoped to the inspected area.

How it works

The Testers.ai CARBON Test Harness Lite MCP server runs locally through an MCP host. The host determines which browser, terminal, and API actions are available, so the server cannot guarantee that every requested test will execute. Skills can also analyze supplied requirements or existing results without requiring a local runtime.

Evidence and state remain in the selected project's .carbon/studio-lite/ directory. Outputs are local files, including normal HTML reports and portable HTML or JSON artifacts. Recorded evidence may contain information from the target, so review files before sharing them. There is no automatic retention or deletion; project evidence must be removed manually when appropriate.

Setup and configuration

Use the repository's standalone local MCP bundle for MCP clients. The README identifies this distribution as a stdio-based package and distinguishes it from the downloadable Codex plugin. The Codex edition is installed by downloading a release ZIP, extracting it, and selecting the folder through Codex's local-plugin flow; that installation path does not install an MCP server.

The optional evidence helper requires Node 22 or newer. Archive users do not need to run npm install for the bundled helper. Demo artifacts may require Python or the test framework documented by the demo. Creating a demo copy does not install its dependencies or start its server.

No account, OAuth setup, hosted service, analytics configuration, or CARBON model call is required for the Lite edition. Host subscriptions, API usage, local compute, and costs associated with the test target remain separate.

Tools and capabilities

The Testers.ai CARBON Test Harness Lite MCP server is built around the Lite investigation methods, including:

• Bounded assessments and individual feature or journey checks
• Bug hunting with protected stateful or persona-oriented effort
• Keyboard, visual, dynamic, and source accessibility checks
• Screenshot-led maps of checks and findings
• Disposable demos with two test profiles and four fixtures
• Project-local testing preferences
• Revisioned evidence and history
• Findings, screenshots, confidence reports, and coverage-gap reporting
• Portable HTML and JSON output
• Screenshot enlargement, map and check inspection, and feedback import/export

The related command set includes /carbon, /carbon-test, /carbon-issues, /carbon-accessibility, /carbon-map, /carbon-demo, /carbon-settings, /carbon-help, and /carbon-studio. Jay conversational entrypoints are also documented as /jay and /j; these provide discussion about risks, results, and next steps rather than additional testing workflows.

Limitations and notes

Testers.ai CARBON Test Harness Lite MCP server is not equivalent to MCP Studio Pro. It excludes hosted sharing, cloud execution, continuous loops, proprietary benchmarks, Pro commands, and cross-build or 3D map features. It also lacks a native global settings entrypoint, automatic host message handoff, unattended UI write-back, and guaranteed live in-chat report rendering.

Browser support varies by host. Some browsers can watch a selected snapshot file, while others require a reload or import. UI feedback remains a draft until it is imported. Optional background checks require explicit per-project choice and trusted host hooks; unsupported hosts retain manual testing. No automatic fixes or live-site testing are performed by those checks.

The project is open source under MIT. The host model provider's processing is separate from the plugin, and any provider consent or allowance requirements remain the user's responsibility.