CARBON — AI testing harness with Jay

Free, open-source AI software testing inside your coding agent. Jay is CARBON's AI test manager: turn project context into risk-based checks, investigate bugs, and collect evidence in live local reports.

CARBON helps your agent with exploratory testing, regression checks, accessibility, test planning, triage, and confidence assessments. The host agent performs the investigation with its available browser, terminal, and model tools; this MCP server organizes workflows, evidence, settings, and reports. Installing it does not supply a model, browser controller, cloud VM, or autonomous testing service.

Quick start

Requires Node.js 22+. Demo creation and full-workflow helpers also need Python 3.10+. No separate CARBON account, payment, or model API key is needed. Your host's subscription/API costs and limits still apply.

npx -y @testers-ai/[email protected] --doctor
npx -y @testers-ai/[email protected] --print-config

Add this version-pinned configuration using your MCP host's documented settings:

{
  "mcpServers": {
    "carbon": {
      "command": "npx",
      "args": ["-y", "@testers-ai/[email protected]"]
    }
  }
}

The configuration shape varies by host. The command is npx; the arguments are -y and @testers-ai/[email protected]. No port or hosted endpoint is required. On Windows, your host may require npx.cmd. Restart/reconnect the MCP server after configuring it.

An unadorned invocation waits for MCP JSON-RPC on stdin; it is not an interactive terminal chat. --help, --version, --doctor, --print-config, and --list-workflows are safe terminal commands. They never install browsers or modify your agent's configuration.

First use

Ask your agent to load CARBON's jay or carbon MCP prompt and test a specific project or URL. Confirm scope and permissions. Jay starts a bounded assessment, records observed evidence, and returns a protected localhost report URL. Ask your host to open that URL as a browser page, not an HTML source file. Unsupported embedded viewers can use the local browser URL while the server is running, or the saved HTML export afterward.

carbon-settings opens the settings page for budgets, scope, analytics, and testing mode. Installing through npm does not register host hooks or automatically start background checks. Selecting an automatic mode does not create an unavailable host hook. Follow the native plugin setup if you want supported automatic-check integration.

Included workflow prompts

• carbon: Run a bounded risk-based assessment of a project or URL, including business value, stateful checks, persona exploration and an evidence-qualified report.
• carbon-demo: Create a disposable demo project, optionally with existing tests, and perform a first CARBON assessment.
• carbon-test: Test one named feature, flow, API or behavior with reproducible steps and evidence.
• carbon-issues: Hunt bugs with stateful multi-step journeys and persona exploration, not just a static checklist.
• carbon-accessibility: Investigate accessibility using source, keyboard, zoom, visual and dynamic browser evidence.
• carbon-map: Open a screenshot-based map of one saved Free assessment, its checks and findings. Not a cross-build planning workspace.
• carbon-help: Explain Free commands, inspect scope when requested, and suggest the smallest useful next assessment.
• carbon-settings: Read or change project testing defaults and analytics preferences in a protected local settings page.
• j: Talk directly to Jay, CARBON's AI test manager: ask questions, discuss evidence, choose what to test, or request an assessment.
• jay: Talk directly to Jay, CARBON's AI test manager: ask questions, discuss evidence, choose what to test, or request an assessment.
• carbon-background: Set up, pause or review bounded background checks with explicit project consent.
• carbon-pro: Open the CARBON Pro marketing page to explore Pro features, pricing, and upgrade options.
• full: Run Jay's comprehensive, resumable test campaign with parallel checks, reviewed findings and live progress, using this edition's capabilities.

j remains a compatibility shortcut for Jay, not a different agent name. Prompt names and slash-command visibility depend on the host. The carbon-pro prompt only opens the public upgrade page; it does not ship paid testing capabilities.

MCP tools

• carbon_start: start a scoped assessment and local live report.
• carbon_update: record checks, findings, evidence, and progress.
• carbon_report: open reports/maps and export saved evidence.
• carbon_settings: inspect or explicitly change settings.
• carbon_demo: create a new disposable demo without overwriting projects.
• carbon_knowledge: load testing-domain and specialist references.

This is a CLI/MCP package, not a stable JavaScript library API. Native Claude/Codex plugins additionally deliver host-specific skills and hooks; npm alone does not make every host expose the same features. Install only one CARBON server/edition per host namespace to avoid duplicate tools.

Privacy and trust

Analytics is off by default for the npm edition. Optional events contain fixed event names, edition, and version—not source code, prompts, findings, target URLs, project paths, or stable user identifiers. CARBON_ANALYTICS=off or DO_NOT_TRACK=1 forces it off. See PRIVACY.md.

Reports and evidence are stored in your project. Local viewers bind to loopback and protect evidence with private tokens. Your coding agent may still send context to its model provider. Local-first is not an air-gap guarantee. Do not expose report URLs publicly; review exports before sharing.

The npm package contains the MIT-licensed free edition only. It does not include CARBON Pro, customer evidence, or a remote license unlock. npm downloads are not verified installations or unique users; first activation and optional command events are separate, best-effort signals.

Documentation and support

• CARBON overview, documentation, and native plugins
• Source
• Report an issue
• CARBON Pro

MIT License. Copyright testers.ai and Jason Arbon. Jay is an AI persona, not a human reviewer. Evidence and confidence assessments are not guarantees that software is defect-free.