CARBON

CARBON Core — free, open-source AI testing for your coding agent

Your AI test manager agent. Jay examines your project and changes, prioritizes risk, coordinates software testing and turns findings into visual evidence. Use CARBON Core for QA, bug finding, accessibility testing, regression-test planning and evidence-backed quality reports. Free under the MIT license.

Start with /carbon:carbon in Claude or /carbon in Codex. Watch a compact execution plan, the current check and recorded screenshots as the investigation progresses. Reports distinguish executed tests, proposed work and untested gaps.

Website and guide · Command reference · Download releases · Codex edition · Report an issue

The product name is CARBON Core. The repository and technical plugin name remain carbon-free and carbon so existing installations keep receiving updates.

CARBON via npm / MCP

The free, MIT-licensed runtime with Jay is also packaged as @testers-ai/carbon-mcp. Requires Node.js 22+; demo and full-workflow helpers also require Python 3.10+.

Use the npm setup guide and verified release artifacts for MCP installation. The source distribution version is not a claim of npm publication.

See the npm setup guide for the 13 MCP workflow prompts, eight tools, host requirements, and privacy boundaries. npm analytics defaults off; npm installation alone does not register native agent hooks or supply model access. The npm release procedure documents source and package checks.

Cursor source candidate: CARBON Core 1.32.23-cursor.5 · Cursor-specific privacy. Bounded native HTTP smoke passed; marketplace review remains pending.

Grok Bot source prerelease: CARBON Core 1.0.2 · Grok-specific privacy. Local protocol tests passed; native Grok testing and marketplace review remain pending.

Your agent's plan. Your local workspace.

Use the model access you already have. In Claude Code, Codex, and other coding agents, CARBON's default testing workflows use the host's configured model access. When the host uses an eligible subscription, CARBON uses that allowance—no separate CARBON model API key is required. If the host uses API credentials or a company LLM proxy, that billing and data policy apply instead. Plan limits and extra-usage charges still apply; CARBON Pro licensing is separate.

Local execution and local reports—not a promise that no data leaves your machine. For local runs, project files, test evidence, and reports are stored in your local workspace. Your coding agent can send selected code, screenshots, prompts, and other context to its configured model provider. Local CARBON is not automatically offline or air-gapped. Cloud runs, external integrations, and optional direct-model checks have separate data flows and may have separate costs. Command-usage analytics vary by edition; review that edition's privacy policy and settings.

To confirm your billing route, check /status in Claude Code or codex login status in Codex. A configured API key can select API billing instead of subscription usage. Standalone MCP clients must supply their own model access and execution tools; installing an MCP server does not include a model subscription.

Free and open source. AI software testing and quality assurance inside Claude.

AI testing inside your coding agent. Built by testers.ai.

CARBON gives your coding agent risk-based testing instructions and a local evidence runtime. The agent inspects your authorized project, runs checks using its available browser, terminal or API tools, and records findings in a live report. CARBON's MCP server does not independently run the investigation or manufacture results.

Visual testing evidence

Four compact report views help you move from a decision to its evidence:

• Release Brief: findings, next actions and outcomes for the selected checks.
• Evidence Lens: page screenshots linked to checks, reproduction and fixes.
• Journey Atlas: recorded paths and step outcomes, without invented connections.
• Live Investigation: the current check, its purpose and recent activity.

Zoom screenshots, keep the full record one click away, and export portable HTML/JSON. Live reports retain their last good snapshot and quietly reconnect after interruption. Unknown and deferred checks remain distinct from passing checks.

Role-based AI testing team

Jay, the AI test manager, now introduces the assessment and appears in reports. Specialist profiles use role-based names and new light/dark icon identities rather than human names and photographs. They remain AI perspectives, not human reviewers. See the AI testing team, visual profile gallery, command guide, and MCP tool reference. Role names describe AI testing perspectives, not independent human review.

Included commands

Twelve distinct commands, with thirteen skill entrypoints including the j alias. Host prefixes differ; Claude normally uses /carbon:command.

All recorded findings, persona journeys, HTML/JSON exports and current-run maps are included. No CARBON account, payment or finding unlock is required. Your coding-agent subscription or API usage is separate. The usual 20-minute/20-check planning default is adjustable, not a coverage guarantee or hard run quota.

Install and start in Claude Code

Requirements: Node.js 22 or later; Python 3.10 or later for the optional demo creator. No npm install is needed for the MCP runtime. Browser automation depends on the tools your coding agent has available. Optional demo tests have their own documented dependencies.

git clone https://github.com/jarbon/carbon-free.gitclaude --plugin-dir /absolute/path/to/carbon-free

Replace the path with your checkout. Claude Code normally prefixes plugin skills with the plugin name. Try these three examples:

1. /carbon:carbon — assess the current project. If the folder is empty or the target is unclear, Jay asks what to test: a URL, folder/repository, API, app/feature, or requirements. With a clear target, he runs a broad assessment and reports findings, evidence-qualified confidence, gaps, and next steps.
2. /carbon:carbon-issues ./my-app — investigate bugs, including state transitions and multi-step journeys, in an authorized project.
3. /carbon:carbon-accessibility http://localhost:3000 — test an app you control; report actual evidence and untested criteria separately.

Use /carbon:carbon-map for the latest assessment or /carbon:carbon-settings to adjust defaults. Names and embedded-browser support can vary by host/version; use the host's skill picker. Install one CARBON edition at a time because editions share the carbon namespace. This repository is not a claim of Anthropic approval or listing.

The bundle also contains a Codex manifest. Other MCP hosts may launch node /absolute/path/to/carbon-free/server/index.mjs, then expose the bundled skills through their supported mechanism. A chat host without local process/browser access cannot provide the same runtime behavior. Test the surface you intend to use.

What runs, writes and connects

• The declared MCP entry point is server/index.mjs, using Node standard libraries.
• Demo creation invokes bundled runtime/carbon/scripts/carbon_demo.py via Python. Fixtures are synthetic examples, some intentionally imperfect; never deploy them as production apps.
• Evidence and report snapshots are saved under your project's .carbon/free/. Existing reports are retained. Uninstalling does not erase evidence.
• Live report/settings views bind to 127.0.0.1 and use random access tokens. The agent opens them with its own browser/panel tool where supported. Saved HTML remains usable after the runtime stops. Review exports before sharing.
• Anonymous command analytics is enabled by default and sends fixed command event names to https://cdn.usefathom.com/. It does not send prompts, source, target URLs, screenshots, findings, credentials or a stable user identifier. Network metadata is visible to that service. Set CARBON_ANALYTICS=off before startup or disable analytics in settings. No telemetry is sent merely by importing the module.
• Global analytics preference and delivery diagnostics use .config/carbon/ in your user home. Project evidence has no automatic expiration; you control retention.
• Your coding agent's model provider still processes the context you share with it. Browser/API checks contact the authorized target and potentially its third parties. Local-first does not mean air-gapped.

See the privacy policy for data handling, retention and contact details. There is no private benchmark fetch, reviewer service, remote MCP backend or Pro unlock code in this bundle.

Interpreting results

Only recorded evidence is displayed. Unexecuted tests remain blocked or deferred; suspected issues are distinguished from demonstrated failures. Confidence is scoped engineering judgment, not a calibrated probability that the whole product is correct. AI persona feedback is simulated, not human research. Accessibility results are not a certification or legal opinion. This Free runtime does not generate formal DOCX conformance documents or run continuous automatic repair.

Troubleshooting and support

• Missing tools: check Node with node --version, reload the plugin and inspect the host's MCP connection status. Run claude plugin validate . from the checkout.
• Expired live link: reopen the saved run with carbon_report or open its saved HTML.
• No browser available: run authorized static/API checks and disclose untested UI scope.
• Demo setup failure: check python3 --version and the selected fixture's README.
• Do not post private code, report tokens, screenshots or customer records in public issues.

Product/security support: [email protected]. Reproducible non-sensitive problems: GitHub issues.

Runtime regression checks

From the repository root, run node tests/smoke.mjs .. The checks use temporary synthetic projects and disable analytics. They exercise the MCP interface, demo copying, report lifecycle, evidence exports, settings persistence, authentication, cross-origin rejection and symlink protection. They do not certify accessibility, security, or performance of an application under test.

License and product information

Open source under the MIT License for CARBON on both Claude and Codex. Read LICENSE and NOTICE.md. Product edition information is available on the CARBON website.

Talk to Jay

Use /jay or /j to talk directly to Jay · AI Test Manager Agent. The MCP prompt router also accepts J. Skill names are lowercase; if a host does not resolve /J, use /j. Claude commonly prefixes these as /carbon:jay and /carbon:j; Codex may show a plugin-prefixed command or $jay in its skill picker. The plugin cannot reserve a global slash name in every host.

• /jay What should we test next? — discuss risks and the evidence we have.
• /j Test this project — run this edition's normal CARBON assessment.
• /jay Explain the last report — interpret recorded results without rerunning.
• /carbon remains the direct broad-testing command.

Jay replies in first person, in a frank, practical tone inspired by Jason Arbon. He is an AI persona, not Jason. With no clear testing target he asks for a URL, folder/repository, API, app/feature, or requirements. He does not create a demo unless asked. Questions do not automatically start tests, and fixes still need authorization. Lite keeps its existing capabilities; these are two conversational entrypoints, not two new testing workflows or an upgrade to Pro.

Jay's bundled icon appears in compatible command pickers and CARBON report views. Host chat avatars and image sizing are host-controlled; text-only chat uses “Jay · AI test manager” instead of a large image.

Jay quick checks

Jay gives changed code a quick second look by default with trusted host hooks. Existing disables remain off. Local idle checks: 2-minute delay, 3-minute runs, one worker, 3 runs and 10 reserved minutes per UTC day. No auto-fixes, installs or live-site requests. Source pattern matches are possible issues, not confirmed bugs. Jay reports category counts in a grayscale report with his icon. Optional Claude-only AI review and exact executable test commands require separate approval. The foreground coding agent uses its configured model access; idle checks make no model calls by default. Use carbon-background to pause, review, re-enable or disable everywhere. See behavior, controls and privacy. Unsupported hosts retain manual testing.

Explore CARBON Pro

Run /carbon-pro to open CARBON Pro features and pricing. In Claude Code or Cowork, the plugin-prefixed command is /carbon:carbon-pro. In Codex, choose carbon-pro from the plugin skill picker (or use $carbon-pro). MCP clients expose the carbon-pro prompt; the full runtime also accepts carbon_pro. This is a navigation shortcut available in every edition, not a paid testing workflow. It does not run tests, start checkout, install Pro, or expose customer downloads.

Optional usage analytics

Analytics records fixed command names, first activation, assessment lifecycle, report/settings use, edition and version. It never includes prompts, code, project paths, target URLs, screenshots, findings, credentials or a persistent user identifier. Enabled events go to Fathom at https://cdn.usefathom.com/; the service necessarily sees connection metadata such as IP address.

Claude editions retain default-enabled analytics. Codex editions default off; enabling requires the user's explicit choice. Run node <plugin>/scripts/usage-analytics.mjs status, on, or off. Global CARBON_ANALYTICS=off, DO_NOT_TRACK=1, or the global analytics-disabled setting overrides edition preferences. An explicit CARBON_ANALYTICS=on enables analytics unless DO_NOT_TRACK is set.

Failed deliveries are retried on later activity, with at most 200 local event records and seven days of retention, in ~/.config/carbon/usage/. Disabling analytics clears pending events on the next helper invocation. There is no background uploader. HTTP acceptance does not prove dashboard ingestion; retries after uncertain failures may duplicate receiver counts. First activation is once per edition/local configuration, not an install or a unique person. Installs come from publisher dashboards. Command telemetry is best effort: hosts that do not execute the skill's helper cannot be counted.

Your first run with Jay

CARBON opens a local settings page on first use in a supported host. Meet Jay, choose Automatic quick checks or Manual only, and adjust your assessment budget, check count, project focus and analytics preference. Save your choice; closing the page does not change it. Reopen it anytime with /carbon:carbon-settings. Automatic checks use the shared, bounded background supervisor; a global pause or environment disable still wins. This choice does not enable AI review, add executable test commands or authorize fixes.

Grok Bot 1.0.3 prerelease

CARBON Core for Grok Bot 1.0.3 uses saved HTML reports delivered through Grok file attachments. Local package and synthetic-state checks passed; native verification of this version is pending. Download the prerelease.

Antigravity prerelease

CARBON Core for Antigravity 1.32.22-mcp.2 includes a bounded native global MCP logic-flow smoke test and desktop/mobile saved-report rendering checks. Strict agy validation remains unverified and workspace registration unresolved; no full-browser or marketplace approval claim. Download the prerelease.

Grok Bot 1.0.4 prerelease

CARBON Core for Grok Bot 1.0.4 clarifies how to open the actual attached HTML report using Preview. Local copy/delivery regressions passed; native results belong to the 1.0.3 baseline, not a fresh 1.0.4 run. Download the prerelease.