Test with intent. Report with evidence.

Start here

Ask CARBON where to start.

Open the project folder in your coding agent so CARBON can see the relevant code, requirements, tests, results, screenshots, and documentation. CARBON Help recommends the best first workflow without changing anything.

$carbon-help

Claude Code:  /carbon-help
ChatGPT/MCP:  /carbon_help
Natural language: Use CARBON Help to assess this project and tell me what to run next

Useful first requests

• $carbon-help — read-only project analysis and a prioritized CARBON starting plan.
• $carbon test this in quick mode — risk-selected local validation of the current change.
• $carbon-browser test checkout with depth exploration — exercise a reachable browser flow.
• $carbon-generate all widgets and checkout flows — generate planned coverage without executing it.
• $carbon-tests add "A signed-out visitor cannot reach billing" — add a prompt-defined intent test.
• $carbon-human — surface only the decision-changing human questions before a confidence or release call.
• $carbon-review review these findings and results — independently challenge the evidence.

Harness map

One runtime across coding agents and execution adapters.

Claude, Codex, Cursor, and other MCP clients route test intent into the same CARBON runtime. CARBON keeps local execution as the default and separates optional testers.ai Premium Cloud capabilities behind an explicit path.

Before execution

See what CARBON will run—and how long it should take.

For an executable local run, CARBON first prepares the test plan and opens a protected link on your computer. The page shows the target, selected breadth or depth, exact tests, priorities, adapters, approval gates, stop conditions, and the initial estimated duration before execution begins.

CARBON Browser v1

CARBON owns the semantics. Adapters operate the browser.

carbon-browser/v1 is the stable boundary between CARBON test intent and browser execution. CARBON now includes an executable bridge to the official Vibium CLI and existing Vibium tests. Other host-agent, Playwright/CDP, Selenium/WebDriver, and managed-device integrations use the same capability contract when their execution bridges are available.

bin/carbon-runtime browser contract
bin/carbon-runtime browser negotiate --manifest adapter.json \
  --require-operation action.click --require-evidence screenshot
bin/carbon-runtime browser init --root . --manifest adapter.json --target https://your-site.example
bin/carbon-runtime browser record --session-dir .carbon/browser/sessions/<id> --result result.json

Executable compatibility

Run native Vibium tests, then use CARBON's higher-level APIs.

CARBON calls the official installed vibium binary and preserves native arguments. It can also run existing JavaScript or Python Vibium tests, or any explicit no-shell project runner. Every invocation records its exit status, redacted output, adapter version, artifact handles, and unavailable evidence channels in a CARBON Browser session.

bin/carbon-runtime vibium doctor --root .
bin/carbon-runtime vibium exec --root . -- go https://your-site.example
bin/carbon-runtime vibium exec --root . -- map
bin/carbon-runtime vibium run-test --root . --file tests/login.vibium.mjs
bin/carbon-runtime vibium generate --root . --target https://your-site.example \
  --vibium-map artifacts/page.map.txt --request "Generate boundary and recovery tests"

Hosts

One workflow, different command syntax.

Specialized workflows follow the same mapping: /carbon_tests in ChatGPT or MCP, $carbon-tests in Codex, and /carbon-tests in Claude Code.

Core workflows

Product testing commands.

AI systems

Confidence engineering workflows.

CARBON operationalizes uncertainty, oracles, judges, statistical design, AI systems, generated software, bias, security, operations, and specialized systems. Use these workflows for prompts, models, RAG, ranking, agents, multimodal features, and AI-generated changes.

Human decision context

Give the harness the judgment code cannot supply.

CARBON can inspect source, execute authorized checks, and qualify the evidence it captures. It cannot silently decide business intent, risk appetite, taste, accountability, or whether a technically correct experience solves a customer’s real problem. Near the end of a broad run, CARBON can therefore add an optional Human step. It does not block technical confidence; it makes the missing human context visible.

What it may ask

• Which customer outcome matters, and is this a demo, staging milestone, or production release?
• What failure is unacceptable, who absorbs the harm, and what recovery is actually available?
• Which domain rule, customer workflow, or credibility expectation would be obvious to an expert but invisible in the repository?
• What should an AI agent never decide or do without a named human owner?

$carbon-human
$carbon-human answer HUM-003: Production launch; incorrect billing totals must hold release.

Decision context: .carbon/human/decision-context.json

Private feedback loop

Put a human perspective directly back into the build.

$carbon-feedback creates a scoped feedback request around a URL, image, document, API, code surface, or any other artifact. It opens a polished local workspace beside the coding agent so a developer can quickly score, categorize, explain, and annotate what they see. The resulting feedback is durable local context for CARBON’s next plan, confidence interpretation, or focused test pass.

$carbon-feedback review this checkout page for first-time-customer trust and clarity
$carbon-feedback ask Priya to review this staging URL for whether the primary value is clear

Execution

Modes and run sizes.

Quick, Medium and Large

• Quick: smallest decision-relevant set; CARBON chooses breadth or depth from risk and normally uses two user perspectives.
• Medium: balanced coverage with four perspectives and adjacent regression checks.
• Large: expands the authorized surface, repetitions and six perspectives while retaining explicit cut lines.

Coverage map

See the complete test surface.

CARBON combines functional, data, experience, security, performance, exploratory, AI-system, and evidence coverage. The map below shows how those checks connect to confidence, remaining risk, coverage gaps, and fix-ready findings.

Truth model

Evidence levels prevent false confidence.

What every result report says

CARBON reports confidence in quality for the selected, executed scope; remaining risk; conclusive and evidence-backed coverage; coverage missing; and the next evidence that would change confidence. Planned, blocked, skipped, canceled, unsupported and excluded checks remain visible instead of being blended into a reassuring score.

Surfaces

Web, API and cross-surface testing.

Every tested page

CARBON records the richest safely available bundle: screenshot, URL/title/viewport, DOM and accessibility state, console and page errors, bounded redacted network traffic, performance, synthetic data, storage-safe metadata, traces, recordings and video. Unavailable channels remain explicit gaps.

First-class API coverage

CARBON can learn from OpenAPI documents and authorized redacted traffic, then generate positive, negative, boundary, wrong-type, auth, tenant, idempotency, retry and resilience cases. It maps API failures back to the UI flows that depend on them.

$carbon map the OpenAPI schema and test the changed checkout endpoints
$carbon-browser test checkout and connect network failures to visible UI states

Continuous confidence

Incremental and scheduled testing.

The framework maps recent code changes to affected components, APIs, flows and adjacent regressions. It can prepare Quick, Medium or Large execution based on assessed risk, install a GitHub Actions workflow after approval, and maintain build history.

• Pull request: narrow, change-focused validation.
• Manual: explicit user-selected scope and size.
• Nightly: broader regression and learning coverage.
• carbon_auto: bounded local test-fix-retest loop with approval before mutation.
• testers_schedule: recurring premium cloud execution tied to an owned report.

Project state

Durable local artifacts.

.carbon/
├── framework/     product model, coverage, regression suite, build history
├── runs/          plans, results, events, findings, evidence and reports
├── learning/      unfamiliar-surface observations and candidates
├── recordings/    redacted browser steps, assertions and playback outcomes
├── tests/         prompt-defined and converted AI-intent tests
│   └── bake-runs/ attempts, repairs, evidence and responsive import reports
└── auto/          bounded loop state and stop reasons

.confidence/
├── risks/         material risks and affected populations
├── claims/        falsifiable behavioral claims
├── evals/         evaluation definitions and results
├── traces/        prompt, model, retrieval, tool and decision evidence
└── releases/      scoped ship/canary/hold/rollback decisions

Optional premium

testers.ai Premium Cloud.

Local carbon workflows use the coding agent and project context. Commands beginning with testers contact testers.ai Premium Cloud and never run automatically from a local workflow.

An email enables the current daily targeted UX pass and weekly two-page sampler routes. Account history, saved cloud tests, schedules and paid execution require a securely stored testers.ai API key.

Authority

CARBON does not take release authority.

• No destructive, paid, public, privacy-invasive, privilege, payment or notification action without explicit scope and approval.
• No credentials in chat. Test passwords and premium keys are stored through the protected settings UI.
• Global prompts and context documents apply across local projects; context files stay local.
• No silent switch from local execution to premium cloud.
• Anonymous product analytics contain only allowlisted major-action names and one first-activation event—never prompts, URLs, credentials, paths, test data, screenshots, logs, or report contents. Disable them in carbon_settings or set CARBON_ANALYTICS=off before startup.
• No generated case, persona opinion or screenshot guess reported as an executed pass.
• Code fixes are proposed first; mutation requires scoped approval.
• Regression promotion and release decisions remain human-controlled.