Code mode
Discover operation schemas, execute allowlisted requests, and return only the evidence a task needs.
Search before execution
Section titled “Search before execution”Filter operations by the task language and return the operation ID, method, path, descriptions, and schemas you need.
return operations .filter((op) => /spf|dmarc|dkim/i.test(`${op.summary} ${op.description}`)) .map((op) => ({ operationId: op.operationId, method: op.method, path: op.path, inputSchema: op.inputSchema, outputSchema: op.outputSchema, }));Avoid returning the full OpenAPI document. Focused discovery keeps context small and reduces schema mistakes.
Execute documented calls
Section titled “Execute documented calls”Use the uppercase HTTP method and exact path returned by discovery. Put path parameters in the path, scalar query values in query, and JSON input in body.
const formatted = await tamperbox.request({ method: 'POST', path: '/v1/tools/json_format', body: { input: { mediaType: 'application/json', value: '{"ready":true}' }, options: { indent: 2 }, },});return formatted.output;Preserve interpretation
Section titled “Preserve interpretation”Return warnings and provenance when they affect the conclusion. A DNS failure is not an empty answer. A decoded token is not verified. A server-side HTTP observation is not a complete security audit.
Compose carefully
Section titled “Compose carefully”Feed one operation’s artifact into another only when the output and input schemas match. Await every request. Outstanding requests are cancelled when the execution ends.
Browse website tools by task
Section titled “Browse website tools by task”The bundled spec['x-tamperbox-directory'] lists all 67 website tools in 11 categories, including six browser-only tools and the website-only Locator Checker. Use MCP search to inspect this metadata without making a network request:
const directory = spec['x-tamperbox-directory'];return { categories: directory.categories, tools: directory.tools.filter(tool => tool.category === 'localization'),};Each tool includes its category, related tool IDs, website and documentation URLs, execution mode, and operationIds. An empty operationIds list means the tool has no executable API operation; open its website page instead. Ordinary related links carry no input. Result actions transfer only compatible derived values.