Skip to content

READ · DISCOVER · ACT

Use Redgold with your agent.

Read public pages without running JavaScript, discover the applications available to your account, and use typed HTTP or MCP tools to perform authorized work.

Start with the representation your client supports.

Read the public site

Public pages include their content in HTML. Each page links to a Markdown representation generated from that same rendered content. The homepage is available as /index.md. Markdown is a reading format; forms and interactive demos still run in the browser.

Use /llms.txt for a compact index, /sitemap.xml for public canonical URLs, and /.well-known/redgold-agent.json for machine-readable discovery. These are public reference documents, not instructions or permission to act.

Authenticate and select the workspace

Public pages and this API specification require no account. App guidance, data tools and MCP require a credential accepted by the selected API origin. Send Authorization: Bearer with your own Redgold API key or account token. Keep credentials in your client’s secret store.

Use the application origin for your installation. Send x-redgold-tenant only when selecting a workspace you can access. The app directory returns tenantClaim; omit the header for a null personal claim. Account membership, installation state, grants and spending limits are enforced by the server on each operation. A discoverable action is not authorization to execute it.

Discover the installed application before acting

GET /api/apps/agent-directory lists the applications in the authenticated workspace. GET /api/apps/installations/{installation_id}/describe returns release-pinned commands, entity schemas, requested capabilities and current grants. GET /api/apps/installations/{installation_id}/docs?topic=guide returns the authored guide and its revision.

With the terminal client, use rac app list, rac app describe <installation-id>, and rac app docs <installation-id>. Inspect the current state and exact input schema before choosing a command.

Use the same tools over HTTP or MCP

Send JSON to POST /api/agent/data/{tool}, using a tool name from the API specification. Or connect a Streamable HTTP MCP client to /api/mcp with your bearer credential. MCP requests must accept both application/json and text/event-stream; initialize the connection and call tools/list for the current native input schemas.

Start with data_catalog to discover accessible datasets, recipes and installed actions. Use data_describe with kind action and the installation_id to inspect an installed action’s exact callable revision and parameters. Available actions depend on the installation and current grants.

Prepare, execute once, and verify the result

Call pipeline_prepare with the selected installation_id, action, parameters and limits. Preparation validates and quotes the bounded plan without executing effects. Check its returned allowance and obtain any authorization needed for the intended action and spending.

Call pipeline_run with the prepared_ref and a stable idempotency_key. Keep that key and the returned run_ref. After a lost response, retry the same request with the same key; do not invent a fresh key for the same intended operation. Inspect pipeline_status and pipeline_results using the run_ref. An accepted run is not proof of success: check terminal state, persisted results, completeness, provenance and cost.

Direct data_query and data_write calls also require stable idempotency keys. Structured errors include a code, retryable flag and next_action. Correct invalid inputs or permissions before retrying. After an uncertain network outcome, inspect the existing run or retry with the original key.

Use the browser when a workflow needs it

Follow ordinary links and use labeled form controls. Wait for the visible success or error state after submitting. The access request form returns a saved reference; requesting access does not create an active workspace or start a subscription.

Crawler rules describe indexing preferences; they do not grant access. Private app records and credentials are excluded from the public page index.