> ## Documentation Index > Fetch the complete documentation index at: https://docs.getmillwork.dev/llms.txt > Use this file to discover all available pages before exploring further. # Check your preview access with Echo > Check your access without paying for a model run, then keep the Echo receipt. **Goal:** Check your organization's access with **Echo**, a test run. Echo does not call an external provider or spend provider credit. This starter check has no platform fee. Live model runs have separate [fees and model costs](/concepts/access). **You are done with this diagnostic when:** the application state is `echo_proved`, `next_action.type` is `done`, and you can retrieve the receipt for `echo_execution_id`. A receipt is a record of the run without its prompt or output. This is not completed tenant setup and must not be reported as `ready`. Full setup still needs a bounded live model result, provenance receipt, remaining-credit view, and a supported public CLI version. Echo does not prove that a model can answer a paid request. ## Before you begin * Your organization is approved for the private preview. * An organization owner invited you and granted `runs:view`, `runs:write`, and `arms:manage`. * You have an organization API key in an approved secret store. * You have Node.js 20 or 22. * If you use the CLI, Millwork has explicitly provided a qualified preview artifact and its exact version. There is no supported public CLI version yet. Do not install a package or copy a version from a registry listing. If your invitation does not name an exact version, stop and contact [support](mailto:support@getmillwork.dev). Run these local checks only from a qualified preview artifact that Millwork explicitly provided. They do not make an API request, open a browser, or read the value of your API key. ```bash theme={null} millwork --version --json millwork docs --json millwork doctor --json ``` `millwork --version --json` returns the local package version and the public support state. `supported_public_version` stays `null` until a published CLI passes the release gate. ```json theme={null} { "schema_version": 1, "package_version": "", "supported_public_version": null, "public_cli_available": false } ``` `millwork docs --json` returns this page with a privacy-safe referral tag. `millwork doctor --json` reports only whether a key is configured. It never prints the key. Set `SOLVERAPI_API_KEY` through your approved secret store. Do not paste the value into a command, chat, ticket, or log. The default API base is `https://api.getmillwork.dev/v1`. Choose `starter` explicitly for this Echo-only check. The CLI's default template is a different flow that can ask you to approve paid model work. In an interactive terminal, run: ```bash theme={null} millwork tenant start --template starter ``` For an agent or other non-interactive session, planning and applying must be two separate commands: ```bash theme={null} millwork tenant start --plan --json --template starter > millwork-plan.json ``` Planning does not create a setup application, a provider resource, or paid work. A setup application records the steps you approve so you can return to the same progress. The `>` in the command saves the plan to a local `millwork-plan.json` file; Millwork does not create that file itself. Check that `template_id` is `starter`, `maximum_spend_usd` is `0`, and the only effects are creating or reusing a saved model option (an **arm**) and submitting Echo. The plan's `digest` identifies the exact plan you approve; `expires_at` is its expiry time. The plan includes these fields: ```json theme={null} { "template_id": "starter", "template_version": "1", "request_preset_id": "starter-public-sandbox-v1", "issued_at": "", "expires_at": "", "digest": "", "qualification": { "state": "approved", "next_action": { "type": "approve_plan", "detail": "" } }, "catalog_row": null, "blockers": [], "effects": [ { "id": "register_or_reuse_starter_arm", "description": "Create or reuse one template-managed model arm. No provider resource is created." }, { "id": "submit_echo", "description": "Submit one Echo execution with routing.required_arm_id and no verifier." } ], "starter_credit": { "balance_usd": 0, "funded_state": "sandbox" }, "maximum_spend_usd": 0, "file_manifest": [ { "path": ".millwork/starter.json", "kind": "secretless_config" }, { "path": ".env.example", "kind": "env_placeholders" }, { "path": "examples/starter-echo.ts", "kind": "node_ts_example" } ] } ``` The starter does not add a verifier, a check applied to a run. It does not connect a provider account or approve live model work. In an interactive session, approve only the starter plan you reviewed. For the separate-command flow, use the command below. Use the `digest` and `issued_at` from the same unexpired plan. The API requires one `Idempotency-Key`, which identifies the application across retries. The current preview CLI does not expose a supported public resume-key input. If an apply loses its response before it returns an `application_id`, stop and contact support instead of starting another apply. ```bash theme={null} millwork tenant start \ --digest "" \ --issued-at "" \ --template starter ``` Without `--write`, the application creates or reuses the template-managed arm and submits one Echo run. It does not write starter files. Add `--write` only after you approve the plan's `file_manifest`. Existing files are never overwritten. Keep the `application_id` so you can inspect the same setup attempt. In the application response, look for `echo_proved`, `done`, and the Echo run ID: ```json theme={null} { "application_id": "", "template_id": "starter", "template_version": "1", "state": "echo_proved", "completed_effects": [ { "id": "register_or_reuse_starter_arm", "at": "" }, { "id": "submit_echo", "at": "" } ], "result": null, "receipt": { "receipt_id": "" }, "managed_arm_id": "", "echo_execution_id": "", "next_action": { "type": "done", "detail": "Echo receipt is retrievable at GET /v1/receipts/:id." }, "diagnostics": { "provider_resource_created": false, "files_written": false, "plan_digest": "" }, "created_at": "" } ``` After the CLI returns an application ID, inspect the same application with `GET /v1/tenant-template-applications/{applicationId}`. Resume the recorded `application_id`; do not start a second application to recover state. Fetch the receipt with `GET /v1/receipts/{echo_execution_id}`. Echo returns no model output: `result` stays `null`. ## Resolve an access result Stop on any state other than `approved`. Follow its typed next action. Do not guess a tenant, invitation, role, or access decision. | State | Next action type | What to do | | ------------------------------ | ------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------------- | | `no_credential` | `apply_for_preview` | Sign in through the browser preview application. The application creates no tenant, key, credit, configuration, or run. | | `preview_pending_or_rejected` | `wait_for_preview_decision` or `contact_support` | Wait when pending. Contact support when rejected. | | `tenant_not_admitted` | `contact_tenant_admin` | Ask the tenant administrator to confirm private-preview admission. | | `organization_invite_required` | `ask_org_admin_to_invite` | Ask an organization administrator to invite you. The response does not confirm that an organization exists. | | `role_lacks_permission` | `ask_owner_for_role` | Ask an owner for `runs:write` and `arms:manage`. This is not a preview-admission denial. | API failures use `application/problem+json`. The shape is: ```json theme={null} { "type": "", "title": "", "status": 400, "detail": "", "instance": "" } ``` Match the final path segment of `type`. Do not parse `title` or `detail` as a stable error code. ## Recover safely * **The digest expired or changed:** request a new plan and approve its new digest. * **The idempotency key conflicts:** use the original plan digest with that key. Do not reuse the key for another plan. * **The apply response was lost:** do not guess a key or start over. Use the returned `application_id` when available. Otherwise stop and contact support. * **The session is non-interactive:** run `--plan --json`, review the result, then use a separate command with `--digest` and `--issued-at`. * **A local file already exists:** inspect it. The CLI skips it instead of overwriting it. * **An API call fails:** stop. Record the HTTP status, problem `type`, `title`, `instance`, and a short redacted excerpt. Do not include request bodies, headers, keys, prompts, or results. Understand what the receipt proves and what Echo does not test. Prepare a redacted support report without sharing a secret.