Founder and manufacturer interviews

Interviews preserve original answers, extract supported facts and ask a focused follow-up. Founder interviews produce production briefs. Manufacturer interviews produce self-reported capability records.

Use your existing agent

Hosted MCP 0.2.14 provides start_interview, recover_interview, get_interview and answer_interview for sandbox or live identities. Ask your connected agent:

Start a founder interview for this project. Ask me the returned question, save my actual answers, keep current samples separate from forecasts, and show the extracted facts and supporting quotes for review.

Keep the returned interviewId and current version with the project. answer_interview takes those values plus a message of at most 4,000 characters. Messages go to Maglev and may be processed by its model; this is not a local-only tool. MCP collection does not confirm or publish a record. The separate CLI confirmation flow below remains an explicit reviewed step.

list_manufacturing_stages reads stage inputs, outputs and availability. Use list_action_kinds for executable action names; a lifecycle stage is not itself an API operation.

Start and answer

Generate a UUID once (for example, node -e "console.log(require('node:crypto').randomUUID())") and save it with the project before starting. Replace REQUEST_UUID below with that saved value. Keep the same UUID for recovery; do not generate a new one just because the connection was interrupted. MCP start_interview requires it as requestId.

maglev interviews start --kind founder --request-id REQUEST_UUID --json

The returned interview includes id, messages, extraction and version. Ask the person the assistant’s question. Save their actual answer in answer.txt:

maglev interviews answer INTERVIEW_ID --file answer.txt --version 0 --json

Use the latest returned version for each turn. The response includes missing, engine and the updated interview. Do not fabricate answers for a person or turn a founder’s requirement into a supplier capability.

Recover creation

maglev interviews recover --request-id REQUEST_UUID --json

Or use MCP recover_interview with the same requestId. Both share the private local receipt in MAGLEV_CONFIG_DIR (default ~/.maglev).

State	Next step
saved	Use its interviewId with get_interview. Repeating the same start reads this original interview.
pending	Creation may have reached the server, but its ID was not saved locally. Reconcile with Maglev before another start; automatic replay is refused.
not_found	Check the original machine, configuration and saved project IDs. This does not prove that no server interview exists.
recovery_locked	Another process is using this request. Wait for it; after a crash the lock expires after one minute. The pending record remains.

Acknowledgments are saved before the tool returns. Two processes using the same request cannot both create an interview. The record contains IDs, kind, hashes and a timestamp; it does not store your key, questions or answers. Changing the input or API key refuses reuse. If a key was rotated, an authorized new key can still read a known server interview ID directly.

Keep the original configuration directory. Containers need a persistent private volume. This is local recovery, not server idempotency or cross-machine recovery. Deleting state or generating replacement IDs defeats this protection. The CLI keeps its older untracked behavior when --request-id is omitted; use the flag for agent workflows.

If the server returned an ID but saving the local receipt failed, interview_receipt_unsaved includes interviewId. Save that ID and read it directly. Do not restart creation.

Review and confirm

Show the person the extracted facts and exact supporting quotes. Unknown values remain null or empty lists. Once they confirm:

maglev interviews confirm INTERVIEW_ID --version 1 --confirm --json

For a founder, the response contains brief, discoverInput and a note. For a manufacturer, it contains a capability record. Its data holds the facts; a discovery result later presents those fields directly.

Sandbox manufacturer records are visible only under the key that created them. Live records initially stay within the organization as self-reported. Maglev review is required before they appear in public discovery.

What readiness means today

The current founder brief captures company, process, materials, quantity and description, with additional supported fields when stated. It is not a complete engineering or production-readiness assessment.

A full readiness review also considers file revisions, critical quality requirements, BOM alternatives, tests, target markets, budgets, timing and sharing permissions. Use existing files to identify gaps before asking the founder. Use the readiness workflow to turn those files and answers into a review. See the lifecycle guide for the larger workflow.

Conflicts and recovery

• HTTP 409: reload with maglev interviews get INTERVIEW_ID. Another answer or publication may have changed the record. Review before submitting against its latest version.
• HTTP 422 with missing_facts: collect the named fields before confirming.
• At 16 saved turns, review the record. A published manufacturer capability requires a new interview to revise it.
• engine: guided means deterministic extraction is being used instead of the language-model interviewer. Follow its labeled format with only facts the person supplied.
• A timeout does not prove that nothing was saved. Reload using get_interview or the CLI before deciding whether the answer needs another write. If the answer is already present, continue from that saved version.
• Do not automatically refresh the version and replay after a conflict; another answer may have changed the meaning of the next turn.
• The server create endpoint is not idempotent, and there is no list-interviews API. Local receipts protect same-machine replay after an acknowledgment is saved. If the response is lost before an ID arrives, the receipt stays pending; stop automatic retries and report the uncertain outcome.

Data handling

Messages and extracted records are stored on Maglev servers. In AI mode, the transcript is sent for AI processing. Keep confidential production files and unrelated sensitive details out of the public sandbox. This API is text-based; it does not place phone calls or clone voices.