MARGINNOOK / DEVELOPER PREVIEW
Formats, limits and retention
Exact boundaries for the local contract and current public preview.
Developer preview. Hosted API is not live yet. Explore fixed synthetic samples; anonymous uploads, hosted execution and payments remain closed.
Supported boundaries
Maximum resource: 2 MiB. Six formats: CSV, JSON rows, text, HTML, ordinary DOCX and JUnit XML. No PDF, OCR, URL fetching, arbitrary SQL or code execution.
Initial anonymous allowance in the local contract: 50 credits per rolling 24 hours and 1,000 per rolling 30 days. Aggregation and JUnit extraction cost 1 credit; document search costs 2. Reading existing evidence is free. Hosted checkout is disabled; US$5 / 10,000 credits is a candidate price, not an offer.
Originals, derived text and result bodies share the original upload’s 24-hour deadline. Redacted job metadata and metering remain up to 31 days. Deletion revokes the entire resource family. Provider point-in-time history in the private qualification environment may remain up to 30 days; downloaded copies cannot be recalled.
Operation bounds
Aggregation: up to 10,000 rows; limit 1–1,000, default 100; precision up to 38 decimal digits. Document search: limit 1–10, default 5; query at most 512 bytes and 24 content keywords. JUnit: max_cases 1–100, default 10. All pages use offset; the core result bound is 128 KiB. Inline HTTP output has a separate 8,000-character boundary.
When a boundary is reached
Malformed input fails explicitly; completed failed jobs charge zero credits. Output may be paginated or fail OUTPUT_TOO_LARGE. A quote does not parse input or guarantee success. 429 free-quota and 503 platform-capacity errors require different recovery; neither opens a checkout.
Questions before you integrate
Do I need an account?
Reading these samples and docs needs no account. Local first use creates an anonymous credential automatically. Public anonymous API access is not open.
Is this semantic search? Can it read PDF or fetch URLs?
Document search is literal any-term FTS5 keyword retrieval. It does not prove that an answer is absent. PDF, OCR and URL fetching are unsupported.
What do incomplete results and pagination mean?
Read scope, incomplete, next_offset, omitted_items and source_read_required. A downloadable returned JSON can still contain only one bounded page. No matching paragraph is not a complete document conclusion.
How do I read evidence and retry?
Use the original or normalized source resource, or MCP read_evidence with offset and length. Keep private handles outside public logs. Reads are free; a new computation page is a new job. Retry a submission with its original key, retain the job ID after timeout, and never automatically repeat an uncertain upload.
Does this prove token savings?
No complete model task usage has been measured. Smaller outputs can help a workflow, but tool definitions, requests, results, retries and answer quality must all be measured before claiming savings.