Built for your workflow
One analysis.
Browser, API or agent.
Send items and criteria. Get typed answers, original rows, supporting passages and calculations your code can use.
Your coding agent, our analysis engine
Give your agent the brief.
Copy this prompt into your coding agent. It can prepare criteria, call the API and bring structured results back into your workflow.
An operator-issued key is required. How access works →
Read https://tenzetta.com/agent.md and help me analyze supplied data with Tenzetta. Use POST https://tenzetta.com/api/v1/analyze, or the analyze_items tool at https://tenzetta.com/api/mcp if connected. Ask me which data to use and what decision I want to make. Send preview: true first. Show me the proposed questions, rubric and limitations before executing the reviewed fields and summaries. Preserve original IDs, source evidence, unknowns and failed rows. API/MCP access requires an API key. Read TENZETTA_API_KEY from my local environment; never print it, embed it in source, or place it in a URL. If it is missing, tell me access is required and stop before calling the service.
1. Prepare your answer columns
Example input: a real public document, evaluated when you run the request. Save this request as analysis.json.
{
"items": [
{
"id": "iana",
"data": {},
"url": "https://www.iana.org/help/example-domains"
}
],
"questions": "Does this page say example domains can be used in documentation?",
"preview": true
}curl "$TENZETTA_URL/api/v1/analyze" \ -H "Authorization: Bearer $TENZETTA_API_KEY" \ -H "Content-Type: application/json" \ --data-binary @analysis.json
Set TENZETTA_URL to this site’s origin. This experimental service uses an operator-issued key; self-service customer keys are not available yet.
2. Run the reviewed columns
The preview returns fields, summaries, items and limitations. Copy the returned fields and summaries into your request and set preview: false. The browser’s “Use via API” button downloads this exact request for you.
Start with a goal instead of writing questions
Optionally pass brief with a use_case, selected choices, additional context and preferred output (table, ranked or summary). The browser downloads the exact request after you choose the goal and options. Sales, vendors, feedback, companies, documents and spreadsheet calculations are supported.
For sales, provide selected ICP choices, your own ICP in context, or your own company_url. The URL describes your business; it is never evidence about a prospect. The planner reads that URL directly and returns a proposed brief.business_context plus self-contained questions for review. Newly planned guided requests always return a preview with review_required: true, even if preview was omitted. Review the returned brief, fields and summaries, then send them back to run. Numeric spreadsheet briefs can specify exact measure_column and group_column names.
Bring your company context
Send { "domain": "company.com" } to POST /api/v1/context/company, or call prepare_company_context over MCP. The same API-key authentication applies. We read up to five relevant pages and return a sourced profile with review_required: true. Show the offering, described customers and suggested target customers to the user for correction. After review, pass it as brief.company_profile when previewing questions.
The accepted profile is reused without reading the site again. Explicit choices and corrections take priority; seller context never proves facts about prospects. Changing context requires reviewing new questions. Save criteria includes the profile and source excerpts. Profiles are kept in the current page, with no account storage or inbox connection.
You can omit preview for a single call. If planning finds unsupported parts, the response has status: needs_review and does not run the analysis. Review those limitations before resubmitting explicit fields.
3. Read results, including failures
Each row includes its original id and data, an answers object, status and warnings. Each answer has value, label, confidence, evidence and note. Unknown values are null. A failed row remains in its original position.
For incremental results, send Accept: application/x-ndjson. Events are plan, row and complete; row events include the original index. A broken connection is incomplete, even if some rows finished.
Ask questions across the dataset
Ask “How many match our criteria?” or “What is the total Revenue for qualified rows, grouped by Region?” The planner prepares any per-row questions plus calculations. The analysis engine evaluates the rows; code calculates counts, percentages, sums, averages, minima and maxima.
Each requested summary has operation, a numeric answer field_key (null for counts/percentages), optional where: { field_key, equals }, and optional group_by naming an original column. Pure input-row counts can use an empty fields array. Numeric math requires a number or rating answer field; number fields name the exact source_column.
Results include summaries[].groups[] with the value, unit, input/known/contributing/unknown/failed row counts and contributing row_ids. Percentages use known answers only, with the denominator shown. Partial results are labeled; mixed units are not combined. The browser can show contributing rows and download a separate summary CSV.
Input shapes
| Input | Item field | Behavior |
|---|---|---|
| Spreadsheet row / JSON | data: { column: "value" } | All strings and duplicate rows preserved. Unique item IDs required. |
| Text | text | One document per item. |
| URL | url | Reads that public URL. No wider search. |
| File | POST /api/v1/analyze/file | Multipart file → parsed items. CSV/TSV rows or readable PDF/text document. Pass the items into analyze. |
MCP access
One tool. The whole analysis workflow.
Connect through MCP.
Add a remote MCP server in your client, select Streamable HTTP and supply your key as an authorization header.
An operator-issued key is required. How access works →
Transport: Streamable HTTP URL: https://tenzetta.com/api/mcp Authorization: Bearer <your API key> Tool: analyze_items Guide: https://tenzetta.com/agent.md
Set Authorization: Bearer … using your client’s secret storage. Client configuration formats vary; these are connection details, not a universal JSON config.
The analyze_items MCP tool at /api/mcp takes the same items, questions, reference, fields, summaries and preview parameters. Use this tool for the supplied-data workflow.
Access & current lab limits
Browser analysis is available in this lab release. API and MCP use an operator-issued shared key for approved internal use. Self-service customer keys, tenant accounts and per-customer quotas are not available yet. Never put a key in browser code, a public repository or a URL.
500 items, twelve answer columns and twelve calculations, 4 MB per upload and 1 MB of extracted data per request. Documents: 60,000 characters; reference: 12,000 characters. Long documents use relevant excerpts and report this limitation. Scanned PDFs require text extraction elsewhere first.
Data and results are not saved by this workflow. Processing uses external AI services; this is not a guarantee about provider retention. Download results before leaving. The server stops starting new rows after four minutes or a judgment-service failure. There is no durable background resume in this experimental flow. Repeating a completed request runs it again.
Numeric columns are copied and calculations use code. Arbitrary numeric extraction, joins and date arithmetic are not supported. Numbers use decimal points and optional thousands commas; ambiguous formats are Unknown. Ratings describe the visible rubric. Confidence is not the probability of a sale or other business outcome. Spreadsheet exports escape formula-like cell values as text.