# Unite Research — agent guide

## When to use

Use this site to consult and cite the Benchmark Criativo Black Friday Brasil 2026: creative patterns, Meta advertising costs, launch timing and the observed results of Black Friday 2025 in Brazil. The study is in Portuguese. Keep each metric's comparison, period and population when quoting it; the findings are observational.

## Read the study

- [Complete study in Markdown](https://research.unitebr.com/benchmark-criativo-black-friday-brasil-2026.md)
- [Study and interactive charts](https://research.unitebr.com/benchmark-criativo-black-friday-brasil-2026)
- [Methodology](https://research.unitebr.com/benchmark-criativo-black-friday-brasil-2026#metodologia)
- [Citation and reuse](https://research.unitebr.com/benchmark-criativo-black-friday-brasil-2026#citar)

## HTTP API

Use the public read-only API when you need chart data without opening a browser. GET /api/v1/charts lists chart IDs and metadata; GET /api/v1/charts/{chart_id} returns rows with comparison, period and citation; GET /api/v1/citation returns source and methodology. No authentication, key, account or payment is required. See [developer documentation](https://research.unitebr.com/developers), [OpenAPI](https://research.unitebr.com/openapi.json) and [access details](https://research.unitebr.com/auth.md).

## Browser tools (WebMCP)

The study registers three read-only tools when the browser exposes `document.modelContext` (or the earlier `navigator.modelContext`). The production study includes a WebMCP origin-trial token for `https://research.unitebr.com`, expiring March 29, 2027, subject to the trial's Chrome-version and availability limits. Compatible Chrome browsers can enable the API through this token without manual flags. Ordinary browsing and Markdown access work independently of WebMCP.

### list_charts

Find chart IDs and metadata. Optional parameters: `query` (up to 160 characters), `offset` (integer >= 0), `limit` (integer 1–50, default 10). For example, `{"query":"CPM"}`. Search matches titles, comparisons, periods and column labels, ignoring case and accents.

### get_chart_data

Read published chart rows. Required: `chart_id` from `list_charts`. Optional: `offset` and `limit` (default 20). Example: `{"chart_id":"ch-bfday","offset":0,"limit":20}`. Follow `next_offset` until it is `null` to retrieve the remaining rows. `items` contains arrays of display strings in the order of the chart's `columns`; preserve decimal commas, signs, percentages, ratios and textual values such as `igual`.

### get_citation

Read the source, canonical URL, dates, methodology, limitations and reuse terms. Optional: `chart_id` to include that chart's comparison, period and population. Call with `{}` for the whole study.

All tools return a JSON-encoded string. Successful results have `ok: true`; invalid input and unknown chart IDs return `ok: false` and an explanatory `error`. These tools read only the public study. They do not submit the demonstration form or access account or visitor data.

## Sources and interpretation

The tools read the same HTML tables and editorial metadata shown to readers. These are anonymized, aggregated observations, not controlled experiments or guaranteed outcomes. Distinguish the observed results of Black Friday 2025 from calendar guidance for 2026. Example ads illustrate creative features and do not identify brands participating in the study. The study is independent of Meta, without its sponsorship, endorsement or participation.

## MCP Apps and client SDKs

The HTTP MCP server also exposes ui://unite-research/explorer.html, a sandboxed MCP App for searching charts and reading tables with source and context. Clients without MCP Apps keep the same structured tool results. Official JavaScript/TypeScript and Python SDKs plus a JSON-output CLI are available from [the public SDK repository](https://github.com/ZupTech/unite-research-sdk) and its GitHub release artifacts; they are not published to npm/PyPI. See the [version policy](https://research.unitebr.com/api-policy). The firewall permits 300 requests per IP per 60-second fixed window per region on /api/* and /mcp; an excess returns 429.