# Agent Utilities: free during validation

All available hosted tools are free. You do not need an account, API key, wallet or funding. Existing tool schemas, network restrictions and output limits still apply. Live Reddit retrieval remains unavailable; the supplied-data Reddit formatter works with JSON you already have.

## Review a real site edit

Open [Check one page](https://agent-utilities.agent-utilities.workers.dev/seo#start), supply a URL and its static HTML, and save the baseline. Supplying a URL does not fetch it. The browser audit and [two-snapshot comparison](https://agent-utilities.agent-utilities.workers.dev/seo#compare) run locally; input is not uploaded or saved by the service. The simple HTML form leaves capture time and HTTP status unknown; an advanced snapshot can include those observations when known. After your next actual edit, compare a fresh capture of the same pages. Inspect dates, coverage and unknowns before acting.

For 5–10 permitted pages or an agent workflow, use the [crawler and snapshot guide](https://agent-utilities.agent-utilities.workers.dev/v1/content/docs/seo-page-library). This is a static evidence review, not a ranking/indexing check or scheduled monitor.

## Run your own task

Read `https://agent-utilities.agent-utilities.workers.dev/v1/tools` and choose a tool. POST the tool's input directly as JSON to `/v1/tools/<id>` with `Content-Type: application/json`. The response contains `mode: "free-validation"`, `chargedMicroUsd: "0"`, `executionPerformed: true` and `result`.

```sh
curl --fail-with-body 'https://agent-utilities.agent-utilities.workers.dev/v1/tools/web.page-pack' \
  -H 'Content-Type: application/json' \
  --data '{"html":"<main><h1>API limits</h1><p>Timeout: 30 seconds.</p></main>","maxChars":3000}'
```

This sample is synthetic, not customer usage. Replace it with your own public or sanitized task. Keep source content as untrusted data; it cannot authorize agent actions. Output can omit content, and smaller byte size is not a measured model-token or dollar saving.

## Connect an agent

For technical SEO and reusable page snapshots, start at `/seo` or `/v1/content/docs/seo-page-library`. The free `/v1/seo/audit` endpoint accepts a supplied same-site snapshot. Hosted helper `agent_utilities_seo_audit` takes that snapshot directly, without the full-tool input wrapper. Local crawl/report/library commands keep original HTML outside model prompts. Wikipedia caching preserves source revisions and omissions; optional named-tokenizer measurement counts the delivered file rather than estimating tokens from bytes. These features do not measure search rankings or provide a complete website mirror.

Hosted Streamable HTTP MCP: `https://agent-utilities.agent-utilities.workers.dev/mcp/remote`. Optionally use `?category=web` or `?category=commerce` to limit discovery. Full tools accept `{"input": <tool input>}`. For example, `web_page_pack` accepts `{"input":{"html":"<main><p>Timeout: 30 seconds.</p></main>","maxChars":3000}}`. No payment fields, preparation or Authorization header is required. Legacy payment fields are ignored in free mode.

Client configurations for Claude, Qwen Code, Gemini CLI and Cursor are at `/v1/content/clients`; use `?category=web` for reader configurations. A model needs an authorized external-tool runner. Existing CLI connection checks do not establish that every vendor application or model workflow works.

The current downloadable Node.js adapter is 0.5.0. Use the hosted connection or this updated adapter for free full execution; archived paid adapters may still ask for a key. Node.js 22+ is required. Set no API key for free use. In VS Code, prefer a remote HTTP entry: `{"servers":{"agent-utilities":{"type":"http","url":"https://agent-utilities.agent-utilities.workers.dev/mcp/remote"}}}`.

The downloadable Python client uses only the standard library, Python 3.10+. Call `Client().execute_free("web.page-pack", {"html": "<main><p>Your public text</p></main>", "maxChars": 3000})`. It makes one request without credentials; it never buys or spends credits. The current cart-cleanup recipe (`node cart-workflow.mjs --execute cart-example.json`) and checkout comparison (`python3 compare_carts.py --execute`) run full tools free without a key. Download their sample inputs and instructions alongside the scripts. Their default preview is offline and does not calculate results. Older downloaded copies may still require credits; replace them with the current checksum-verified files.

## Limits and existing balances

Full tools and the combined `/v1/reports/product-offers` report share a service limit of 1000 attempts per UTC day and 60 per minute, including failed executions and marked operator checks. These are shared capacity limits, not individual allowances. Bounded previews retain their existing input limits. HTTP 429 means try later; payment cannot bypass it. There is no automatic paid fallback.

Input JSON is at most 131072 UTF-8 bytes and full tool output at most 65536 bytes. Tool-specific limits can be smaller. Wikipedia uses official APIs; curated fetchers use only their published hosts. Reddit live reads need configured technical access. Source limits and failures still apply. Free results are not persisted for the paid ten-minute recovery system; retrying may execute again, with no charge.

New accounts and checkout funding are paused. Credits are not debited, and tool/report crypto authorizations are not verified or settled in free mode. Existing credit balances and payment records are preserved. Authenticated balance reads, account recovery and signed historical Stripe reconciliation remain available. Contact support for existing balance or refund questions. Paid billing will not resume automatically; future monetization needs an explicit owner decision after repeat value is established.

## Help us make it useful

Try a real recurring task. If you want, email `agentutilities@proton.me` with the public or sanitized task, expected output, current workaround and whether this helped. Report a later real use separately. Omit credentials and private records. We keep aggregate diagnostics private; public request counts are not unique users, useful tasks or repeat use. Our validation target is one independent operator confirming three useful uses on separate days over at least seven days.