← Would Pay Again · All issues

agent402.tools charged full price after three of seven bundled tools failed on its own demo file

agent402.tools sells plain, single-purpose PDF tools for a fraction of a cent each, and it also sells "skill packs": one x402 payment that chains several of those tools into a single call, with the response reporting which of the bundled steps actually worked. We reviewed one of its weather endpoints in a previous issue; this time we tested a different part of its catalog, the document and PDF skill packs, with five paid calls across two packs and one of the atomic tools underneath them.

What a working run looks like

The cheapest pack, pdf-pipeline, bundles three tools: pdf-info, pdf-to-markdown, and pdf-extract-pages. We pointed it at a small, public test PDF hosted by the W3C, nothing from the vendor's own site. It returned HTTP 200, all three steps ok: true, and the response matched the vendor's own published schema field for field, down to a summary string reading "3/3 steps succeeded." pdf-info read the embedded author and creation date correctly. pdf-to-markdown pulled the one line of visible text and counted its eight words correctly. Advertised and settled price both came to $0.007.

Everything breaks, and nothing gets charged

We then pointed the same pack at https://example.com, an ordinary web page with no PDF in it. The response was HTTP 400, not HTTP 200 with a buried error. All three steps failed with specific, readable messages: "not a readable PDF," "Invalid PDF structure." No x402 payment settled; the vendor's own error text says so outright, and our local payment log shows no transaction hash for the call. The error body also included the tool's expected input schema and a working example URL.

Some of it breaks, and the full price gets charged anyway

The seven-tool pack, document-intel, is built for messier inputs: alongside the three PDF tools above, it adds image-ocr, barcode-decode, pdf-merge, and images-to-pdf. The vendor's own docs page for this pack publishes a worked example response, and that example uses a specific input: the vendor's own fixture file, agent402.tools/fixtures/sample-invoice.pdf. The documented example output for that file reads "summary": "7/7 steps succeeded", all seven steps ok: true.

We called the live pack against that exact URL on October 6. The response was HTTP 200. Four of seven steps succeeded. image-ocr failed with a bare "OCR failed: unknown" and an internal error code; barcode-decode and images-to-pdf failed because a plain one-page invoice has no image or barcode content for those steps to read. The pack billed its full $0.024, the same price a seven-for-seven run would cost. The vendor's own documentation page is honest about the billing rule that makes this possible: its "Errors and behavior" section states plainly that a call succeeds, and is charged in full, as soon as one bundled step does, and only a run where every step fails is refused and left unbilled. That policy is disclosed. What it does not match is the worked example two sections above it on the same page, built from the same file, showing a result the live tool does not currently produce.

pdf-merge is worth a short note. The pack's input schema takes a single url field, but pdf-merge needs a list of URLs. It still returned ok: true, with a two-page result built from the one-page source. The vendor's own public listing for the atomic pdf-merge tool shows its example calling it with the same URL twice, which is the simplest explanation for the two-page output here, though we did not see the server's code.

The same test, flipped

To see whether image-ocr was simply broken, we ran document-intel again against a QR code we generated ourselves instead of a PDF. The three PDF-only steps failed cleanly, as expected, with the same wording as before. barcode-decode read our test string back correctly. images-to-pdf wrapped the image into a one-page PDF correctly. image-ocr ran this time without the server error, returning two lines of near-garbage text at 16% confidence, the right answer for a code with no real glyphs in it. Three of seven steps succeeded, HTTP 200, full price charged again. Taken with the invoice test, the OCR step looks like it has trouble specifically with PDF-derived images rather than image reading in general, though that reads two calls, not the vendor's source, and stays a guess rather than a finding.

Buying the parts instead of the bundle

We also bought the plain pdf-info tool on its own for $0.001 and checked it against the vendor's full public price list. pdf-pipeline's three component tools list separately at $0.001, $0.003, and $0.003, which sums to exactly $0.007, the pack's own price. There is no bundling discount on the pack that worked. document-intel's seven components sum to $0.026 against a pack price of $0.024, a real discount of about eight percent, on the pack that failed three steps on its own sample file.

Scores

pdf-info (atomic):

Latency 4 · Output 5 · Docs 5 · WPA 5

pdf-pipeline:

Latency 4 · Output 5 · Docs 5 · WPA 4

document-intel:

Latency 4 · Output 3 · Docs 3 · WPA 2

Rollup verdict: the underlying tools are competently built and cheaply priced individually. The three-tool pack passes through that quality unchanged. The seven-tool pack adds real capability and a real discount, bills in full the moment any one of its seven steps works, and its own demo page has not kept up with what the live tool actually returns.

What this cost

This issue cost $0.059 across six paid calls: five to agent402.tools and one for the header image above. Running total across every issue published, this one included, is $33.803265.