Apify’s 30-day account of TOON in its Model Context Protocol (MCP) server makes a seemingly narrow change harder to judge: the proposal, release and follow-up tests did not examine the same intervention. The result is a case study in how to test an encoding choice, not a verdict on TOON across MCP. Apify says it ran TOON in production on five tools, then removed it after two live A/B rounds showed no consistent benefit.
The first comparison was a proposal, not a live model evaluation. Its 18-fixture study used 16 captures from Apify’s API and two synthetic cases, measuring UTF-8 bytes. It proposed an adaptive picker: encode each response as JSON and dot-flattened TOON, then return the smaller. Apify’s proposal reported a 6.2% byte reduction across the fixtures under that best-of-two policy. That figure does not establish token or task-quality gains.
Payload shape is part of the story. The proposal says TOON’s tabular encoding is most useful for flat rows with uniform keys; nested or nonuniform fields can lose the advantage. Its sampled Google Maps and scraped-data payloads included nested fields and optional data, so “dataset items” did not represent one stable shape. This is one reason a format result from a narrow fixture set may not travel to other payloads.
The release differed from the proposal: five tools always returned TOON rather than whichever encoding was smaller. Apify’s account says its fixture results for dataset items depended on Actor shape; the picker chose JSON for three of four sampled Actors, where forcing TOON was 2% to 4% larger. The proposed fixture result therefore cannot stand in for the policy that actually shipped. Apify also says only the text representation changed; structuredContent and outputSchema stayed unchanged, so the experiment concerned the text supplied to the model, not a change to the structured data interface.
The live tests asked another question. In round one, Apify used 20 tasks and five runs per format. It reported TOON at 4.8% more tool-result bytes and 0.7% more agent tokens than JSON, with pass rates of 64% for TOON and 67% for JSON. Six tasks never called a TOON-encoded tool. In round two, Apify used 16 storage tasks and three runs per format, added real scraped-data tasks, and changed the get-dataset-items description to encourage field selection. It reported TOON at 1.4% fewer bytes and 1.7% fewer tokens, with pass rates of 92% and 88%, respectively. The signs flipped, but this was not a controlled repeat: both the task mix and tool guidance changed, so the round cannot isolate serialization.
There are further limits. The benchmark record says Claude Haiku 4.5 was the agent, DeepSeek V4 Flash graded answers, and live Actor outputs varied in unseeded runs. Round-one task pass counts ranged from 8 to 18 out of 20. That spread is a reason to read the three-point difference between aggregate pass rates cautiously, not evidence that one encoding improved task success. The raw round-one and round-two files preserve Apify’s run results; they are not independent replications. The rounds should not be averaged into one score.
Output selection is a separate lever. In the proposal’s Google Maps fixture, the full payload was 1.01 MB; full TOON reduced it 14% to 893 KB, while projecting three fields cut it 98.8% to 12.3 KB. Projection plus the adaptive picker reached 9.5 KB. Those are fixture results, not a production A/B of field selection. Apify also reports that dataset items made up 27.5% of all tool calls and 99.8% of calls to the five changed tools; the other four together were 0.07% of calls. Those are call counts, not bytes or token share.
After removing TOON, Apify says the text serialization returned to JSON while structuredContent and outputSchema remained unchanged. It also kept a 20-item default for dataset items, with field selection or pagination guidance for large results. For builders, the bounded lesson is to test the exact encoding policy that will ship against representative payloads and call patterns, and to evaluate projection and item limits as separate changes. To isolate a format effect, hold the tasks and tool instructions steady. Apify’s results show why those distinctions matter; they do not predict what another server or workload will see.
