WebMCP browser tools cut tokens 52% on one task but increase them on another, DeepDeck experiment finds
j032 · reddit · 2026-09-14
- The maintainer of DeepDeck (MIT-licensed macOS app, independently built on DeepSeek Harness) compared ordinary browser access vs. saved WebMCP-style reusable tools on two fixed tasks, 5 interleaved runs per condition with fresh sessions, both using the openai-codex/gpt-6-astra route.
- Books to Scrape (scan 65 fiction books, return the 3 cheapest): median time 46.8s→29.9s, total tokens 92,559→44,440, tool calls 5→3.
- X structured extraction (3 fixed posts): median time 62.2s→55.7s, but total tokens rose 134,038→145,301 and calls went 7→8.
- The baseline could use JavaScript and DevTools, so it's not a pure screenshot/computer-use comparison. Measurements exclude tool construction/maintenance cost; totalTokens includes cached reads and isn't a dollar estimate.
- Conclusion: reusable tools can reduce extraction work, but the X result is a concrete counterexample to claims that they always reduce tokens. Generated local adapters can break when sites change; compatibility with other WebMCP hosts is untested.
More from coding & agent
- Agent hacks stem from scraping local private files, not alignment failure, researcher argues — ryunuck · 2026-09-14
- Diversity-Aware Skill Routing Uses DPP to Cut Redundant LLM Agent Skill Picks — Wang Wei · 2026-09-14
- Contextual Bandit Algorithms Route Prompts to LLM Experts with Sublinear Regret — Wang Wei · 2026-09-14
- Step-by-Step Guide to Becoming an Inference Optimization Engineer, From Quantization to Speculative Decoding — ashishllm · 2026-09-14
- DeepDeck ships a WebMCP directory with pinned revisions: build inspectable browser tools for sites without APIs — j032 · 2026-09-14
- StepFun Launches Startup Program Offering API Credits and Ecosystem Support for Early AI Teams — StepFun_ai · 2026-09-14