pub.dev with an AI agent
The short answer: pub.dev is the package site for Dart and Flutter. We gave Claude Code 2 real tasks there, each in a brand-new session. With Playwright MCP a task used 167K tokens on average. With Dejo it used 230K the first time and 79K on every run after, 53% fewer, because Dejo replays a learned task without the model reading the pages again.
The tasks
- “find the Dart package mustache_recase, open its Changelog tab and tell me what changed in version 2.0.1”
- “find the Dart package mustache_recase, open its Versions tab and tell me the minimum Dart SDK listed for version 2.0.3”
What each tool used
| task | Dejo, first run | Dejo, every run after | Playwright MCP | Playwright CLI |
|---|---|---|---|---|
| 1 | 252K tokens, 23 s | 79K tokens, 1 call, 14 s, right | 164K tokens, 2 calls, 15 s, right | 214K tokens, 3 calls, right |
| 2 | 208K tokens, 17 s | 79K tokens, 1 call, 11 s, right | 170K tokens, 2 calls, 16 s, right | 214K tokens, 3 calls, right |
Calls are browser tool calls the agent made. Right or wrong is checked against the answer read off the live site before any tool ran.
Learned and replayed with Dejo on 13 November 2026
We ran the task again with Dejo on 13 November 2026, so you can see what it saves today. These are the steps Dejo kept, exactly as it reported them. Each one carries a check that it landed, and a replay walks them in plain code.
1. go https://pub.dev/
2. fill ref "e20" (by name)
3. press Enter
4. click link "mustache_recase" (by name)
5. click button "Changelog" (by name)
6. answer (by position)Replay: 6 steps in 2.0 s, with no page reading. Answer: “Removed Generator+ dep” (right).
A new task, once Dejo knows the site
In our site-memory test, the task “find the Dart package mustache_recase, open its Versions tab and tell me the minimum Dart SDK listed for version 2.0.3” ran twice in fresh sessions: once with an empty memory, and once after Dejo had learned a different task on pub.dev. It used 183K tokens the first way (4 turns) and 234K the second (5 turns), right both times.
How we measured
Claude Sonnet 5 at medium effort, a brand-new Claude Code session per task and tool, @playwright/mcp 0.0.80, October 2026. The full method, and the sites that blocked every tool, are on the benchmark page. Raw data: benchmark-real-2026-10-21.json, benchmark-real-2026-10-23.json, benchmark-real-2026-10-23-fixes.json, benchmark-site-memory-2026-11-10.json.
Other websites we measured
- scrapethissite.com: a sandbox site for practising web scraping, with tables, forms and paging
- packages.debian.org: Debian's package directory, searched by name and release
- rfc-editor.org: the official home of the RFC series of internet standards
- catalog.ccv.edu: the course catalog of the Community College of Vermont
- ctan.org: the Comprehensive TeX Archive Network, where LaTeX packages live
- practicesoftwaretesting.com: a demo tool shop built for practising test automation
- demowebshop.tricentis.com: Tricentis's demo web shop for test automation
- collection.internationalfolkart.org: the online collection of the Museum of International Folk Art in Santa Fe
- gutenberg.org: Project Gutenberg, a library of free ebooks
- hackage.haskell.org: the Haskell community's package archive
FAQ
How many tokens does an AI agent use on pub.dev?
In our test, Claude Code with Playwright MCP used 167K tokens per task on pub.dev on average. Dejo used 230K the first time and 79K on every run after, 53% fewer than Playwright MCP once the task was learned.
Can Claude Code automate pub.dev?
Yes. In fresh Claude Code sessions on 2 tasks there, Dejo answered 2 of 2 right, Playwright MCP 2 and Playwright CLI 2. With Dejo, a task done once is replayed later without the model reading the pages again.
Does a new task on pub.dev get cheaper once Dejo knows the site?
In our site-memory test, a new task there used 183K tokens with an empty memory and 234K after Dejo had learned a different task on the same site (4 turns against 5).