We're moving beyond 0.0.x and into 0.1.0!
This release brings major improvements across scraping, crawling, AI-powered data extraction, document processing, and MCP. We've also rebuilt parts of the experience to make Maxun more capable, reliable, and useful as an open-source web data platform.
Here's what's new.
⚡ Faster scraping, better results
Scraping is now 4× faster, dropping from 2.5 seconds to 0.63 seconds in our local benchmark covering text, links, Markdown, HTML, and screenshots.
We've also improved the quality of extracted content across 91 real-world pages:
- 26% higher precision - less irrelevant content in the output.
- 12% higher F1 score - a better balance between precision and completeness.
- 20% smaller output - less data to process downstream.
- 99% recall maintained - virtually all expected content is still captured, with no page losing recall in our benchmark.
The result: faster scraping, cleaner output, and less unnecessary data.
Content extraction is more precise, output is smaller, and no page lost recall in the benchmark.
🤖 A new AI Mode
AI Mode has been rebuilt to make natural-language data extraction more useful and reliable.
Previously, AI Mode relied on captured lists (based on recorder mode) from the page. It now builds an extraction robot based on your prompt and the data available on the specified page.
Describe the data you want and Maxun builds a reusable robot to extract it.
Once created, the robot can be run repeatedly without making LLM calls on every run, keeping subsequent extractions fast and cost-efficient.
Want more autonomy? Maxun Cloud also offers Agent, which can discover relevant pages and configure websites to find the data you need. We'll continue improving both AI-powered workflows in upcoming releases.
Learn more here: https://docs.maxun.dev/robot/extract/llm-extraction, https://docs.maxun.dev/robot/agent
🕷️ Crawling is faster and cleaner
Crawling is now 77% faster, with our benchmark dropping from 411 seconds to 96 seconds.
We've also reduced duplicate and irrelevant data, so crawls return cleaner, more useful results:
- 100% fewer duplicate links in the output, down from 28% to zero.
- 50% smaller HTML output per page, from 280 KB to 139 KB.
- 16% higher Markdown precision, reducing irrelevant content in extracted pages.
- No duplicate articles in the benchmark, down from 23.
The result: faster crawls, less redundant data, and cleaner output for downstream processing.
📄 Better document extraction
Document Robots are now better at extracting structured data from large and complex files, including PDFs, spreadsheets, CSVs, and DOCX documents.
- More complete results: Preserve document metadata alongside line items and other records.
- Better large-file handling: Recover from truncated model responses and process document chunks in parallel.
- Improved table parsing: Preserve table structure, headers, and row order across file formats.
- More accurate data: Better handling of dates, international number formats, and duplicate records.
In tests with simulated model responses and realistic output limits:
- Invoices: Recovered all 4 header fields and all 60 line items.
- Bank statements: Extracted all 300 transactions, up from 25.
- CSV files: Extracted all 1,500 orders, up from 5 rows.
Document Robots are now substantially more reliable when processing large files and datasets.
🔌 A new MCP server
The previous MCP implementation has been replaced with a new, more capable server that exposes Maxun's core functionality directly to AI assistants and MCP clients.
New one-shot tools include:
maxun_scrape- Scrape a page as Markdown, HTML, text, links, a summary, or screenshots, with optional Smart Queries.maxun_crawl- Crawl a website with configurable scope, depth, and include/exclude paths.maxun_search- Search the web in discovery or scrape mode, with time-range support.maxun_agent- Build an AI-powered extraction robot from a prompt and run it.maxun_cancel_run- Cancel a running job.maxun_get_run_diff- Inspect changes between runs.
Existing tools have also been renamed with the maxun_ prefix. For example, get_run_details is now maxun_get_run.
One-shot tools can reuse a robot when called again with the same arguments, building a run history instead of creating a separate robot for every repeated request.
The new MCP server makes Maxun's scraping, crawling, search, extraction, and run-management capabilities available through a much more unified interface. Learn more: https://docs.maxun.dev/mcp/tools
🗑️ Bulk-delete robot runs
Managing run history is easier. You can now bulk-delete runs for a robot directly from the UI, making it simpler to clean up old runs without deleting them one by one.
What's next?
0.1.0 is an important step toward making Maxun a more capable and reliable open-source web data platform. We're continuing to improve extraction accuracy and performance.
What's Changed
- feat: use snake_case keys in captured-data JSON downloads by @amhsirak in #1235
- fix: keep leading digits in snake_case keys by @amhsirak in #1239
- feat: return snake_case keys for captured data in API responses by @amhsirak in #1240
- feat: show llm extract prompt in settings by @amhsirak in #1241
- perf(scrape): main-content isolation, single page snapshot, live DOM left intact by @amhsirak in #1242
- perf(crawl): skip non-page links, clean capture, ignore tracker requests by @amhsirak in #1243
- fix(i18n): add 10 English keys that render as raw keys by @theluckystrike in #1234
- chore(swagger): update metadata + add tags by @amhsirak in #1246
- fix(documents): keep header fields, use PDF tables, survive cut-off answers by @amhsirak in #1244
- feat: hide target URL, list limits, actions, duplicate and retrain for AI/Agent robots by @amhsirak in #1247
- fix: show retrain option only for extract recorder mode robots by @amhsirak in #1248
- feat: sync new cloud ui for robot & agent creation by @amhsirak in #1249
- feat: ai extract v0 [cloud sync] by @amhsirak in #1250
- chore(deps): upgrade i18next-http-backend to 4.0.4 by @amhsirak in #1253
- feat: Add delete all runs function by @Wootark-Kim in #1233
- feat: move delete all runs button to top by @amhsirak in #1255
- fix: allow bulk run deletion for any robot type by @amhsirak in #1256
- feat(mcp): add scrape, crawl, search and agent tools to the MCP worker by @amhsirak in #1257
- feat(proxy): expand {session} placeholder in proxy username for per-run sticky IPs by @amhsirak in #1259
- fix(documents): stop defaulting Anthropic to retired Claude 3.5 Haiku by @amhsirak in #1260
- fix(sdk): skip temperature and forced tool_choice for Claude models that reject them by @amhsirak in #1261
- docs: add Simplified Chinese README translation by @WanderLandWalker in #1254
- chore: -rm legacy code & add maxun oss security scope by @amhsirak in #1262
- fix: handle nullable proxy username by @amhsirak in #1263
- fix: wrap username helper text in by @amhsirak in #1264
- chore: remove browse auto robots button by @amhsirak in #1265
- chore: pre-release v0.1.0 by @amhsirak in #1266
New Contributors
- @theluckystrike made their first contribution in #1234
- @Wootark-Kim made their first contribution in #1233
- @WanderLandWalker made their first contribution in #1254
Full Changelog: v0.0.48...v0.1.0