Developer builds meta-server to search 75,000 MCP servers using BM25 ranking

·2 views

A developer has released mcp-anything, a single MCP server that lets AI agents search and connect to tools across a fragmented ecosystem of roughly 75,000 indexed MCP servers. The tool aggregates four catalogs — the official MCP registry, PulseMCP, npm, and Glama — deduplicating entries and ranking results using BM25 lexical search combined with a popularity boost from GitHub stars and download counts. Rather than requiring users to manually find, configure, and restart servers, the meta-server exposes five tools that allow the model itself to discover, inspect, and call external servers on demand. Security features include SSRF protection, a disabled-by-default code execution mode requiring an explicit allowlist, and labeling of third-party tool output to reduce tool-poisoning risk. A public hosted instance is available in discovery-only mode, while full execution capability is intentionally restricted to local deployments.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Two-Year Snapcraft Build Bug Fixed by Switching to kde-neon-6 Extension

The Snapcraft build for open-source app Packet Sender had been broken for two years, crashing on launch due to a missing shared library file called libpxbackend-1.0.so. Initial attempts to fix the issue using LD_LIBRARY_PATH environment variables and swapping Qt library versions, including a suggestion from AI tool Grok, failed to resolve the crash. The breakthrough came when contributor Dan Nagle recommended adopting the kde-neon-6 Snapcraft extension, which bundles a fully configured Qt 6 stack with correctly staged shared libraries. Because the extension already manages the required Qt 6 dependencies, most manually listed stage-packages and build-packages could be removed, and the build system was migrated to CMake. The fix reduced the final executable size by roughly 96 percent, shrinking it from 165MB to just under 6MB.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Waymo's Gemini In-Cabin Feature Remains a Beta With Phased Rider Access

Waymo has been rolling out its Gemini AI integration inside the Ojai cabin through a limited beta program rather than a full public launch. The company first announced the rollout in May 2026, offering free rides to a restricted group of early users across San Francisco, Phoenix, and Los Angeles to gather feedback. By July 2026, Waymo still described Gemini in the Ojai cabin as a beta feature, noting a redesigned user interface and continued product enhancements. The phased approach means broader access is expected to expand gradually over time, with no universal availability date announced. Waymo has clarified that current rider communications should be consulted to determine actual access, as the beta label signals the experience is still being actively refined.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI Models Can Memorise and Leak Verbatim Fragments of Training Data

Large AI models do not simply learn general patterns from training data — they can memorise and reproduce identifiable fragments of it verbatim, according to a decade of documented research. Two main attack methods exploit this: membership inference, which determines whether a specific record was included in a training dataset, and data extraction, which recovers memorised content word-for-word. Memorisation tends to increase with model scale and dataset size, meaning the industry's push for larger models and bigger datasets has amplified rather than reduced the risk. Membership inference alone poses serious privacy concerns, as confirming a person's record was in a training set can reveal sensitive information — such as medical history — even without exposing the record's contents. These vulnerabilities underpin ongoing legal disputes over AI-reproduced copyrighted content and privacy research showing that personal data can leak from production AI systems.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer fixes 58%-reproducible race condition in Vesuvius scroll-reading Python package

A developer testing the open-source Vesuvius Challenge Python package on Windows 11 discovered a race condition in its shared disk cache that caused failures in 58% of test runs. The bug stemmed from a Windows file-locking incompatibility: the zarr library's write-then-replace cache commit pattern, legal on Linux, triggers an access-denied error on Windows when multiple processes try to overwrite the same file simultaneously. Any multi-worker PyTorch DataLoader using the package would randomly crash mid-training run on Windows due to this issue. The fix wraps cache write and delete operations so that permission errors degrade gracefully to cache misses rather than aborting the read, with no platform-specific code branches. After the patch, the test suite went from a 58% failure rate to zero failures across 12 consecutive runs, and the Windows test results improved from 47 passed/2 failed to 52 passed/0 failed.

0 comments Read more at DEV Community