SymbolPeek MCP Server Cuts AI Coding Context Usage by 95% Using Semantic Navigation

·1 views

A developer built SymbolPeek, an MCP server that gives AI coding agents symbol-level access to codebases instead of loading entire files. Rather than fetching a full source file to locate a single function, agents can request only the specific symbol they need, dramatically reducing token consumption. Real-world usage data from day-to-day development showed over 1.6 million tokens saved, 352,000 lines avoided, and an average context reduction of 95.7% across 162 requests. SymbolPeek leverages the TypeScript Compiler API for semantic analysis of TypeScript and JavaScript, while using Tree-sitter for syntax-aware navigation in Rust, Python, Java, Go, JSON, and Markdown. The tool is designed to expose compiler-level knowledge — such as inferred types, call hierarchies, and module resolution — directly to AI coding agents through the MCP protocol.

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)

How Top Tech Firms Build Systems That Speed Up as They Scale

Contrary to common assumption, software does not have to slow down as it grows — companies like Google, Netflix, and Amazon have built systems that actually get faster with scale. The core principle is eliminating unnecessary work rather than simply speeding up existing operations. Caching plays a central architectural role: as more users access the same content, cache hit rates rise, reducing database load and improving response times. Large systems also exploit data popularity by keeping frequently accessed items in fast memory while leaving rarely used data on slower storage. Precomputing results on a schedule, rather than recalculating them per request, further shifts workload away from users and onto background processes.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Professionals in Stable Jobs Still Deserve Space to Explore Career Options

Many experienced professionals exist in a middle ground between actively job hunting and being completely disengaged from the market, yet hiring systems rarely accommodate this nuance. People can be satisfied in their current roles while still holding legitimate curiosity about other opportunities, teams, or problems worth exploring. However, the traditional application process demands significant commitment upfront — résumés, portfolios, multiple interview rounds — before candidates learn the details that actually matter to them. This high barrier often discourages employed professionals from engaging at all, creating a paradox where companies miss out on their most desirable candidates. The article argues that career conversations should be built on mutual respect and early transparency rather than requiring full commitment before either side knows if there is a genuine fit.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Practical Code Review Habits to Make the Process Faster and Less Painful

A software developer has shared a set of refined code review practices aimed at making the process more productive for engineering teams. Key recommendations include keeping pull requests under 200 lines, reviewing overall architecture before line-level details, and using a consistent checklist covering correctness, error handling, and security. The author also emphasizes respectful communication, such as asking questions instead of making blunt criticisms, and acknowledging what has been done well. Automating style checks through linters and CI tools is advised so reviewers can focus on logic and design. Timely reviews within 24 hours and clear responses to all comments are highlighted as ways to maintain team momentum and mutual respect.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Lessons from running a web scraper processing millions of pages daily

A developer who built a large-scale scraping platform processing millions of pages daily at 95% extraction success shared the key failure points encountered along the way. Unbounded queues caused memory overflows in the message broker, while fixed retry backoffs accidentally triggered DDoS-like traffic spikes on struggling target sites. URL deduplication at enqueue time, rather than after fetching, proved critical to cutting costs and reducing block risk from duplicate content. Silent data degradation — where extraction appeared successful but key fields quietly returned null — was caught only by tracking field-level fill rates and alerting on deviations from baseline. Using multiple extraction strategies in parallel, such as embedded JSON, microdata, and DOM selectors, helped maintain high success rates when site redesigns broke individual methods.

0 comments Read more at DEV Community