Benchmarks Show MCP Outperforms REST for GitLab AI Agents on Tokens and Latency

·1 views

A developer published benchmarks comparing the Model Context Protocol (MCP) with GitLab's traditional REST API for autonomous AI coding agents. The REST API returns bulky JSON payloads full of UI fields that waste LLM context tokens, while MCP resources deliver compact, model-oriented context. In the reported tests, MCP reduced payload size by about 92% and cut prompt token cost roughly 12.6 times, with parsing latency falling from 185 ms to 12 ms. Schema-constrained MCP search also eliminated the parameter errors seen in unconstrained REST queries, which occurred in about 22% of 100 tasks. The author, who contributed to GitLab's MCP implementation, presented these results as evidence that MCP suits agentic workflows better than REST.

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)

GitLab CI/CD Caching Methods Can Cut Docker Build Times by Up to 80%

Developers often face long CI/CD pipeline runs because GitLab's ephemeral runners start with an empty Docker cache, forcing layers to be rebuilt on every commit. Pulling base images from Docker Hub also frequently hits rate limits, causing 429 errors. GitLab offers two built-in features to address both issues: Container Registry with BuildKit layer caching and the Dependency Proxy, which acts as a local pull-through cache for upstream images. The author's guide shows how combining these techniques, along with a layer-friendly Dockerfile ordering, can reduce container build times by as much as 80 percent.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

n8n 2026 Pricing: Cloud Plans Versus Self-Hosting Real Costs

Standard n8n pricing comparisons focus on sticker prices, but overall cost depends on how many workflow executions run each month and what self-hosting really involves. n8n bills per execution rather than per user, and includes unlimited users and active workflows on every plan. Frequent scheduled triggers can push usage past the Starter or Pro limits quickly, with the next tier costing far more. Self-hosting the free Community Edition still requires a server, backups, and ongoing maintenance time, which can add up to more than a Cloud subscription. The article advises checking n8n's pricing page directly, since prices change often.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Open-Source LocalRAG Lets Teams Query Documents Without Cloud AI APIs

A developer has released LocalRAG, an open-source document question-answering system designed to run entirely within a private or air-gapped environment. The stack lets users upload PDF, DOCX, TXT, or Markdown files, which are chunked, embedded with a local bge-m3 model, and stored in Qdrant for retrieval. Answers are generated by a local LLM through Ollama and returned with source references, with no cloud AI calls at runtime. The project is aimed at sectors like legal, healthcare, and finance that cannot send confidential data to external services. Initial setup still requires network access to pull Docker images and models, and the developer noted that the large unpacked image size can be a problem on slow connections.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

CDP-6001 Exam Breakdown: PySpark Code That Covers 58% of the Cloudera ML Exam

A DEV Community article explains that Cloudera's CDP-6001 Machine Learning Engineer exam has 45 questions to be answered in 90 minutes, with a 60% passing mark. Three domains, Spark, Spark MLlib, and deploying models, together make up 58% of the test, and all are practiced through code in a Cloudera Machine Learning session. The author walks through a sample churn project covering loading events, window features, pipelines, tuning, MLflow tracking, and deployment. Because the exam is proctored and forbids reference material, the author argues that hand-writing the code beforehand helps candidates recognize errors within the roughly two minutes available per question. The piece also covers Spark DataFrame practices such as explicit schemas and partitioned Parquet files.

0 comments Read more at DEV Community