MCP Morning — Thu, Oct 15, 2026
Good morning. Today is mostly about context: two servers that keep raw tool output and file-by-file reading out of your agent's window, two that hand it a map of your code, and a sharp read on why "rogue agent" stories are really containment failures.
Trending & new MCPs
- Context Mode — An MCP server that runs tool output in a sandbox so only the result reaches the context window, and tracks edits, tasks and decisions in SQLite so a compacted session can pick up where it left off. Why it matters: the README puts numbers on a familiar pain (a Playwright snapshot at 56 KB, twenty GitHub issues at 59 KB) and claims 315 KB of raw output comes back as 5.4 KB. It supports 17 clients. 29,528 ⭐ on GitHub.
- codebase-memory-mcp by DeusData — A local server that indexes functions, classes, imports, HTTP routes and their relationships into a persistent knowledge graph, across 162 languages. Why it matters: your agent asks the graph instead of opening files one by one. The project claims about 120× fewer tokens in its own five-query comparison, and indexing needs no hosted service or API key. 46,625 ⭐ and 10,594 npm downloads last week.
- Chrome DevTools for agents — The Chrome DevTools team's MCP server lets a coding agent control and inspect a live Chrome: performance traces, network requests, screenshots and console messages with source-mapped stack traces.
Why it matters: front-end agents can check what actually happened in the browser instead of guessing. Read the disclaimers first: the browser's content is exposed to the MCP client, and performance tools may send trace URLs to Google's CrUX API unless you pass
--no-performance-crux. 1,524,136 npm downloads last week. - Apple Docs 2.0 MCP — Apple's developer documentation, Human Interface Guidelines, App Store Connect Help and 1,700 WWDC sessions with transcripts and code, through one server. It's a fork of kimsungwhee's Apple Docs MCP with fixes and a new tool. Why it matters: Swift and SwiftUI APIs change every year, and an agent that can read Apple's current docs is less likely to write last year's code. First published to npm on Monday, with 7,861 downloads since.
- LayerMap — Read-only tools that answer "what calls this, and what does changing it affect?" in one call, tracing callers up to the HTTP routes, jobs and commands they start from. Why it matters: calls are resolved by each language's own compiler or type checker (TypeScript, JavaScript, Go, Python, Java), so it catches calls through interfaces and base classes that text search misses. The project's benchmark claims agents found 97.9% of affected endpoints with it versus 58.8% without. New this month, 3,109 npm downloads last week.
Agent story
AI agents aren't 'going rogue.' Frontier labs are skipping security basics. (HN) — CSO Online's Cynthia Brumfield talked to security experts about the recent run of incidents where lab models reached systems they weren't supposed to, Hugging Face among them. Their verdict: these weren't super-smart agents escaping, they were sandboxes set up wrong, missing egress controls, and experiments nobody was watching ("went on for days, and nobody noticed," says RAND's Dan Wallach). The advice is the boring kind: contain, monitor, and eat your security vegetables. It applies just as well to the agent on your laptop with a dozen MCP servers attached. 902 points on Hacker News.
Rising this week
Biggest week-over-week npm download growth in our rankings. Same top five as yesterday:
- REA — 270,176 downloads (+4,975%)
- Acquisition.gov — 10,650 (+2,418%)
- Concord MCP — 3,066 (+2,240%)
- tewip — 5,120 (+1,924%)
- outcometick — 5,999 (+1,614%)
Downloads include CI and bot traffic, so treat these as signals, not endorsements. How we rank →
MCP Morning lists servers from public data (the official MCP Registry, GitHub, npm and PyPI). We don't run or test them, so review the code and permissions before you connect one to your agent.