MCP Morning — Thu, Oct 29, 2026
Good morning. Today's picks: a local search engine your coding agent can query instead of reading files one by one, a server that makes agents plan in seven phases before they code, a self-hosted Firecrawl alternative that checks every page it reads, Korean tax math done in code, and a shared context graph for the agents inside a product. The story is a measurement of what Claude Code's background-agent status line costs.
Trending & new MCPs
- XERJ — One local Rust binary that indexes code, docs, logs, JSONL/CSV and PDFs for full-text, vector and hybrid search, with an MCP server and an Elasticsearch-compatible API.
Why it matters:
xerj initwires it into Claude Code or Cursor, and its main use case is "reference coding": the agent clones and indexes open-source repos close to the task and searches how they solved it before writing code. The project claims that cut output tokens 2.7x at the same solve rate in its own case study. The quick start pipes a remote install script into your shell, so read it first. Apache-2.0, 17,455 ⭐ on GitHub. - M7 MCP — Makes a coding agent go through discovery, architecture, solution, impact analysis and an implementation simulation, writing an artifact for each, before it touches code.
Why it matters: instead of trusting the model to remember the rules, the server issues a gate ID with every phase, rejects reflections that don't quote it, re-injects a short rule summary after a number of tool calls and keeps cycle state on disk so a new chat can resume. It's a published snapshot of a private tool (38 tools), and it currently refuses to start without
OPENAI_API_KEYandANTHROPIC_API_KEY, though placeholders work if you skip its multi-model feature. Apache-2.0, published two days ago, 50 ⭐ on GitHub. - FrankenSurf — A free, self-hosted alternative to Firecrawl: one container with the same API, which reads each page with the cheapest tool that works and checks it got the real page.
Why it matters: a challenge page, login wall or empty search page counts as a failure with a reason, so your agent doesn't answer from a CAPTCHA screen. It climbs from plain HTTP to browsers and stealth browsers, and to paid APIs only if you allow them. The project claims 96% of 150 blind-picked sites read correctly with paid fallbacks and 88% with free tools only. Available as an MCP server (
frankensurf-mcp). Apache-2.0, 4,082 PyPI downloads last week. - Korea Tax Calculator — 85 tools from Upstage for Korean corporate, income, VAT, withholding and inheritance tax, where the math is done in code and every result cites the statute it follows. Why it matters: it uses year tables for 2016 to 2026 and stops instead of estimating when a year isn't covered, and it rejects unknown or missing inputs and says what it needs. The calculation tools run locally with no API key; only the PDF tools call Upstage. The README is in Korean, with an English version. MIT, 487 PyPI downloads last week.
- Keel — A context layer for the agents inside a product: domain events go in, a per-tenant context graph with history comes out, and every agent reads and writes it through the same five tools. Why it matters: the pitch is that your second agent shouldn't need a second copy of the customer's data and its own memory. You describe your domain once in a YAML "cartridge", and each fact keeps its source and when it was valid and recorded. It runs over MCP, HTTP or SDKs for TypeScript, Python, Go and PHP, on SQLite or MySQL. The author labels it alpha (0.34.0). Apache-2.0, 2,003 PyPI downloads last week.
Agent story
Claude Code burns ~15–20% of your tokens on a 3-word status line (HN) — Flamerix AI writes that every 30 seconds or so, for each background agent, Claude Code sends that agent's entire transcript to its model and asks for a 3-to-5-word description of its latest action, which only feeds the status line in the agents panel. Grouping Claude Code's OpenTelemetry events by query_source, they found these agent_summary requests were 16.4% of their Claude Code spend the week before, slightly more than the main conversation, with 96% of it on Opus. They link a GitHub issue that collects four other measurements of 18 to 26%. Since there's no setting to turn it off, the gist offers an 80-line Bun preload that keeps the summary timer from firing without modifying the binary, and warns that an update can silently undo it. 378 points on Hacker News.
Rising this week
Biggest week-over-week download growth in our rankings:
- REA — 478,884 downloads (+53%)
- claude-flow — 18,616 (+343%)
- mobile-mcp — 52,895 (+258%)
- e2e — 473,877 (+163%)
- playwright-mcp — 6,908,874 (+207%)
npm counts now run through Oct 27. 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.