StackRadar MCP

Ask which developer tools are actually heating up.

Comment

Elevator

I run StackRadar, a small desk that scores about 30 developer tools from GitHub, Hacker News, Dev.to, Reddit, and RSS. The number is absolute, 0 to 100: stars 60 percent, forks 15 percent, community mentions 25 percent. It works on a screen. It is clumsy out loud.

StackRadar MCP puts that formula on a self hosted MCP server. A host can ask what is trending in AI this week, compare FastAPI with Next.js, or hear a weekly digest. Every tool returns a short spoken line plus the breakdown. If an API is down, a saved snapshot still answers, and the line says so.

Description

StackRadar MCP is a Python server on the official MCP SDK. It speaks spec revisions 2025-11-25 and 2026-07-28 over Streamable HTTP. Four tools: trending_tools, compare_tools, tool_score, weekly_digest.

The score is calculate_all_tool_scores from my StackRadar repo, including the log floors, the 25 mention cap, and the word boundary matcher. The catalog slugs match that repo too.

The new part is the voice layer. Leaders follow the momentum score. The "who got louder" line is the change in community mentions against a saved baseline. A week of stars barely moves a log scale that tops out near 300,000, so the digest reports the thing that can actually move.

Live reads use the public GitHub, Hacker News, and Dev.to APIs, plus RSS. Reddit's public search returned 403 Blocked from this environment, so that slice falls back to the snapshot. Successful fetches cache for six hours. A failed source is remembered for 15 minutes. A live refresh gives up after about eight seconds and answers with whatever it has, so a spoken question does not hang.

There is a demo client that opens a real Streamable HTTP session and calls all four tools, a sample host config, tests, a Dockerfile, and a SAM template for a Lambda Function URL. Lambda's buffered URL cannot carry the default SSE stream, so each POST returns one JSON body and the server keeps no session.

The Alexa+ add-on CLI and web simulator are select partner only. I cannot attach this server to a device today. The FAQ says a self hosted server plus your own client is the submission. That client is in the repo.

How to try it

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pytest
STACKRADAR_MODE=seed STACKRADAR_PORT=8741 stackradar-mcp
python clients/demo_client.py

Health: http://127.0.0.1:8741/health

The demo prints a protocol version and one spoken line per tool. Snapshot mode is deterministic. Set STACKRADAR_MODE=auto when you want live GitHub, Hacker News, and Dev.to numbers. Optional GITHUB_TOKEN if you have one.

Live MCP endpoint on AWS Lambda: https://54v2yptdgkqlv7bu2mfjbcjaaa0raiuw.lambda-url.ap-south-1.on.aws/mcp

Built With

Share this project:

Updates

Submission history