SIGNAL RADAR

Open source · local first · personal use

Turn your AI news intake into signals you can act on.

Signal Radar reads the newsletters and feeds you already follow, builds a knowledge graph and a time-series signal store from them, and tells you what is gaining momentum, what is new, what is fading, and what that means for your work. Every signal carries its evidence and a confidence level.

MCP server with 13 typed tools 18 architecture decision records 162 automated tests Local or cloud models, one router 0 cloud services required

What it answers

Six questions, every week, from your own sources

Not a news reader. An instrument that counts, compares and explains, and shows you the articles behind every number.

Rising

What is gaining momentum?

Entities whose weekly mentions outrun their own baseline, confirmed by independent source families.

New

What just appeared?

First-seen entities and pairs of entities that start turning up in the same stories: clusters forming.

Fading

What is losing attention?

Topics dropping below their trend, so you stop investing time in yesterday's conversation.

Lens

What matters to me?

Each item is scored against a reader profile you write in plain text. The default lens: financial services, risk and the EU AI Act.

Learn

What should I learn?

Skills rising across sources, ranked by momentum and fit, each with one concrete learning project.

Build

What should I create?

Product, service and content ideas, each tied to the signals and articles that justify it.

Architecture

One pipeline, every stage inspectable

Each stage writes plain files or SQLite you can open, diff and rebuild. The corpus is the source of truth; everything to its right is derived. A default install runs the solid parts; the dashed ones are optional tiers.

SOURCES Email newslettersoptional · IMAP, Graph RSS / Atomfeedparser, page meta Web archivespolite, cached backfill INGEST signal-sourcesMCP server 13 typed toolssource adaptersnewsletter parsersdedupe on url + titlesubscribe by config STORE CorpusMarkdown + YAML,one file per item Extractionentities, relations,category, relevance MODEL Signal storeSQLite: velocity, novelty Knowledge graphoptional · Graqle, Neo4j Vector indexoptional · hybrid search USE Console · Today, Explore, Ask Weekly signal brief Claude via MCP tools and prompts Static public edition LLM routerlocal or cloud model per role, in one file
Core: runs in every install Optional: turn on with an install extra or a setting Integration points

Components

What is inside

Each component sits behind a small interface and has its design decision written down.

MCP server

signal_radar/mcp_server.py

Built on the official Python mcp SDK 2.x, stdio transport. Typed inputs, docstrings written for the model, structured JSON out.

  • Tools: top_movers, trend_metrics, emerging_clusters, reading_list, semantic_search, run_ingest, subscribe_source and more
  • Resources: corpus://items/{id}, corpus://digest/{date}
  • Prompts: weekly_signal_brief, learning_radar, content_ideas
ADR-0005 →

Source adapters

signal_radar/sources/

One interface, fetch(since) → list[Item]. Subscribing to a feed or newsletter is a config entry, not a code change.

  • Gmail over IMAP, read-only, app password in the OS keychain
  • Microsoft 365 over Graph, delegated Mail.Read, device-code sign-in
  • RSS/Atom with per-feed cleanup rules; web-archive backfill
  • Parsers for TL;DR, AlphaSignal and Bay Area Letters
ADR-0004 →

LLM router

signal_radar/llm.py

Every model call goes through one router that reads config/models.yaml. No model name lives anywhere else.

  • Roles: extraction, reasoning, writing, embedding; override per run with --model role=provider/model
  • Local (Ollama, LM Studio, llama.cpp, vLLM) or cloud (Anthropic, OpenAI, OpenRouter), chosen per role; ranked fallbacks
  • cheapest_passing: the cheapest model that met a quality bar on your own evals
  • Monthly budget cap; every call logged with tokens, cost and input hash
ADR-0013 →

Extraction and ontology

signal_radar/extract.py

A local model reads each item once and returns schema-checked JSON. Resumable, cached by input hash, re-run when the taxonomy changes.

  • 8 entity types, 7 relation types, editable in config/taxonomy.yaml
  • Relevance 0 to 3 against your written lens, calibrated so most news scores low
  • Alias table folds "GPT-5" and "gpt5" into one entity
ADR-0010 →

Signal store

signal_radar/signals.py

A SQLite sidecar for the time-series maths a graph does not do: weekly mentions, velocity, novelty and co-occurrence growth.

  • Confidence from volume and independent source families, not one loud newsletter
  • Fully rebuildable from corpus plus extractions, no LLM calls
  • Watchlist blind spots: a zero means your sources miss it
ADR-0018 →

Knowledge graph and vector search Optional

graph_export.py · vectors.py

An optional tier. Documents and entities go into a Graqle graph file, optionally mirrored into Neo4j for queries, story clusters and lens links. Without it, Explore and clusters come from the signal store.

  • Neo4j sync with a restricted user and a delete-fraction guard
  • In-process ONNX embeddings, cached by text hash, bounded per run
  • With vectors on, Ask is hybrid: keyword plus vector, merged by reciprocal rank fusion
ADR-0019 →

Console and static site

signal_radar/site/

Static-first: JSON data files plus templates. The local console adds Ask and feedback on top of the same pages.

  • Today: lens themes, ranked reading list, rising and fading with sparklines
  • Explore: radial ego graph per entity with its news
  • Method: every number on the page explained
ADR-0011 →

Evaluation, brief and operations

evals.py · brief.py · schedule.py

Models are compared on this project's own tasks before they are trusted, and the jobs run unattended.

  • LLM-as-judge plus blind human grading sheets
  • Weekly brief: signals computed in code, a model writes only the prose, uncited ids removed
  • launchd scheduling with catch-up after sleep; optional publishing behind Cloudflare Access
ADR-0014 →

The console

An instrument, not a dashboard

Screens from the bundled demo: synthetic news about fictional companies, generated by scripts/demo.py. Click through the live version.

Today page: themes for your lens, a ranked reading list, rising entities with sparklines
Today. Themes are found in the data, a model only names them. The reading list ranks by lens relevance, cross-source coverage and momentum. Confidence shows as filled dots.
Explore page: a radial graph of entities mentioned with Contoso AI and the news behind them
Explore. Pick any entity: its neighbours in the graph, its stated relations, its 12-week trend and the articles behind it.
Weekly brief page: rising signals with evidence, items through your lens, source coverage and watchlist
Weekly brief. Rising signals with evidence and source-family counts, the week through your lens, and a watchlist that makes blind spots visible.

Design principles

Decisions that shape every component

 

Evidence or it did not happen

Every signal links to the items behind it. Generated text cites ids; invented ids are stripped. Thin evidence is labelled thin.

Code computes, models narrate

Counts, velocity and clusters come from SQL. Models classify and write, never decide what is rising.

Local first

Runs on a laptop with a local model by default. Cloud models are your choice per job, capped and logged.

Models are swappable

Roles, not model names. Change provider in one YAML file and re-run the evals.

Respect publishers

Store summaries and links, never full articles. Sources whose terms forbid sharing stay out of any public edition.

Write the decision down

18 ADRs record context, options, the choice and its consequences, including the gaps.

Quickstart

Try it in two minutes

The demo needs Python 3.11+ and nothing else: no mailbox, no API key, no model.

# get the code
git clone https://github.com/jpshetty-ai/signal-radar-community
cd signal-radar-community
python3 -m venv .venv && source .venv/bin/activate
pip install -e '.[mcp,web,dev]'

# synthetic demo, opens on http://127.0.0.1:8765
python scripts/demo.py --serve

Your own radar

Point it at your sources

Public EU tech and policy feeds are on by default. Add newsletters and your own lens when you are ready.

# a model: local by default, or set a cloud
# model per role in config/models.yaml
ollama pull qwen3:8b

# fetch, extract, score, build the console
signal-radar daily
signal-radar web

# use it from Claude
claude mcp add signal-sources -- .venv/bin/signal-sources-mcp

Built with

Python 3.11+MCP SDK 2.xPydanticSQLiteOllamaAnthropicOpenAIOpenRouterGraqleNeo4jfastembed / ONNXFastAPIJinja2feedparserMSALpytestlaunchdCloudflare Pages + Access