ResearchRadar

Turn a reviewed research topic into a daily article whose claims point back to the source.

ResearchRadar watches a topic for new work, chooses the papers worth reading, reads the full text, and checks public claims against exact source anchors. It runs locally and leaves the finished article where you want to review it: the WeChat draft box, a public web archive with RSS, a Markdown export, or your private inbox.

简体中文 · Live Archive · RSS · Usage · Providers · macOS App Guide (local beta)

Local-first · Full-paper reading · Evidence-gated · No auto-publish

v0.3.0b1 is available as a pre-release. Download macOS App · CLI installation · Release notes and checksum

The App requires macOS 26 and Apple Silicon and includes its Python engine; it does not require the CLI, Python or uv. It is locally signed, not notarized; downloaded Gatekeeper behavior and a complete fresh first-use run remain unverified. See the App guide before installing.

What You Get

• A focused daily read. Paper-first discovery and topic-aware ranking keep the brief centered on the research question instead of filling it with generic web results.
• Full-paper explanations. Selected papers are read from usable source text or PDFs, with the problem, method, experiments, limitations, and figures explained in plain language.
• Claims you can inspect. Public factual claims need complete evidence anchors. Weak, broad, or unmatched claims stay in local audit artifacts instead of leaking into the article.
• One verified draft, several outputs. The same ArticleDraft can become a WeChat draft, a static Archive/RSS report, a Zhihu-ready Markdown export, or a private email.

How It Works

flowchart LR
    A["Reviewed topic"] --> B["Discover and rank"]
    B --> C["Acquire full papers"]
    C --> D["Read and explain"]
    D --> E["Split claims and anchor quotes"]
    E --> F["Verify conservatively"]
    F --> G["ArticleDraft"]
    G --> H["WeChat draft"]
    G --> I["Archive and RSS"]
    G --> J["Zhihu Markdown"]
    G --> K["Private email"]

Loading

Search expands recall, but snippets do not become publishable facts. The public article is built only after full-text acquisition, claim splitting, anchor checks, and verifier review.

Try It Once

These steps install the CLI from the fixed v0.3.0b1 source tag. There is no standalone CLI executable or PyPI release for this beta. App users can follow the App guide instead. You need Git, Python 3.12+, uv, a configured reader API, and the Codex CLI used by the default verifier. WeChat credentials are optional until you create a WeChat draft.

Install the project and create a private local config:

git clone --branch v0.3.0b1 --depth 1 https://github.com/SleepyLGod/research-radar.git
cd research-radar
uv sync --locked
uv run research-radar init

Use a new directory for this checkout. Development/test dependencies are optional; see Detailed Usage if you want to work on the project rather than just run it.

Edit config.yaml and replace example-topic with a topic you have reviewed. The file is gitignored; keep real topics and provider settings there rather than in config.example.yaml.

Store the secrets needed by the default discovery and reading path in macOS Keychain:

uv run research-radar secrets set deepseek
uv run research-radar secrets set web-search
uv run research-radar secrets status

Run the topic once:

uv run research-radar run daily \
  --topic <topic-id> \
  --config config.yaml \
  --root research-radar-data \
  --language zh \
  --model-cache

The command prints Created run: <RUN_DIR>. Treat that exact path as the input to later compose, archive, and publishing commands. Start with <RUN_DIR>/wechat.html or <RUN_DIR>/daily.md.

Publishing And Automation

• WeChat: upload safe paper figures and create a draft for review. ResearchRadar does not publish or mass-send it.
• Public Archive and RSS: export static files that can be hosted on GitHub Pages or another static host, or publish through a configured clean Git checkout. The live archive is one deployment.
• Zhihu: export constrained Markdown with local or public image URLs for manual import.
• Private email: send the same verified report to one personal inbox through TLS-protected SMTP.
• Daily scheduler: the CLI generates a macOS launchd job. The App uses its own schedule while it remains running; quitting stops App scheduling. Resolve overlapping schedules before enabling both.

App delivery is off by default. Enabling a channel allows independent delivery jobs after eligible research: WeChat creates drafts only, but private email actually sends to the configured recipient.

See Detailed Usage for the exact commands and deployment steps.

Defaults, Not Lock-In

The current quality path uses DeepSeek deepseek-flash with explicit thinking and high reasoning for DeepSeek-backed tasks, Tavily for web recall, and Codex gpt-5.6-luna with xhigh reasoning for verification. Provider instances and task routes are configurable; see Provider Configuration.

Trust Boundary

Reader-facing reports use only supported claims with complete evidence anchors. Rejected claims, weak evidence, source-selection details, provider diagnostics, and timing data remain in local audit artifacts. Renderers may reorganize verified content, but they cannot invent research claims.

ResearchRadar is self-hosted software, not a hosted research service. It stores secrets locally and does not automatically publish to WeChat, the public Archive, or Zhihu.

Documentation

• macOS App Guide (local beta)
• Detailed Usage
• Provider Configuration
• Architecture
• Security
• Roadmap

License

MIT