web-research

End-to-end web research for MiniMax Code: search the web via Tavily (keyless, no API key required) and fetch each URL with a free-first cascade — defuddle → Tavily /extract → mcode in-app browser. One Skill, one Python entry point, no bash required.

The problem it solves

Research inside an agent usually means three separate tool calls, three failure modes, and a manual paste-into-context step. This plugin collapses the whole flow into a single Skill invocation that the agent runs end-to-end and then summarises.

Try it

After installing the plugin, ask the agent:

调研一下向量数据库的最新进展，5 篇文章，输出到 ./research-output

Expected result: a directory ./research-output/ containing 01-…md through 05-…md, each with a one-line tag ([defuddle] / [tavily extract] / [browser required]) on stderr. The agent then summarises the files in chat. Direct invocation:

python3 skills/web-research/scripts/research.py "向量数据库 最新进展" 5 ./research-output

Single-URL form:

fetch https://example.com/article and summarise

The agent uses the cascade directly. If defuddle is missing or fails, it transparently falls back to Tavily /extract; if both fail, the agent re-fetches with the mcode in-app browser tool.

Dependencies and supported platforms

Tested on Windows (Python 3.13, Node 22, no bash required), macOS 13+, and Linux (Python 3.11+).

The plugin never installs any of these for you. It only checks for defuddle with command -v defuddle and prints a one-time notice to stderr if the binary is absent.

Network and data behavior

This plugin talks to three network destinations, in order:

1. https://api.tavily.com/search (Tavily Search, keyless by default) — sends the user's query to Tavily. Tavily then forwards the query to the upstream search engines it is configured with (Google, Bing, etc.). The script never follows HTTP redirects; a 3xx is an error. No telemetry is sent by this script. Authorization: optional TAVILY_API_KEY Bearer, or X-Tavily-Access-Mode: keyless.
2. https://api.tavily.com/extract (Tavily Extract, keyless by default) — sends the URL chosen by the user to Tavily for raw-content extraction. Same redirect policy and auth model as above.
3. The user-provided URL (via defuddle or mcode browser) — defuddle fetches the page server-side via its own HTTP client; the mcode browser tool navigates the page in the in-app browser panel.

No third-party telemetry. No silent network calls. The plugin writes only to:

• The user-specified output directory (default ./research-output).

No credentials in the repo. TAVILY_API_KEY is read from the process environment only and is never logged or echoed. There are no native binaries, no symlinks, no installer scripts, and no ~/.mcode* writes.

Rate limits.

• Tavily keyless (/search and /extract) is free but rate-limited. On 429, the script surfaces Tavily's message and exits that URL with the browser-required marker.
• defuddle has no rate limit; it just times out per the user's network.

If you bring your own TAVILY_API_KEY, the key is sent only to https://api.tavily.com/* over HTTPS. It is never sent to defuddle, the browser tool, or any other destination.

Files

plugins/<owner>/web-research/
├── plugin.json
├── README.md
├── LICENSE
└── skills/
    └── web-research/
        ├── SKILL.md
        └── scripts/
            ├── research.py
            ├── tavily_search.py
            └── tavily_extract.py

Compliance

• No secrets, private endpoints, hidden telemetry, native binaries, or symlinks.
• All third-party network destinations are disclosed above.
• All third-party code is original to this plugin (the Tavily keyless header convention is documented publicly by Tavily).