Searchplex 1.1.0 — Perplexity-style web research for OpenCode (MCP)

Searchplex connects a model in OpenCode (or any other MCP client with the stdio transport) to the internet. The model decides when to search, Searchplex finds and reads the pages, and the model answers with numbered sources, as in Perplexity. It works with a local model (LM Studio, Ollama, llama.cpp, vLLM) or a cloud OpenAI-compatible API, and needs no subscriptions or API keys.

• One file, srchplex.py, Python 3.9+ standard library only.
• Search engines: your own SearXNG (installed with one command), then DuckDuckGo, Yandex and Bing; Bing News, arXiv and OpenAlex; Stack Overflow, GitHub and Hacker News APIs for code and discussions; optionally Tavily, after SearXNG by default or first if you prefer (see Search priority).
• Pages are read as clean Markdown with code blocks and tables; PDF, GitHub, Stack Overflow, Hacker News and Reddit are read through their APIs or feeds.
• Ranking with BM25 or embeddings, deduplication, iterative research rounds with reflection.
• Search inside local PDF, DOCX, ODT, PPTX and EPUB files; an exact calculator.

Tested with the local models qwen3.8-27b and qwen3.8-flash-next.

The research pipeline is ported from Vane (Perplexica).

Tools

In OpenCode the names get the server prefix: srchplex_deep_search and so on.

Tool	What it does
deep_search	Research: evidence passages with numbered sources; mode = speed, balanced (default) or quality (also a cited report, 1–5 min)
search	Quick results without reading pages: title, URL, snippet
fetch	Up to 3 pages as Markdown; query keeps only the relevant parts, start continues reading, extract extracts facts with the LLM
local_search	Passages from PDF, DOCX, ODT, PPTX and EPUB files in paths
calculate	Exact calculator with variables and math functions

focus: web, news, academic, discussions, code. time_range: day, week, month, year. domains: only these sites, e.g. ["docs.python.org"].

Install

You need OpenCode, Python 3.10+ (3.9+ for Searchplex itself; SearXNG needs 3.10) and a model in LM Studio, Ollama or a cloud API. On Windows write python instead of python3.

1. Get the project: git clone <repository URL> ~/searchplex, or download the release archive. Run the commands below from that folder.

2. Optional, recommended: install SearXNG. The script downloads it, enables autostart and runs a test search:

   python3 searxng_setup.py

3. Register the server in ~/.config/opencode/opencode.json (%USERPROFILE%\.config\opencode\opencode.json on Windows). OpenCode 2.x:

   {
     "mcp": {
       "servers": {
         "srchplex": {
           "type": "local",
           "command": ["<path to python>", "<project folder>/srchplex.py"],
           "environment": {
             "LLM_BASE_URL": "http://localhost:1234/v1",
             "TAVILY_API_KEY": "{file:~/.config/opencode/.tavily_key}"
           }
         }
       }
     }
   }

   OpenCode 1.x puts the servers directly under "mcp", without "servers". LLM_BASE_URL is LM Studio (http://localhost:1234/v1), Ollama (http://localhost:11434/v1) or any OpenAI-compatible server. Drop the TAVILY_API_KEY line if you do not use Tavily. Keep the srchplex name: AGENTS.md refers to the tools by it.

4. Copy the rules for the model, or append them to your own AGENTS.md:

   mkdir -p ~/.config/opencode && cp AGENTS.md ~/.config/opencode/AGENTS.md

5. Restart OpenCode and check everything:

   python3 srchplex.py --doctor

Search priority and Tavily

Tavily is an optional cloud search API (1000 free searches a month). Without TAVILY_API_KEY it is never used. With a key, SRCHPLEX_TAVILY in the server's environment sets its place in the order. Engines are asked in turn; the next one only when the previous one fails or finds fewer than 3 results (then the results are merged).

SRCHPLEX_TAVILY	Web	News
fallback (default)	SearXNG → Tavily → DuckDuckGo → Yandex → Bing	SearXNG → Bing News and Tavily together
primary (Tavily first)	Tavily → SearXNG → DuckDuckGo → Yandex → Bing	Tavily → SearXNG → Bing News
last (save credits)	SearXNG → DuckDuckGo → Yandex → Bing → Tavily	no Tavily
off	no Tavily, even with a key	no Tavily

Papers, code and discussions never use Tavily: they have their own sources. Restart OpenCode after changing it.

A wrong key or used-up credits pause Tavily for an hour and the search goes on with the next engine.

Settings

Set them in the environment of the server entry in the OpenCode config; for the CLI, in the terminal.

Variable	Default	Purpose
LLM_BASE_URL	http://localhost:1234/v1	OpenAI-compatible model API: LM Studio, Ollama, llama.cpp, a cloud API
LLM_API_KEY	lm-studio	API key, if the server needs one (cloud)
SRCHPLEX_LLM_MODEL (LMSTUDIO_MODEL, LLM_MODEL)	the loaded model	The model to use; required for cloud APIs and servers with several models. LM Studio and Ollama load a model that is not loaded yet
SRCHPLEX_LLM	auto	off: no LLM at all
SRCHPLEX_REASONING_EFFORT	none	Sent as reasoning_effort; off: not sent
SRCHPLEX_LLM_CONCURRENCY	2	Parallel LLM requests
EMBEDDING_MODEL	the loaded embedding model	The embedding model; required for cloud APIs
SRCHPLEX_EMBEDDINGS	auto	off: BM25 only
SEARXNG_URL	http://localhost:8080	SearXNG address; off: do not use it
GITHUB_TOKEN	—	Higher GitHub API limits
TAVILY_API_KEY	—	Tavily key; without it Tavily is not used
SRCHPLEX_TAVILY	fallback	Tavily's place: fallback, primary (first), last, off; see Search priority
TAVILY_TIMEOUT_MS	10000	Tavily request timeout
SEARCH_PROVIDER	—	Compatibility with configs of other search MCP servers: ddg = no Tavily, tavily = Tavily first; SRCHPLEX_TAVILY wins
SRCHPLEX_BROWSER	auto	off: do not render JavaScript pages with Playwright
SRCHPLEX_REGION	—	DuckDuckGo region (kl), e.g. us-en
SRCHPLEX_CACHE_DIR	<tmp>/srchplex_cache	Cache folder
SRCHPLEX_BLOCK_PRIVATE	—	1: fetch refuses localhost and LAN addresses

Diagnostics

python3 srchplex.py --doctor       # certificates, cache, SearXNG, Tavily, every search source, LLM, embeddings, PDF
python3 searxng_setup.py status    # SearXNG: installation, autostart, test search

Optional packages, picked up when installed into the Python that runs Searchplex: pypdf (PDF in local_search and fetch), playwright with playwright install chromium (pages that are empty without JavaScript), certifi (root certificates for python.org builds on macOS).

Limitations

DuckDuckGo, Yandex, Bing and Bing News are read as ordinary web pages, without an official API: their pages change, they rate-limit frequent queries and their terms may not allow automated access. For steady use install SearXNG or add Tavily. Searchplex is meant for personal research, not bulk data collection.

Search improves facts, not reasoning: the model can still misread a source. Ask for the official docs when details matter.

License

MIT. Includes the MIT notice of Vane (Perplexica).