pi-research

What pi-research does

pi-research is a tool that gives coding agents free, unlimited web search and a knowledge store. The README states that search and scraping run locally through a stealth browser with no search provider and no monthly cap, and that the only cost is LLM tokens. It is built as an extension for the Pi Coding Agent and also as a standalone agent skill for Claude Code, Codex, and other skills compatible agents, plus an SDK. The core loop is a research run that moves through agent teams: a coordinator plans and starts with a search, researcher agents scrape and read in parallel, and a research lead decides whether to do another round before writing a cited Markdown report, with findings optionally saved to the knowledge store. Three depth levels, normal, deep, and ultra, set the team size and number of rounds, and the tool picks the right one from a natural language request. The README frames the value as avoiding the quota and monthly fees of hosted AI search providers, which serve results from their own index on their servers, while pi-research searches and scrapes locally through DuckDuckGo in a stealth browser. The index is local: findings can be saved to a LanceDB knowledge store that seeds future runs, so repeat questions get faster and cheaper answers. This design targets agents that need to read the web without hitting a paid search API limit.

Safety and design choices

The README emphasizes that pi-research is read only by design. The research agent cannot run shell commands or write, edit, or delete anything, which the authors argue is the right shape for an agent whose whole job is reading untrusted web pages. They note that prompt injection picked up during a run has nothing to act on, because the agent lacks write or execute capabilities. This is a deliberate containment choice rather than a side effect, and it addresses the common risk that a web browsing agent gets manipulated by page content. The tool also supports a persistent and searchable knowledge store scoped globally or per project, so a team can build a reusable corpus of findings. Use cases listed in the README include researching inside Pi with no API credential needed, researching from Claude Code or Codex while a cheaper or local model drives the run so it does not spend the main agent's budget, keeping a knowledge store of findings, and building agent systems that find and read web content or populate a dataset of sources. The README presents these as concrete workflows rather than abstract benefits. The local index and the read only stance together mean the tool trades the convenience of a managed search API for a self hosted scrape that the operator controls, with the caveat that it still depends on an LLM with a large context window to plan and write the report.

Install, requirements, and limits

Installation depends on the channel. As a Pi extension the README shows pi install npm:@lincoln504/pi-research, and standalone it shows npm install with a global flag followed by pi-research skill install. The README adds that on npm 12 or newer the user must pass --allow-scripts to either install, because npm 12 stopped running dependency install scripts by default and the stealth browser's native dependencies are built by one; without it searches fail with a missing module error. The first install downloads the stealth browser engine and takes a few minutes. Requirements are Node.js 22.19.0 or newer, an LLM with a 100k plus context window via an API credential or local model, and internet access on a residential IP because search, scraping, and YouTube transcripts get bot blocked from datacenter, VPS, or cloud IPs. Cloudflare Turnstile and similar systems block scraping on some sites, and the README says a run compensates with a wide pool of search results to scrape. The npm package @lincoln504/pi-research is the stable channel kept current with breaking Pi changes, while a git install is the development channel with the latest commits that breaks first. The package identity note warns that the unscoped pi-research name on npm is unrelated and deprecated. With 30 stars in the metadata, the project is early stage but clearly documented about its limits.

Editorial conclusion

The package is distributed under the MIT license and its most recent commit was on 2026-08-26.