Skip to content

rrrrrredy/research-toolkit

v2.0.0MIT

Complete research methods for in-depth analysis, long-form reports and revisions.

Research Toolkit plugin

English | 简体中文

One installation bundles the research Skill, detailed methods and nine local MCP tools. The agent clarifies the brief, researches and writes with its own capabilities. When task management is needed, the bundled MCP exposes the same local operations as the CLI.

Use the host's supported plugin installation mechanism. The guide below covers Python 3.10+ and MCP dependencies. Bundled manifests launch python: the interpreter found by the host process must have the dependencies installed. If the host cannot select that environment for the plugin, connect MCP separately with the absolute virtual-environment interpreter path from the guide.

Open a new session, select Research Toolkit, and confirm the nine research_* tools appear. Supply a question or existing report, sources and intended use; the agent uses the complete research methods. Installing the plugin makes tools available; it does not require creating a managed task for every report.

Full installation and usage guide · Research instructions · Complete standard

New tool-managed tasks without a backend use explicitly degraded self-review. Configure a CLI or compatible HTTP backend for external/independent review. Declared independent reviews, evaluations and audited plans never silently downgrade. The usage guide explains data destinations, account usage and configuration checks.

The portable plugin.json / mcp.json and Codex-compatible .codex-plugin/plugin.json / .mcp.json start the same implementation. MCP is local stdio, not a hosted HTTPS endpoint. The package includes no credentials. Mechanical checks do not certify research quality.

For maintainers, the package's core files are generated from the repository with python scripts/build_plugin.py; --check detects divergence. Edit the shared source first.

Generated ownership is recorded in .generated-files.json. Read-only --check compares contents and managed paths; builds delete obsolete copies only when their recorded hash still matches. Handwritten metadata and unmanaged content remain; inspect unknown ownership or modified obsolete files before proceeding.