Research Toolkit

What research-toolkit does

The research-toolkit MCP server supports research tasks in the 🔬 Research category, including industry research, product comparisons, company analysis, and technical research. An agent provides the research question, audience, scope, sources, or existing report; the toolkit supplies methods and workflow support for analyzing evidence and writing a deliverable.

The project can be used in several forms: its methods can be followed directly from the repository, it can be installed as a Skill or plugin, or its local MCP tools can be connected to an MCP-capable client. The MCP option is intended for work that benefits from saved progress and explicit workflow records.

How it works

Research Toolkit defines two main workflows. Lite is intended for bounded research where author checks are sufficient. Full is intended for version-bound review, tracked revisions, and delivery-consistency checks. Full does not automatically require independent review, and the appropriate workflow depends on the task rather than report length.

The agent remains responsible for researching and writing. Toolkit provides guidance for required reading, evidence handling, analysis, and writing, along with records for claims and sources. When revising an existing report, the workflow can reuse completed research and its sources, update conclusions when new material arrives, and add a short change note.

Review configuration can use the author context, a configured external backend, or an independent review in a separate context. An explicit independent-review requirement must not be silently reduced to a different review tier.

Setup and configuration

For method-only use, no installation is required: give an agent access to the repository and sources, then direct it to read SKILL.md and follow the relevant methods. The repository also documents Skill and plugin installation for supported agents.

To use the MCP implementation, install and configure it as a local process through the client’s MCP settings. The README states that MCP requires Python 3.10 or later and a client capable of starting a local process. Installation guidance covers Windows, macOS, and Linux. The project also includes a runnable MCP example that replays a supplied self-review without making an external model call.

Tools and capabilities

The local MCP implementation provides eight tools. Based on the documented grouping, they manage task records, provide stage guidance, query registered release logs, run reviews, and perform delivery checks. These tools support maintaining working records during a research task and stating which checks were completed.

The toolkit can support a new study, continuation of Lite work in Full, report revision against sources, and updates based on newly available material. Its documented examples include reports with source lists, reviews, revisions, and research decisions.

Limitations and notes

The toolkit does not replace the agent’s research or writing process. Users still need to supply an appropriate question, scope, cutoff, and source material, and should select the workflow that matches the task.

Public evaluations document reports, failures, reviews, repairs, and tested workflow behavior. They do not establish a general quality or efficiency advantage over a native agent. The published case studies are based on public-source analysis and should not be treated as deployment tests.

The research-toolkit MCP server is local rather than a hosted research API. The README does not identify external data integrations or guarantee support for particular named MCP clients; compatibility depends on whether the client can start the local process and configure MCP.