Kansei MCP Server

What kansei-link/kansei-mcp-server MCP server does

kansei-link/kansei-mcp-server MCP server helps an agent research SaaS integrations before it attempts an API workflow. Its bundled SQLite database contains information on more than 11,000 SaaS services and MCP servers, including connection methods, authentication guidance, API details, known pitfalls, workarounds, trust signals, and usage information. The data covers 23 categories, with 2,257 entries identified as MCP-verified through a handshake.

The package also includes about 200 multi-service workflow recipes and 199 API guides. Search is intended to identify a relevant service or workflow, while lookup returns the practical information needed to use it. After an external operation, report lets the agent record success, failure, feedback, API changes, or qualitative observations. Submitted personal information is automatically masked according to the project description.

How it works

The normal interaction pattern is search_services, followed by lookup, the agent’s own API operation, and then report. Search supports intent-based discovery using full-text, trigram, and category-aware matching. Lookup can return default tips, a more complete connection guide, usage insights, recipes, service combinations, history, feedback, or community voices depending on the request.

The service records are shipped inside the npm package and read from a local SQLite database. The README states that normal lookup requires no API calls, hosted service dependency, signup, or network connection. Use kansei-link/kansei-mcp-server when an agent should consult a packaged knowledge base before guessing at authentication or request parameters.

The package also includes optional local measurement hooks. These parse session transcripts to calculate token totals, cache splits, failed calls, retry chains, stuck time, and KanseiLink call statistics. A monthly report can display measured values and separately labeled estimates. Sharing a rank or live tool-call events requires explicit consent; local mode is the default.

Setup and configuration

Run the MCP server with:

npx -y @kansei-link/mcp-server

Configure the same command and package in an MCP client such as Claude Desktop, Cursor, Cline, or Windsurf. The README also documents Claude Code support. The package includes separate commands for installing measurement hooks, generating the monthly report, installing the Claude Code skill, and managing live-update consent.

The usage hook stores local files beneath ~/.kansei-link/usage/. Measurement can be disabled with KANSEI_USAGE_HOOK=off. The optional reporting hook is controlled by the consent file and KANSEI_REPORT_HOOK; setting it to off, or disabling live updates, prevents transmission.

Tools and capabilities

• search_services: find services by name, intent, or category.
• lookup: retrieve tips, detailed guides, insights, recipes, combinations, history, feedback, or voices.
• report: submit outcomes, feedback, API change events, or experience reports.
• inspect: review documented colony-health and update-management data.
• analyze: inspect token savings, cost audits, and AEO reports or articles.

The first three tools form the documented standard flow. The latter two are additional administrative and analytics tools.

Limitations and notes

The knowledge base is local and packaged with the npm release, so its accuracy depends on the shipped dataset and update cycle. The README describes trust scores as weekly and says error-resolution intelligence is still forthcoming. Reported token savings are based on the project’s stated benchmark, not a guarantee for every workflow.

Optional central reporting is not enabled merely by installing a hook. When enabled, the documented event payload includes a service or server name, outcome, tool name, error category, and fixed context; it excludes prompts, inputs, outputs, record names, API keys, file paths, and free text. The monthly share feature sends scalar aggregates and uses a cohort size instead of a rank when fewer than 20 users are measured.