kansei-mcp-server

KanseiLink MCP Server is a local-first MCP navigator designed to reduce token waste by AI agents in SaaS integration research. It is officially positioned as a tool that lowers token consumption through collective intelligence.

Purpose AI agents often consume tokens in three ways: looking up SaaS documentation that could be queried locally, retrying errors that other agents have already solved, and repeatedly reading context that has already been processed. KanseiLink primarily addresses the first two and measures all three, so users can understand where tokens go. The officially reported measured savings are: 89–97% on SaaS integration research across 7 services (an average of about 16,800 tokens reduced to about 950).

Main capabilities

1. Measurement: automatically installed hooks track each session, counting total tokens, cache breakdown, error loops, and stuck time, with data never leaving the local machine.
2. Reduction: SaaS queries can eliminate trial and error in API integrations; error resolution intelligence (coming soon) is used to prevent repeated failures within the community.
3. Comparison: an optional monthly Wrapped report shows where tokens go and rankings among measured users.
4. SaaS integration intelligence: the core reason is that agents have to search documentation, guess authentication flows, and recover from errors every time. KanseiLink provides a local SQLite database so agents can get answers on the first attempt. All data is distributed with the npm package as a local SQLite database, with no API calls, no server dependency, and no registration required.
5. Tool consolidation: v1.0 consolidates the tool surface from 25 separate tools into 5 unified tools and supports automatic mode detection. The standard flow has 3 tools, and the management tools have 2.
6. Categories: covers 23 categories, including CRM, project management, communications, accounting, HR, e-commerce, legal, marketing, groupware, productivity, storage, support, payments, logistics, booking, data integration, BI/analytics, security, developer tools, AI/ML, databases, design, and DevOps.

How to use Supports any MCP client such as Claude Code, Cursor, Cline, Zed, and Windsurf. It can be added to configuration files (such as claude_desktop_config.json, .cursor/mcp.json, etc.), or added using the Claude Code CLI.

Claude Code skill installation: Installing only the MCP will not let Claude Code know when to call KanseiLink; the bundled skill must also be installed. This command copies SKILL.md to ~/.claude/skills/kansei-link/, where Claude Code will discover it automatically and invoke the skill when triggered by phrases such as "connect to Stripe", "Slack MCPある?", or "send invoice via freee", without needing to explicitly say "use KanseiLink".

Measurement hook installation: one command, idempotent, backs up settings first. Adds Stop / SessionEnd hooks, parses each session record, and writes total token counts and KanseiLink call statistics to ~/.kansei-link/usage/, without uploading anything. Reports can be viewed at any time. The report distinguishes measured figures (total tokens from the user's own records, number of KanseiLink calls, and response sizes) from estimated figures (avoided web research costs based on the 2026-04-16 freee/kintone/smarthr benchmark), with each interface labeled. The report also shows where the agent got stuck: failed tool calls, retry chains (the same tool failing 2 or more times in a row), tokens consumed while stuck, and the tools that failed most often. --share submits only scalar monthly aggregates (anonymous id + token counts, never content) and returns a ranking among measured users ("top X% saver"). When fewer than 20 measured users exist in the current month, it returns the cohort size instead of a ranking. Measurement can be disabled at any time: export KANSEI_USAGE_HOOK=off, or kansei-link-install-hooks --remove.

Optional PostToolUse hook: automatically captures success/failure after each MCP call (agents often forget to report).

Consent mechanism (v1.2, breaking change): installing the hook alone no longer transmits anything. All central transmission is governed by a consent gate (~/.kansei-link/consent.json), with priority: DO_NOT_TRACK=1 / explicit OFF → explicit ON (KANSEI_REPORT_HOOK=on) → Live Updates sameIntent (npx -y -p @kansei-link/mcp-server kansei-link-live-updates --enable) → default OFF (local mode, zero transmission). Existing hook users are OFF until they re-consent. Management commands: kansei-link-live-updates --status|--enable|--disable, kansei-link-privacy --status|--reset-id.

What the hook sends when enabled: a small pseudonymous event is sent to the KanseiLink hosted endpoint after each MCP tool call. The payload is a fixed 7-field set, frozen by a snapshot test (scripts/smoke-hook-payload.mjs): it sends the service slug (or MCP server name), success/failure, tool name, error category (such as auth_error), and a fixed context string; it never sends prompts, tool inputs/outputs, page/customer/record names, API keys, file paths, or any free text. No account or machine identifier is attached. Installing the hook is not the same as opting in — transmission requires the above consent gate (Live Updates consent or an explicit KANSEI_REPORT_HOOK=on). It can be disabled at any time: kansei-link-live-updates --disable or export KANSEI_REPORT_HOOK=off.

Example workflows: find and integrate services; multi-service workflows; share real opinions.

For SaaS companies KanseiLink also serves as an Agent Readiness Index (ARI) evaluation platform. Real agents using real APIs produce objective telemetry — success rates, latency, error patterns, and resolution paths — which no survey or benchmark can replicate. You can see: how your API's agent success rate changes over time, error patterns and how agents work around them, Agent Voice (why agents choose or avoid your service), category rankings against competitors, and API change impact (before-and-after analysis). This data comes from the same MCP that saves tokens for individual developers — the collective intelligence that helps agents is the same signal as the evaluation service.

Privacy and data handling Privacy protection by default: local-first, the complete service database is distributed with the npm package, no API calls; measurement stays local, and usage hooks are written to the local ~/.kansei-link/usage/, and are not uploaded unless you opt in with --share (only scalar aggregate token counts are sent, never content); PII is automatically redacted, and each reported call is stripped of email addresses, phone numbers, IP addresses, and Japanese names before storage; agent identity is anonymized, retaining only the agent type (claude / gpt / gemini), never a user ID; no telemetry by default, and the local stdio server does not communicate externally. See SECURITY.md for details.

Troubleshooting Verify the skill is installed: ls ~/.claude/skills/kansei-link/SKILL.md; if missing, run npx -y -p @kansei-link/mcp-server kansei-link-install-skill; restart Claude Code, and the skill will be indexed at session startup; check whether MCP is registered under the name kansei-link: claude mcp add -s user kansei-link -- npx -y @kansei-link/mcp-server; try category filtering: search_services({ intent: "...", category: "accounting" }); try English equivalents — most entries are indexed bilingually, but some are English-only; if the service truly does not exist, submit feedback: report({ subject: "Missing: ServiceX", body: "..." }); start with lookup({ service_id: "..." }) — it returns known OAuth pitfalls and refresh token workarounds; report failures: report({ service_id: "...", success: false, error_type: "auth_error", workaround: "..." }) — your fix helps the next agent.

Other information PRs are welcome. If you find a missing service or incorrect information, the fastest path is to submit a PR. MCP Registry: io.github.kansei-link/kansei-mcp-server. License MIT — Synapse Arrows PTE. LTD.。