OpenClaw Plugin

KanseiLink ships as an OpenClaw plugin that turns every tool call into telemetry — across Claude, GPT, Gemini, or any model running inside OpenClaw.

Package: @kansei-link/openclaw-plugin (v0.1.0) License: MIT Repo: github.com/kansei-link/openclaw-plugin

What It Does

The plugin hooks into OpenClaw's tool-call lifecycle:

Hook / Tool	When	What
before_tool_call	Before any SaaS tool runs	Injects known pitfall warnings
after_tool_call	After any SaaS tool completes	Records success/failure/duration
kansei_diagnose	Agent hits an error	Returns diagnosis + workaround
kansei_confirm	Agent succeeds	Returns optimization tips

This means every MCP tool call (freee, Notion, Slack, Stripe...) is automatically observed — the agent doesn't need to do anything special.

Installation

1. Add to your OpenClaw config

{
  "plugins": [
    {
      "id": "kansei-link",
      "package": "@kansei-link/openclaw-plugin"
    }
  ]
}

2. (Optional) Configure telemetry mode

{
  "plugins": [
    {
      "id": "kansei-link",
      "package": "@kansei-link/openclaw-plugin",
      "config": {
        "telemetryMode": "local",
        "dataDir": "~/.kansei-link/telemetry/"
      }
    }
  ]
}

Option	Values	Default	Description
telemetryMode	"local" / "sync"	"local"	local = disk only. sync = send anonymized data to KanseiLINK central.
dataDir	path string	~/.kansei-link/telemetry/	Where telemetry JSONL files are stored.

How It Works

Automatic Observation

Once installed, the plugin silently observes every tool call:

Agent calls freee_api_get(...)
  ├── before_tool_call: check known pitfalls → warn if any
  ├── (tool executes normally)
  └── after_tool_call: record { tool, duration, success/fail, error type }

No agent code changes needed. The plugin runs at priority 40 (below default 50), so it never blocks other plugins.

Telemetry Storage

Data is stored locally as append-only JSONL:

~/.kansei-link/telemetry/
  2026-06-09.jsonl    # Daily log files
  stats.json          # Running aggregates per service

Privacy by design:

• Parameter values are stripped — only parameter keys are recorded
• No full response content stored
• Agent identity is not tracked
• telemetryMode: "local" (default) keeps everything on disk

Agent-Facing Tools

kansei_diagnose

When an agent hits a SaaS error, it calls kansei_diagnose:

kansei_diagnose({
  "toolName": "freee_api_get",
  "error": "company_id 不整合",
  "context": "Trying to fetch invoice list"
})

Response:

{
  "diagnosed": true,
  "diagnosis": "company_idが初期値'0'のまま。freee_set_current_companyを先に呼ぶ必要がある。",
  "workaround": "freee_set_current_company を最初に呼んで事業所を選択してから API を叩く",
  "confidence": 0.95,
  "evidenceSource": "source_scan:config.ts:160 + client.ts:107-121",
  "successRate": "47/50 (94%)"
}

If the error isn't in the knowledge base, it returns an honest "don't know" with suggested web search keywords — never vague or stale info.

kansei_confirm

After a successful connection:

kansei_confirm({
  "toolName": "freee_api_get",
  "tokensUsed": 1200
})

Returns optimization tips based on telemetry patterns (e.g., "average response time is high, consider batching").

Knowledge Base

The diagnosis engine ships with known pitfalls from source-code scans:

Service	Pitfall	Severity
freee	companyId defaults to '0' until freee_set_current_company is called	High
freee	401 after refresh = token revoked, not just expired	Medium
freee	company_id in POST body must match current company	Medium
freee	Rate limit: 3600 req/h, check Retry-After header	High

This knowledge base grows automatically as:

1. More source-code scans run (confirmed data production)
2. Agents report errors via kansei_diagnose
3. Telemetry accumulates failure patterns

KanseiLink MCP vs OpenClaw Plugin

	KanseiLink MCP Server	OpenClaw Plugin
Install	npx @kansei-link/mcp-server	Add to OpenClaw plugin config
Works with	Any MCP client (Claude, Cursor, Cline...)	OpenClaw only
Data source	11,000+ pre-built service evaluations	Live observation of your tool calls
Agent action	Agent calls search_services / lookup	Automatic (hooks observe silently)
Best for	"Which service should I use? How do I connect?"	"Why did this call fail? What's my success rate?"

Use both together for maximum coverage: KanseiLink MCP gives agents the map before they start; the OpenClaw plugin watches the journey and catches problems in real time.

Development

git clone https://github.com/kansei-link/openclaw-plugin
cd openclaw-plugin
npm install
npm run build    # tsc → dist/
npm run dev      # tsc --watch

Requires OpenClaw plugin API >=2026.3.24-beta.2.