mcp-obsidian-cli: Skill for Claude Code

Desktop Sniff Test

End-to-end smoke test of the MCP server in Claude Desktop. Claude Desktop can't be automated — it requires manual copy-paste — so this skill orchestrates the config swaps and result evaluation around that manual step.

Prerequisites

• Obsidian must be running
• Claude Desktop must be installed
• The project must be the mcp-obsidian-cli repo

Workflow

Step 1: Swap config to local server

Read the Claude Desktop config file:

~/Library/Application Support/Claude/claude_desktop_config.json

Save the current mcpServers.obsidian block so it can be restored later. Then replace it with the local dev server:

"obsidian": {
  "command": "node",
  "args": ["<project-root>/server.js"],
  "env": {
    "OBSIDIAN_VAULT": "<keep existing vault value>"
  }
}

Preserve the existing OBSIDIAN_VAULT value — don't hardcode it.

Tell the user:

Config updated to local server. Restart Claude Desktop to pick up the change.

Step 2: Give the test prompt

Tell the user to paste this into Claude Desktop:

Read the file <project-root>/test/sniff-test-prompt.md and run every test in it.

Use the actual absolute path to the project root, not a placeholder.

Then say:

Paste that into Claude Desktop and come back with the results when it finishes.

Wait for the user to return with results. Do not proceed until they paste results back.

Step 3: Evaluate results

When the user pastes the results, use a Sonnet agent to evaluate them. This saves Opus tokens since evaluation is mechanical pattern-matching work.

Launch the agent with:

• The full pasted results
• The contents of test/sniff-test-prompt.md (so it knows the pass criteria)
• Instructions to evaluate as described below

The agent should:

1. Parse the summary table — extract test name, result, criteria counts
2. Count passes and failures — report X/N (N = total tests in sniff-test-prompt.md)
3. For any failures, check:
   • Was the tool call correct?
   • Did the pass criteria match what the test expects?
   • Is it a real bug or a test issue?
4. Check content samples — verify the quoted output looks correct (right headings, expected syntax patterns, reasonable data)
5. Flag anything suspicious even in passing tests (e.g., truncated output, unexpected content, wrong sort order)