@jiffylabs/jiffy-chat-mcp

MCP server that lets Claude Desktop / Cursor / Claude Code ask natural-language questions about your Jiffy org and get cited answers.

Three tools:

• jiffy_query({ question, time_range?, tier_filter? }) — answers questions like "what are the riskiest artifacts on my fleet in the last 24 hours". Returns an answer plus typed citations (inventory items, approvals, audit rows).
• jiffy_recommend({ finding_id }) — given an inventory or approval UUID, proposes the next admin action (approve / quarantine / escalate / remediate) with a deep link back into the Jiffy UI.
• jiffy_next_step({ context }) — open-ended "now what?". Maps the current UI context (page, artifact id) to the most likely next action.

Answers come from a bounded templated intent router on the server side — no free-form LLM-to-SQL, no prompt injection surface into your database.

Install

pnpm add -g @jiffylabs/jiffy-chat-mcp

Or run without installing:

pnpm dlx @jiffylabs/jiffy-chat-mcp

Configure

The server reads an API key from JIFFY_API_KEY or ~/.jiffy/config.json — the same shape the Jiffy Intake CLI uses, so one key covers both.

export JIFFY_API_KEY=jtp_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
export JIFFY_API_URL=https://jiffylabs.app   # optional, this is the default

Or drop a file at ~/.jiffy/config.json:

{
  "apiKey": "jtp_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",
  "apiUrl": "https://jiffylabs.app"
}

Wire into Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "jiffy-chat": {
      "command": "jiffy-chat",
      "env": {
        "JIFFY_API_KEY": "jtp_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Restart Claude Desktop. The three jiffy_* tools appear in the tool picker.

Wire into Cursor

Cursor uses the same MCP shape. Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "jiffy-chat": {
      "command": "jiffy-chat"
    }
  }
}

Wire into Claude Code

Claude Code picks up MCP servers from ~/.claude/settings.json (or the project-local equivalent):

{
  "mcpServers": {
    "jiffy-chat": {
      "command": "jiffy-chat",
      "env": { "JIFFY_API_KEY": "jtp_live_..." }
    }
  }
}

Example prompts

• "Jiffy, what are the riskiest artifacts on my fleet in the last 24 hours?"
• "How many approvals are pending?"
• "Which endpoints are running drift?"
• "Which publishers have the lowest score?"
• "Is clawdbot installed anywhere?"
• "Who installed @acme/web-search?"
• "What changed in the last 7 days?"

Supported intents (V.12a)

The server routes questions to one of ~15 bounded intents. Unknown questions return a canned "here are the questions I can answer" response.

• top_risky_artifacts_last_24h
• pending_approvals_count
• endpoints_with_drift
• publishers_with_lowest_score
• find_artifact_by_name
• what_changed_last_7d
• who_installed_artifact
• compliance_coverage_by_framework
• recent_vet_denies
• agents_by_tier
• artifacts_missing_attestation
• runtime_invocation_rate
• critical_findings_open
• clawdbot_installed
• fallback_general

Non-goals (V.12a)

• No free-form LLM-backed SQL. Every query is a hand-written Supabase filter.
• No streaming responses.
• jiffy_recommend + jiffy_next_step are read-only. execute() exists on the server with dry_run: true as default — wiring it to live mutations is a follow-up.
• No conversation memory across tool calls.
• No Slack surface — that's V.12b.

License

Apache-2.0.