jarvis-setup

Jarvis Infrastructure Setup

Set up and configure Jarvis infrastructure components.

Environment Detection

This skill requires local shell, Docker, and filesystem write access.

• Claude Code (CLI): Full access — can install MCP, configure Docker, create credentials → proceed with all steps
• Claude Cowork (web/desktop): No shell, no Docker, no ~/.claude.json access → can only review current config and plan setup steps. Tell Chadron: "Infrastructure setup requires a Code session. I can review your current config and create a setup checklist."

Context

• Workspace root: resolved from CLAUDE.md canonical root. In Cowork sessions, use mounted workspace (visible in session environment).
• Console project: Jarvis/console/
• napa-mcp install: servicenow_data/mcp/napa-mcp/ (isolated venv)
• napa-mcp creds: .env.test / .env.dev / .env.prod (in mcp dir, gitignored — .env.example is template)
• Backend .env: Jarvis/console/backend/.env
• MCP config source of truth: Jarvis/console/backend/config/mcp_config.json

Setup Menu

JARVIS SETUP
─────────────────────────────────
[1] MCP Server          Install or update ServiceNow MCP server
[2] Docker              Set up or fix Docker environment for frontend
[3] Credentials         Configure ServiceNow instance credentials
[4] AI Assist           Configure Anthropic API key for NL interpreter
[5] Health Check        Verify all infrastructure components
[6] Instance Pipeline   Configure dev/test/prod instances
[7] Story Mode          Validate router-first execution readiness
─────────────────────────────────

Option 1: MCP Server Setup

napa-mcp (sole MCP server — Snow-Flow RETIRED Wave 48)

Install Path: servicenow_data/mcp/napa-mcp/ (isolated venv) | GitHub: chadronbryant/napa-mcp (private)

MCP servers are registered by instance name, NOT generic "servicenow":

Environment	Instance	MCP Server Key	Tools	Lazy?	Purpose
Test	napaanesthesiatest	napaanesthesiatest	388	Yes	Reads / analysis (default)
Dev	napaanesthesiadev	napaanesthesiadev	388	Yes	Writes / builds
Prod	napaanesthesia	napaanesthesia	388	Yes	Read-only, OAuth svc_napa_mcp

Lazy mode: only tool_search, tool_execute, list_tool_packages load at startup (not all 388 tools). Register in ~/.claude.json (Claude Code). Config source of truth: Jarvis/console/backend/config/mcp_config.json. The brain_reader.py health check looks for the instance-keyed name. napaanesthesia (prod) requires manual add by Claude Code — AG cannot edit ~/.claude.json.

Hallucination guardrail: Never use napaanesthesia-dev variant — only officially configured URLs.

Router-First Tool Selection

Prefer batch/aggregate tools over chained individual calls:

• batch_request / snow_batch_request — collapse multi-record reads
• aggregate_query — replaces query + manual count in one call
• snow_record_manage — unified CRUD, avoid switching specialized tools
• nl_query_search — build encoded queries without iteration

Option 4: AI Assist Setup

1. Create .env in Jarvis/console/backend/:

   ANTHROPIC_API_KEY=sk-ant-api03-...
   

2. Backend loads via load_dotenv(override=False) at startup — container env vars win over baked values.
3. Verify: GET /api/mcp/nl/status → {"available": true}
4. Cost: ~$0.003–0.01 per NL interpretation call (Claude Sonnet)

Option 5: Health Check

1. MCP Server: Check ~/.claude.json for instance key → venv exists → API connectivity
2. Docker: Docker daemon running → both console-backend-1 and console-frontend-1 containers running
3. Backend: Docker container on port 8000 → curl http://localhost:8000/api/health
4. AI Assist: /api/mcp/nl/status → Anthropic API key configured
5. Brain Files: Core brain files exist and readable
6. Data Freshness: last_data_pull vs today
7. Local Inference: POST http://100.86.223.86:8000/api/inference (jarvis-brain, Tailscale)
8. Agent Mesh: GET /api/mesh → all nodes healthy

Option 7: Story Mode (Router-First)

Use this before ServiceNow story execution in lite profile:

scripts/claude/story_mode.sh ready

Expected:

• Profile can stay light.
• jarvis-router is up and exposes execute_tool.
• ServiceNow reads/writes should route via router instead of direct MCP server calls.

If router preflight fails:

1. Restart session and re-run readiness check.
2. Switch to full only when needed:

   scripts/claude/switch_mcp_profile.sh full
   

3. After direct-debug task, return to light profile.

Important

• Never store passwords in brain files or logs
• Backend runs in Docker (console-backend-1) — rebuild with docker compose -f docker-compose.headless.yml after changes
• macOS requires python3 not python
• MCP server runs via venv Python — always use venv's bin/python
• CoWork VM cannot make outbound HTTP — MCP tools only work from local
• Snow-Flow is RETIRED (Wave 48) — do not reference or install it
• Prod instance (napaanesthesia) is read-only by default — confirm test validation before any write
• Plugin update procedure: edit source → bump plugin.json + marketplace.json → uninstall+reinstall → restart session