Niobe

Runtime intelligence for AI agents — local-first runtime observation via MCP.

Niobe answers: "What is the system doing right now, and did my change make it worse?"

Take snapshot A → deploy → snapshot B → diff. That's it.

Install

pip install niobe

Quick Start

# Register a service
niobe register myapp --port 8080 --log /var/log/myapp.log

# Take a snapshot (captures metrics + ingests logs)
niobe snapshot myapp

# Make a change, then snapshot again
niobe snapshot myapp

# Compare before/after
niobe compare <snapshot-id-1> <snapshot-id-2>

# Search logs
niobe logs --query "connection refused"

# View recent errors
niobe errors --service myapp --since 10

CLI Reference

MCP Tools

Niobe exposes 5 tools via Model Context Protocol:

Claude Code Integration

Add to ~/.claude/.mcp.json:

{
  "mcpServers": {
    "niobe": {
      "command": "niobe-mcp",
      "args": []
    }
  }
}

Then in Claude Code:

• "Register my web server on port 3000" → niobe_register
• "How's my app doing?" → niobe_snapshot
• "Did my last deploy break anything?" → niobe_compare
• "Show me recent errors" → niobe_errors
• "Search logs for timeout" → niobe_logs

Log Format Support

Niobe auto-detects log formats:

• JSON — {"level": "error", "message": "...", "timestamp": "..."}
• CLF — Apache/NGINX common log format (HTTP status mapped to levels)
• Python — 2024-01-15 10:30:45,123 - myapp - ERROR - message
• RAW — Everything else (stored as-is with level=unknown)

Configuration

Optional. Create .niobe/config.toml in your project root:

[store]
db_name = "niobe.db"

[ingestion]
tail_lines = 100
max_line_length = 8192

[snapshot]
error_window_minutes = 5
cpu_sample_interval = 0.5

[mcp]
default_error_since_minutes = 5
default_query_limit = 50

All values can be overridden with NIOBE_* environment variables.

Architecture

• SQLite + WAL — single-file store, concurrent-read safe
• FTS5 — full-text search across log messages with porter stemming
• psutil — process metrics (CPU, memory, threads, connections)
• Frozen dataclasses — immutable models, no Pydantic dependency
• Snapshot-based — no background monitoring, entirely on-demand

Part of the EvoIntel MCP Suite

Niobe solves AI Blindness #2: Runtime Behavior — CPU, memory, error rates, and log patterns that no model can observe from code alone.

Part of the EvoIntel MCP Suite by Evolving Intelligence AI: five tools for five blindnesses no model improvement will ever fix.

License

MIT