I'm the project owner. I've been running multiple AI coding agents simultaneously and had no way to answer basic questions: which one is using the most tokens? Why did that session end with 40 tool calls and no working code? Is any agent touching files it shouldn't?
Riva is a local-first monitor for AI agents running on your machine. No cloud, no telemetry.
What it does:
- Detects running agents (Claude Code, Cursor, Kiro, Codex CLI, Gemini CLI, Cline, Windsurf, LangGraph, CrewAI, AutoGen, and more) by process name, exe path, and
config dir
- Live TUI (riva watch) and web dashboard with CPU, memory, uptime, child processes, network connections
- Session forensics — parses agent JSONL transcripts into turn chains with timeline, tool calls, dead-end detection, and efficiency metrics. Works for both
interactive and headless/API sessions
- Skills tracking — detects slash commands (/commit, /review-pr) across sessions and computes usage count, success rate, backtrack rate, avg tokens per invocation
- Security audit — scans agent configs for exposed API keys, world-readable credential files, suspicious MCP stdio commands, and agents running as root
- Boundary policies — define allowed file paths, network domains, and max child processes; violations fire configurable hooks
- OpenTelemetry export — push metrics, logs, and traces to Grafana/Datadog/Jaeger
- riva --mcp-help — outputs a Markdown tool description so any agent can introspect and use riva itself as a tool
Install:
pip install riva
riva # TUI dashboard
riva web start # web dashboard at localhost:7821
The session forensics piece turned out to be the most useful part in practice. You can replay any past session as a structured timeline and see exactly where an agent went in circles, what it read vs wrote, and what decisions it made in its thinking blocks.
I'm the project owner. I've been running multiple AI coding agents simultaneously and had no way to answer basic questions: which one is using the most tokens? Why did that session end with 40 tool calls and no working code? Is any agent touching files it shouldn't?
Riva is a local-first monitor for AI agents running on your machine. No cloud, no telemetry.
What it does:
Install: pip install riva riva # TUI dashboard riva web start # web dashboard at localhost:7821The session forensics piece turned out to be the most useful part in practice. You can replay any past session as a structured timeline and see exactly where an agent went in circles, what it read vs wrote, and what decisions it made in its thinking blocks.
Website: sarkar.ai/riva/