riva

AI Agent Command Center

Observe, monitor, and control local AI agents running on your machine.

Riva is a local-first observability and control plane for AI agents.
It helps you understand what agents are running on your machine, what they are doing, and how they are behaving in real time.

As agent frameworks push toward autonomy, visibility often disappears.
Riva exists to restore clarity, safety, and trust.

Getting Started · How it works · CLI Reference · Web Dashboard · Security · Contributing · Discord

Demo

Agent Overview	Infosec Monitoring

Event Stream (v0.3.14)	Events → Forensic Drill-in

Forensic Timeline — Tool Calls, Patterns, Dead Ends

How it works

Local Agents (Claude Code / Codex CLI / Gemini CLI / LangGraph / CrewAI / AutoGen / ...)
        |              |                        |
        | hook events  | session JSONL tail     | OTLP push
        v              v                        v
        +----------------------------------------------+
        |                    Riva                      |
        |  discovery · metrics · logs · lifecycle      |
        |  hook_events (tool calls · sessions · spans) |
        +--------------------+-------------------------+
                             |
        +--------------------+-------------------------+
        |          |         |           |             |
      CLI         TUI    Web Dashboard  REST API   System Tray
                Events tab  /api/events  /otlp/v1/*  (macOS)

Riva runs entirely on your machine.
It observes agent behavior but does not execute agent actions.

Highlights

• Agent discovery — detect locally running agents across 15 frameworks and growing (Claude Code, Kiro, Cursor, Cline, Windsurf, Continue.dev, Codex CLI, Gemini CLI, Strands, and more)
• Lifecycle visibility — see when agents start, stop, crash, or hang
• Resource tracking — CPU, memory, and uptime per agent in real time
• Token usage stats — track token consumption, model usage, and tool call frequency
• Event stream (new in v0.3.14) — real-time tool-call visibility via three ingestion paths: Claude Code hook events, JSONL session file tail-watching (Claude Code + Cursor), and OTLP HTTP receiver (any OTel-instrumented agent); all surfaced in the new Events tab (3/e) in riva watch
• Claude Code hooks (new in v0.3.14) — riva hooks install registers Riva into Claude Code's hook system in one command; every PreToolUse, PostToolUse, SessionStart, SubagentStop, and Stop event is streamed to Riva in real time
• OTLP receiver (new in v0.3.14) — POST /otlp/v1/traces|metrics|logs accepts protobuf and JSON; LangGraph, CrewAI, AutoGen and any OTel-instrumented agent can push directly to Riva
• Environment scanning — detect exposed API keys in environment variables
• Sandbox detection — detect whether agents run inside containers (Docker, Podman, containerd, LXC) or directly on the host
• Session forensics — riva forensic deep-dive analysis of agent session transcripts — timeline, patterns, decisions, efficiency metrics; works for both interactive and non-interactive (API/MCP) sessions
• Skills system — riva skills tracks workflows across sessions with usage count, success rate, backtrack rate, and avg token cost; discoverable from all agents (Claude Code commands/skills, Kiro hooks/specs, Cursor rules, Continue slash commands, Windsurf memories, Codex/Gemini instructions, Cline rules)
• OpenTelemetry export — riva otel pushes metrics, logs, and traces to any OTel-compatible backend (Datadog, Grafana, Jaeger) via OTLP
• Boundary monitoring — continuous policy evaluation every poll cycle — flag violations for file access, network connections, process trees, and privilege
• Compliance audit log — tamper-evident JSONL log with HMAC chain, CEF export for SIEMs, integrity verification via riva audit verify
• Security audit — riva audit checks for config permission issues, exposed secrets, and dashboard misconfiguration
• System tray — native macOS menu bar app for quick access to TUI, web dashboard, scan, and audit (compiled Swift)
• Web dashboard — Flask-based dashboard with REST API, security headers, optional auth token, forensic drill-in, Skills tab, and Events feed
• Skill description — riva --skill-help outputs structured Markdown so any AI agent can use riva as a tool; invocable as an MCP skill
• Framework-agnostic — works across multiple agent frameworks and custom agents
• Local-first — no cloud, no telemetry, no hidden data flows

Supported Frameworks

Riva ships with built-in detectors for these agent frameworks:

Python-based frameworks (LangGraph, CrewAI, AutoGen) are detected by matching python processes whose command line references the framework.

Adding more frameworks — Riva is extensible via:

1. Built-in detectors in src/riva/agents/
2. Third-party pip packages using [project.entry-points."riva.agents"]
3. Plugin scripts dropped into ~/.config/riva/plugins/

What Riva Is Not

Riva is intentionally not:

• An AI agent
• An orchestration framework
• A cloud monitoring service
• A replacement for agent runtimes

Riva does not make decisions.
It makes agent behavior visible.

Requirements

• macOS (Ventura, Sonoma, Sequoia) or Linux
• Windows via WSL2
• Python 3.11+
• System tray (optional, macOS only): Xcode Command Line Tools (xcode-select --install)

Quick Start

Install from PyPI

pip install riva

Install via bash script

curl -fsSL https://raw.githubusercontent.com/sarkar-ai-taken/riva/main/install.sh | bash

Install from source

git clone https://github.com/sarkar-ai-taken/riva.git
cd riva
pip install -e ".[test]"

Verify

riva --help

CLI Reference

riva scan

One-shot scan for running AI agents.

riva scan              # Rich table output
riva scan --json       # JSON output
riva scan --otel       # Scan and export metrics/logs to OTel collector

riva watch

Launch the live TUI dashboard with real-time resource monitoring.

riva watch

Tab keys:

riva tray

Native macOS system tray (menu bar app). Provides quick access to both the TUI and web dashboards, plus scan and audit actions.

riva tray                      # Start daemon (background)
riva tray start                # Same as above
riva tray start -f             # Foreground mode
riva tray start --port 9090    # Custom web port
riva tray stop                 # Stop the tray daemon
riva tray status               # Show running state and PID
riva tray logs                 # View tray logs
riva tray logs -f              # Follow log output

Requires Xcode Command Line Tools (xcode-select --install). The Swift binary is compiled on first launch and cached at ~/.cache/riva/tray-mac.

Menu items:

riva stats

Show token usage and tool execution statistics.

riva stats                        # All agents
riva stats --agent "Claude"       # Filter by name
riva stats --json                 # JSON output

riva list

Show all known agent types and their install status.

riva list

riva config

Show parsed configurations for detected agents.

riva config

riva audit

Security audit and compliance commands.

riva audit                              # Run security audit (PASS/WARN/FAIL table)
riva audit run --json                   # JSON output
riva audit run --network                # Include network security checks
riva audit log                          # Show recent audit log entries
riva audit log --hours 48 --type boundary_violation  # Filter by time and type
riva audit verify                       # Verify HMAC chain integrity
riva audit export --format jsonl        # Export for compliance (default)
riva audit export --format cef          # CEF format for SIEM (Splunk, QRadar)
riva audit export --hours 72 -o report.cef  # Custom time range and output

Checks performed:

• API key / secret exposure in environment variables
• Config directory and file permissions (group/other-accessible)
• Web dashboard bind address and status
• Plugin directory existence and permissions
• MCP server configs — HTTP endpoints, shell commands, temp-dir references
• Plaintext token scanning across all agent config files
• Agent processes running as root (UID 0)
• Agent binary writability (group/world-writable binaries)
• Suspicious launcher detection (unknown or script-interpreter parents)
• Orphan process detection
• Network checks (with --network): unencrypted connections, unknown destinations, excessive connections, stale sessions

See Security Audit Details for the full list and rationale.

riva forensic

Session forensics — deep-dive analysis of AI agent session transcripts (JSONL).

riva forensic sessions                    # List recent sessions
riva forensic sessions --project riva     # Filter by project
riva forensic summary latest              # Summary of latest session
riva forensic timeline <slug>             # Event-by-event timeline
riva forensic patterns <slug>             # Detected anti-patterns
riva forensic decisions <slug>            # Key decision points
riva forensic files <slug>                # File access report
riva forensic trends                      # Cross-session aggregate metrics
riva forensic trends --limit 50 --json    # JSON output

Session identifiers: latest, a session slug (e.g. witty-shimmying-haven), a UUID prefix, or a full UUID.

riva skills

Track reusable agent workflows (slash commands, tool sequences) with forensic stats.

riva skills list                        # All skills with usage count, success rate, avg tokens
riva skills list --agent claude         # Filter by agent
riva skills list --json                 # JSON output
riva skills scan                        # Parse recent sessions and record invocations
riva skills scan --all-sessions         # Process all historical sessions
riva skills stats /commit               # Per-skill breakdown
riva skills add "deploy"                # Create a skill in the global registry
riva skills share /commit --to cursor   # Mark shared, reassign agent
riva skills export skills.toml          # Export all skills to TOML
riva skills import skills.toml          # Import skills from TOML (deduped by id)

Skills are defined in ~/.riva/skills.toml (global) or .riva/skills.toml (workspace):

[skill.commit]
name = "commit"
description = "Stage and commit with a conventional commit message"
agent = "Claude Code"
invocation = "/commit"
tags = ["git", "workflow"]
shared = true

Riva auto-discovers skills from agent-specific sources:

Slash commands (e.g. /commit, /review-pr) in session JSONL are automatically detected and linked to skill invocations during riva skills scan.

riva skills send — Cross-agent Skill Export

Share skills between workspaces and across agent frameworks. Riva converts each skill into the target agent's native format automatically.

# Same agent, different workspace
riva skills send frontend-design --to /path/to/project2

# Claude Code → Cursor
riva skills send frontend-design --to /project2 --agent cursor

# Claude Code → Gemini CLI
riva skills send frontend-design --to /project2 --agent gemini-cli

# Disambiguate when multiple agents have same skill name
riva skills send commit --from claude-code --to /project2 --agent codex-cli

Supported target formats:

Web dashboard: click Send on any skill row to open the export modal:

Skills Tab with Send buttons	Export Skill Modal

riva hooks

Install and manage Riva's real-time hook integrations with AI agents.

riva hooks install                    # Register Riva hooks in ~/.claude/settings.json
riva hooks install --agent claude-code  # Explicit agent (default)
riva hooks uninstall                  # Remove Riva hook entries
riva hooks status                     # Show which events have Riva hooks installed

Once installed, every Claude Code tool call (PreToolUse, PostToolUse), session start, subagent stop, and stop event is streamed to the running Riva server in real time.

Quick setup:

pip install riva
riva hooks install        # one-time setup
riva watch                # launch dashboard — press 3 for the Events tab
# now run Claude Code in any terminal

Environment variables for the hook script:

Event Stream — Real-time Tool-call Visibility

Riva v0.3.14 adds a unified event stream that captures every tool call, session lifecycle event, and agent action in real time. Three ingestion paths feed into one view:

Claude Code hooks ──→ POST /api/events ──→┐
JSONL tail-watching ──→ (automatic)  ──→  ├──→ hook_events table ──→ Events tab + API
OTLP receiver ──→ POST /otlp/v1/* ──→    ┘

1. Claude Code hooks (push — real-time)

Every PreToolUse, PostToolUse, SessionStart, SubagentStop, and Stop event is pushed to Riva as it happens.

riva hooks install claude-code    # one-time setup
riva web start                    # start the server
# Use Claude Code normally — events appear instantly

2. JSONL tail-watching (automatic — no setup)

Riva watches ~/.claude/projects/**/*.jsonl and ~/.cursor/projects/**/*.jsonl for new lines every 5 seconds. No configuration needed — if the Riva server or TUI is running, tailing is active.

Events appear as jsonl:tool_use, jsonl:tool_result, jsonl:assistant, jsonl:user, etc.

3. OTLP receiver (push — any OTel agent)

Any agent instrumented with the OpenTelemetry SDK can push traces, metrics, and logs directly to Riva:

# From curl
curl -X POST http://127.0.0.1:8585/otlp/v1/traces \
  -H "Content-Type: application/json" \
  -d '{"resourceSpans":[{"resource":{"attributes":[{"key":"service.name","value":{"stringValue":"my-agent"}}]},"scopeSpans":[{"spans":[{"traceId":"abc","spanId":"def","name":"search","startTimeUnixNano":"1712859232000000000","attributes":[]}]}]}]}'

# From Python — point any OTel exporter at Riva
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
exporter = OTLPSpanExporter(endpoint="http://127.0.0.1:8585/otlp/v1/traces")

Events appear as otlp:trace, otlp:metric, otlp:log.

Viewing events

Web dashboard — click the Events tab in the sidebar:

• Filter by agent, event type, or time range
• Auto-refreshes every 3 seconds
• Click any session ID to jump to the forensic session detail

TUI — press 3 or e in riva watch for the Events tab.

API — query events programmatically:

# All events from the last hour
curl "http://127.0.0.1:8585/api/events?hours=1&limit=100"

# Filter by agent
curl "http://127.0.0.1:8585/api/events?agent=Claude+Code"

# Filter by type
curl "http://127.0.0.1:8585/api/events?type_prefix=otlp"
curl "http://127.0.0.1:8585/api/events?type_prefix=jsonl"

Events → Forensics deep-link

Clicking a session ID in the event stream navigates directly to the Forensics tab with the full session analysis:

Forensic Session Summary	Forensic Timeline

The forensic view shows:

• Summary — model, duration, turns, tokens, efficiency score
• Patterns — dead ends, search thrashing, retry loops, write-without-read
• Timeline — every turn with prompt, tool calls, duration, and success/failure
• Files — all files read and written during the session
• Decisions — key decision points with reasoning previews

riva otel

OpenTelemetry export — push metrics, logs, and traces to any OTel-compatible backend.

OpenTelemetry is included with pip install riva — no extra install needed.

riva otel status                            # Show SDK availability and config
riva otel status --json                     # JSON output
riva otel export-sessions                   # Export forensic sessions as traces
riva otel export-sessions --limit 5         # Limit session count
riva otel export-sessions --project myapp   # Filter by project

Configuration via .riva/config.toml:

[otel]
enabled = true
endpoint = "http://localhost:4318"
protocol = "http"
service_name = "riva"
export_interval = 5.0
metrics = true
logs = true
traces = false

[otel.headers]
Authorization = "Bearer <token>"

Or via environment variables: OTEL_EXPORTER_OTLP_ENDPOINT, OTEL_SERVICE_NAME, RIVA_OTEL_ENABLED.

OpenTelemetry Observability

Riva exports all three OTel signals — metrics, logs, and traces — to any OTLP-compatible backend.

Quick Start with Grafana LGTM

# 1. Start the all-in-one Grafana stack (Loki + Grafana + Tempo + Mimir)
docker run -d --name otel-lgtm -p 4318:4318 -p 3000:3000 grafana/otel-lgtm

# 2. Install Riva (OTel support included)
pip install riva

# 3. Initialize workspace config (if not already done)
riva init

# 4. Enable OTel in .riva/config.toml
# [otel]
# enabled = true
# endpoint = "http://localhost:4318"
# metrics = true
# logs = true
# traces = true

# 5. Run a scan — metrics and logs are exported automatically
riva scan --otel

# 6. Export forensic session traces
riva otel export-sessions --limit 10

# 7. Import the bundled Grafana dashboard
curl -s -X POST http://localhost:3000/api/dashboards/db \
  -H 'Content-Type: application/json' \
  -u admin:admin \
  -d "{\"dashboard\": $(cat grafana-dashboard.json), \"overwrite\": true}"

# 8. Open Grafana at http://localhost:3000 (admin/admin)

Signals

Metric Names (OTel → Prometheus)

Grafana Dashboard

A pre-built Grafana dashboard is included at grafana-dashboard.json. It provides 12 panels covering all three signals — stat panels, time series, log viewer, and trace table.

Environment Variables

Boundary Policies

Configure continuous boundary monitoring via .riva/config.toml:

[boundary]
enabled = true
allowed_paths = []
denied_paths = ["~/.ssh/**", "~/.gnupg/**", "~/.aws/**"]
allowed_domains = ["api.anthropic.com", "api.openai.com"]
denied_domains = []
max_child_processes = 50
denied_process_names = ["nc", "ncat"]
deny_root = true
deny_unsandboxed = false

Boundaries are evaluated every poll cycle (default 2s). Violations fire the BOUNDARY_VIOLATION hook and are recorded in the tamper-evident audit log.

Web Dashboard

Start / Stop

riva web start                  # Background daemon
riva web start -f               # Foreground
riva web start --auth-token MY_SECRET   # With API auth
riva web stop                   # Stop daemon
riva web status                 # Check status
riva web logs                   # View logs
riva web logs -f                # Follow logs

Custom host and port

riva web --host 0.0.0.0 --port 9090 start

A warning is printed when binding to a non-localhost address.

API endpoints

POST /api/events — event ingestion schema

{
  "agent_name": "Claude Code",
  "session_id": "550e8400-e29b-41d4-a716-446655440000",
  "event_type": "PostToolUse",
  "timestamp": 1712859232.5,
  "tool_name": "Bash",
  "tool_input": { "command": "pytest tests/ -x" },
  "tool_output": "5 passed in 0.4s",
  "success": true,
  "duration_ms": 412,
  "metadata": {}
}

OTLP receiver — configure any OTel SDK

# Python — opentelemetry-sdk
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace.export import BatchSpanProcessor

provider = TracerProvider()
provider.add_span_processor(
    BatchSpanProcessor(OTLPSpanExporter(endpoint="http://127.0.0.1:8585/otlp/v1/traces"))
)

Authentication

When started with --auth-token, all /api/* and /otlp/* routes require an Authorization: Bearer <token> header. The index page (/) remains accessible without authentication.

# Start with auth
riva web start --auth-token secret123

# Access API
curl -H "Authorization: Bearer secret123" http://127.0.0.1:8585/api/agents

# OTLP with auth (set in your OTel SDK)
OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer secret123" your-agent

Security headers

All responses include:

• X-Content-Type-Options: nosniff
• X-Frame-Options: DENY
• Content-Security-Policy: default-src 'self' 'unsafe-inline'
• X-XSS-Protection: 1; mode=block
• Referrer-Policy: strict-origin-when-cross-origin

Security

• Runs locally — no network exposure by default
• Web dashboard binds to 127.0.0.1 by default
• Non-localhost binding triggers a visible warning
• Optional bearer token auth for the web API
• Security headers on all HTTP responses
• riva audit performs 15+ automated security checks (see below)
• No agent execution privileges — read-only observation

See SECURITY.md for the full security policy.

Security Audit Details

riva audit runs a comprehensive set of checks designed to catch real-world threats to local AI agent deployments. Each check below links to supporting evidence for why it matters.

Credential & Token Exposure

Covered config files per agent: settings.json, config.json, mcp.json, .env, config.toml (Codex CLI), config.ts (Continue.dev), OAI_CONFIG_LIST (AutoGen), langgraph.json (LangGraph), mcp_config.json (Windsurf). Also scans VS Code extension directories (Cline, Copilot, Continue) and macOS Application Support paths.

Permissions

Process Safety

MCP (Model Context Protocol) Supply Chain

Scans 6 well-known MCP config paths plus dynamically discovered mcp.json/mcp_config.json in every installed agent's config directory.

Network (with --network flag)

Dashboard

Architecture

src/riva/
├── agents/              # Agent detection and parsing
│   ├── base.py          # AgentInstance, AgentStatus, BaseDetector
│   ├── registry.py      # Agent registry
│   ├── claude_code.py   # Claude Code detector
│   ├── codex_cli.py     # Codex CLI detector
│   ├── gemini_cli.py    # Gemini CLI detector
│   ├── openclaw.py      # OpenClaw detector
│   ├── langgraph.py     # LangGraph / LangChain detector
│   ├── crewai.py        # CrewAI detector
│   └── autogen.py       # AutoGen detector
├── core/                # Core logic
│   ├── audit.py         # Security audit checks
│   ├── audit_log.py     # Tamper-evident JSONL audit log (HMAC chain)
│   ├── boundary.py      # Continuous boundary policy engine
│   ├── env_scanner.py   # Environment variable scanning
│   ├── forensic.py      # Session forensics (timeline, patterns, decisions)
│   ├── monitor.py       # Resource monitoring (CPU, memory)
│   ├── sandbox.py       # Sandbox / container detection
│   ├── scanner.py       # Process scanning
│   └── usage_stats.py   # Token/tool usage parsing
├── otel/                # OpenTelemetry exporter
│   ├── config.py        # OTel configuration loading
│   ├── metrics.py       # Metrics exporter (gauges, counters)
│   ├── logs.py          # Logs exporter (audit, lifecycle)
│   ├── traces.py        # Traces exporter (forensic sessions)
│   └── exporter.py      # Main coordinator
├── tray/                # System tray (macOS)
│   ├── manager.py       # Swift binary compilation, spawn, IPC
│   ├── daemon.py        # Background daemon management (PID, start/stop)
│   ├── run.py           # Daemon subprocess entry point
│   └── tray_mac.swift   # Native macOS NSStatusBar app
├── tui/                 # Terminal UI (Rich)
│   ├── components.py    # Rich table builders
│   └── dashboard.py     # Live dashboard
├── web/                 # Flask web dashboard
│   ├── server.py        # Flask app, REST API, security middleware
│   └── daemon.py        # Background daemon management
├── utils/               # Shared utilities
│   ├── formatting.py    # Display formatting helpers
│   └── jsonl.py         # JSONL file parsing
└── cli.py               # Click CLI entry points

Riva is modular by design.
New agent detectors can be added without changing the core.

Development

Setup

git clone https://github.com/sarkar-ai-taken/riva.git
cd riva
python -m venv .venv
source .venv/bin/activate
pip install -e ".[test]"

Running tests

pytest                                    # All tests
pytest --cov=riva --cov-report=term-missing  # With coverage
pytest tests/test_cli.py                  # Specific file

Linting

pip install ruff
ruff check src/ tests/
ruff format --check src/ tests/

Type checking

pip install mypy types-psutil
mypy src/riva/ --ignore-missing-imports

Release Process

1. Update version in pyproject.toml
2. Update HISTORY.md with changes
3. Run full test suite: pytest --cov=riva
4. Build the package: python -m build
5. Verify: twine check dist/*
6. Create a git tag: git tag v0.x.x
7. Push with tags: git push --tags
8. Create a GitHub Release — this triggers automatic PyPI publishing

Manual publish (if needed)

python -m build
twine upload dist/*

Uninstall

If you installed Claude Code hooks, remove them first:

riva hooks uninstall --all      # remove hook entries from ~/.claude/settings.json
riva web stop                   # stop the web daemon if running
pip uninstall riva

Without riva hooks uninstall, the hook entries remain in ~/.claude/settings.json. They're harmless (the hook script fails silently when Riva isn't installed) but leave orphaned config. Clean them up manually if needed:

# Manual cleanup (if you already uninstalled without removing hooks)
python3 -c "
import json; p = __import__('pathlib').Path.home() / '.claude/settings.json'
d = json.loads(p.read_text())
h = d.get('hooks', {})
for ev in list(h):
    h[ev] = [e for e in h[ev] if not any('riva.hooks' in x.get('command','') for x in e.get('hooks',[]))]
    if not h[ev]: del h[ev]
p.write_text(json.dumps(d, indent=2))
print('Cleaned')
"

Or use the uninstall script:

curl -fsSL https://raw.githubusercontent.com/sarkar-ai-taken/riva/main/uninstall.sh | bash

Early Stage Project

Riva is early-stage and evolving rapidly.

Expect:

• Rapid iteration
• API changes
• Active design discussions

Feedback is highly encouraged — join the conversation on Discord.

Philosophy

If you cannot see what an agent is doing, you cannot trust it.

Riva exists to make local AI agents inspectable, understandable, and safe.

Contributing

We welcome contributions and design discussions.

• See CONTRIBUTING.md for guidelines
• Join the Discord for questions and discussion

License

MIT — see LICENSE for details.

Share

If you find Riva useful, feel free to share:

• Share on X
• Post to Hacker News