Setting Up a Self-Hosted AI Observability Platform in 10 Minutes

Agentic Resource Registry (ARR, formerly Observal) runs entirely on your own infrastructure. This guide takes you from nothing to traced sessions from your team’s coding agents.

Prerequisites

• A Linux or macOS machine with Docker and Docker Compose
• Python 3.11 or newer for the CLI, and pipx to install it
• About 10 minutes

Step 1: Start the Server

curl -fsSL https://raw.githubusercontent.com/AgenticResourceRegistry/AgenticResourceRegistry/main/install-server.sh | bash

The installer downloads the latest release, runs a short guided setup, and brings up the whole stack with Docker Compose:

• The API server (FastAPI) and a background worker
• PostgreSQL for the registry: users, agents, components, reviews
• ClickHouse for session events and traces
• Redis for queues, live updates and auth
• The web app, behind an nginx load balancer
• Prometheus and Grafana for the server’s own metrics

When setup finishes, open the web app in your browser.

Step 2: Install the CLI

pipx install observal-cli

The CLI is still called observal. It pulls agents into your coding tools, publishes components and manages the server.

Step 3: Log In

observal auth login

On a fresh server, the first login on the same machine creates the admin account. Everyone else signs in through the browser with device authorization, or with SSO once you configure it.

Step 4: Capture Sessions

observal doctor patch

This installs session hooks into the coding agents on your machine. From then on, every session is pushed to your server when it ends: prompts, tool calls, MCP round trips, token usage and errors. If a laptop was offline, observal reconcile backfills what was missed.

Step 5: Pull Your First Agent

observal agent pull your-team/pr-reviewer --harness claude-code

ARR resolves every component the agent bundles (MCP servers, skills, hooks, prompts and sandboxes) and writes the right configuration files for the harness you name.

What Gets Collected

• Session metadata: agent, model, harness, user and timestamps
• Tool calls: every Read, Write, Edit, Bash and MCP call, with inputs and outputs
• Token usage: input, output and cache tokens
• Errors: failed tool calls, retries and timeouts
• Timing: per call and per session

All of it stays on your infrastructure.

Supported Coding Agents

Claude Code, Codex CLI, Cursor, Kiro, GitHub Copilot, Copilot CLI, OpenCode, Goose, Pi and Antigravity.

Next Steps

1. Publish your team’s skills and MCP servers, and route them through review before anyone can install them
2. Create a teamspace so each team gets its own namespace
3. Generate an Insights report on an agent once it has a couple of weeks of sessions
4. Invite your team and let them pull approved agents into whichever tool they use

Want a walkthrough for your team? Book a demo.