Skip to content

rLLM UI

Repository: rllm-org/rllm-ui

Web interface for monitoring and analyzing training runs in real time. Think of wandb dedicated to rLLM, with powerful features such as episode/trajectory search, observability AI agent and more. Only supports training runs using Unified Trainer.

Features

  • Real-time Dashboard — Live metrics charts with SSE streaming, multi-experiment overlay with custom colors
  • Episode & Trajectory Inspection — Browse/search episodes, inspect agent trajectories step-by-step (observations, actions, rewards), view it in trajectory groups
  • Training Logs — Live stdout/stderr capture with ANSI color support, search with match navigation
  • Code & Config Visibility — View extracted workflow/agent source code, Hydra config snapshots
  • Observability AI Agent — Query your training data using natural language

Getting Started

There are two ways to access rLLM UI:

  1. Cloud — Use our hosted service at ui.rllm-project.com. No setup required.
  2. Self-hosted — Run locally from the repository (see below).

Regardless of the service you use, add ui to your trainer's logger list in your rLLM training script:

trainer.logger="['console','ui']"

How It Works

rLLM connects to the UI via the UILogger backend, registered as "ui" in the Tracking class (rllm/utils/tracking.py).

On init, the logger:

  1. Creates a training session via POST /api/sessions
  2. Starts a background heartbeat thread (for crash detection)
  3. Wraps stdout/stderr with TeeStream to capture training logs

During training, the logger sends data over HTTP.

So the overall flow looks like: rLLM UI Architecture

Cloud Setup

  1. Sign up at ui.rllm-project.com
  2. Copy your API key (shown once at registration)
  3. Set the key in your training environment (either through export or in .env)

That's it. Run your training script with 'ui' included, and you will see your training runs real-time.

Variable Required Scope Default Description
RLLM_API_KEY Yes Training script env — API key for authenticating training data ingestion (shown once at registration)
RLLM_UI_URL No Training script env https://ui.rllm-project.com Defaults to cloud URL when RLLM_API_KEY is set

AI Agent

The observability AI agent can be enabled from the Settings page in the UI by entering your ANTHROPIC_API_KEY there.

Self-hosted Setup

git clone https://github.com/rllm-org/rllm-ui.git
cd rllm-ui

# Install dependencies
cd api && pip install -r requirements.txt
cd ../frontend && npm install

# Run (two terminals)
cd api && uvicorn main:app --reload --port 3000
cd frontend && npm run dev

Open http://localhost:5173 (or the port shown in the Vite output).

Custom API port

If you run the API on a port other than 3000, update both sides so they know where to find it:

  • rLLM training side — export RLLM_UI_URL="http://localhost:<port>"
  • rllm-ui frontend — set VITE_API_URL=http://localhost:<port> in frontend/.env.development

Database

rLLM UI stores sessions, metrics, episodes, trajectories, and logs in a database so they persist across restarts and are searchable.

  • SQLite (default) — No setup required. A local file (api/rllm_ui.db) is created on first run.
  • PostgreSQL — Adds full-text search with stemming and relevance ranking. Set DATABASE_URL in api/.env:
DATABASE_URL="postgresql://user:pass@localhost:5432/rllm"

Observability AI Agent

rLLM UI includes a built-in AI agent that can query your training data using natural language. Currently experimental — more support coming soon. To enable it, set your Anthropic API key in api/.env:

ANTHROPIC_API_KEY="sk-ant-..."

Configuration

Variable Required Scope Default Description
RLLM_UI_URL No Training script env http://localhost:3000 URL of your local rllm-ui server
DATABASE_URL No api/.env SQLite PostgreSQL connection string. Defaults to SQLite if unset.
ANTHROPIC_API_KEY No api/.env — Enables the built-in AI agent
VITE_API_URL No frontend/.env.development http://localhost:3000 Only needed if the API runs on a non-default port