Skip to main content

​
CLI

OpenRTC ships a console script named openrtc (Typer + Rich) for discovery-based workflows. The Typer application and main() live in openrtc.cli.commands (with helpers in openrtc.cli.livekit, openrtc.cli.dashboard, openrtc.cli.reporter, openrtc.cli.types, and openrtc.cli.params). The lazy entrypoint and missing-extra hints are in openrtc.cli (the package __init__ plus openrtc.cli.entry). The programmatic entry is typer.main.get_command(app).main(...) (Click’s Command.main), not the test-only CliRunner.

​
Installation

The library (AgentPool, discovery, routing) installs with: 
pip install openrtc
The CLI stack (Typer, Rich) is the optional extra cli: 
pip install 'openrtc[cli]'
If Typer/Rich are not importable, openrtc.cli:main exits with 1 and prints an install hint.

​
Typical usage

  1. Set the same variables as a standard LiveKit worker: 
    export LIVEKIT_URL=ws://localhost:7880
export LIVEKIT_API_KEY=devkey
export LIVEKIT_API_SECRET=secret
  2. Run a worker subcommand with an agents directory (plus any provider defaults your agents need). You can pass --agents-dir or use the first positional argument on start / dev / console. A second positional is optional and only sets --metrics-jsonl when you want JSONL metrics; skip it if you only need the agents directory (unless you already passed --metrics-jsonl). 
    openrtc dev ./agents
openrtc dev ./agents ./openrtc-metrics.jsonl
# equivalent to:
openrtc dev --agents-dir ./agents --metrics-jsonl ./openrtc-metrics.jsonl
openrtc start --agents-dir ./agents
Defaults are conservative: no Rich dashboard unless you pass --dashboard, and no metrics files unless you pass --metrics-json-file or --metrics-jsonl. Refresh intervals default to 1 second where applicable. Subcommands mirror the LiveKit Agents CLI (python agent.py dev, start, console, connect, download-files). OpenRTC adds --agents-dir for discovery, then delegates to livekit.agents.cli.run_app. OpenRTC-only flags are stripped from sys.argv before that handoff so LiveKit does not see options like --agents-dir.

​
Connection overrides

You can override LIVEKIT_* per invocation:
  • --url
  • --api-key
  • --api-secret
  • --log-level (also LIVEKIT_LOG_LEVEL)
These appear under Connection or Advanced in --help depending on the flag.

​
Commands

Across list, connect, download-files, and start / dev / console, you can often pass paths positionally instead of --agents-dir or --metrics-jsonl (see each command below). The first non-flag token after the subcommand is rewritten before parsing; use --agents-dir when you need a different argument order.

​
openrtc list

Discovers agent modules and prints each agent’s resolved settings.
  • Default: Rich table (human-friendly).
  • --plain — Stable, line-oriented text (no ANSI/table borders); good for grep and CI.
  • --json — Stable JSON. Top-level fields include schema_version (bump when the shape changes) and command: "list". Combine with --resources for resource_summary.
  • --plain and --json together are rejected (non-zero exit).
  • --resources — Footprint and memory hints (grouped under Advanced in --help).
openrtc list ./agents
openrtc list --agents-dir ./agents --plain
openrtc list ./agents --json

​
openrtc start / openrtc dev / openrtc console

Worker subcommands. start is production-style; dev adds reload; console is a local-only sandbox. All three accept the same option shape (see --help per command for the full list). 
openrtc start ./agents
openrtc dev ./agents
openrtc console ./agents

​
Coroutine-mode runtime knobs (v0.1)

Two flags pick the worker isolation mode and the coroutine-mode backpressure threshold. They show up under the OpenRTC panel in --help and are accepted by start, dev, and console.
  • --isolation — coroutine (default) or process.
    • coroutine runs every session as an asyncio.Task inside one worker process and shares the prewarmed VAD / turn detector across sessions. Use this when density is the goal.
    • process keeps the v0.0.x behavior of one OS subprocess per session via livekit-agents’s default ProcPool. Use this when you need hard process isolation between sessions or per-session memory caps via livekit-agents’ job_memory_limit_mb.
  • --max-concurrent-sessions — Integer ≥ 1, default 50. The coroutine pool reports load >= 1.0 to LiveKit dispatch once this many sessions are in flight, so the dispatcher routes new jobs elsewhere. Ignored under --isolation process (livekit-agents’ own load math applies there).
# Default: coroutine mode at 50 concurrent sessions per worker.
openrtc start ./agents

# Opt back into v0.0.x behavior:
openrtc start ./agents --isolation process

# Tune the coroutine threshold for a 1-vCPU / small-memory host:
openrtc start ./agents --max-concurrent-sessions 12
See docs/concepts/architecture.md for the coroutine-mode lifecycle and the README’s “Isolation modes” section for a side-by-side comparison table.

​
openrtc connect

Connect the worker to an existing room (LiveKit connect). Requires --room. 
openrtc connect ./agents --room my-room

​
openrtc download-files

Download plugin assets (LiveKit download-files). Only needs the agents directory (for a valid worker entrypoint) plus connection settings—no --default-stt / --default-llm / --default-tts / --default-greeting. 
openrtc download-files ./agents

​
Runtime visibility and automation

  • --dashboard — Live Rich summary (RSS, sessions, routing, savings estimate). Off by default.
  • --metrics-json-file PATH — Overwrites a JSON file each tick with the latest PoolRuntimeSnapshot (good for scripts). Grouped under Advanced.
  • --metrics-jsonl PATH — Appends versioned JSON Lines (truncates when the worker starts). Each line is one record: schema_version, kind (snapshot or event), seq, wall_time_unix, payload. Snapshots match PoolRuntimeSnapshot.to_dict(); events carry session lifecycle hints (session_started, session_finished, session_failed). Tail it with tail -f ./openrtc-metrics.jsonl, pipe it through jq, or point any log shipper or script at it.
  • --dashboard-refresh — Interval in seconds for dashboard, metrics file, and JSONL when --metrics-jsonl-interval is not set (Advanced).
  • --metrics-jsonl-interval — Override JSONL cadence only (Advanced).

​
Shared default options (discovery)

Worker commands that load agents accept optional defaults applied when a discovered agent does not override them via @agent_config(...):
  • --default-stt
  • --default-llm
  • --default-tts
  • --default-greeting (Advanced)
Example: 
openrtc list \
  --agents-dir ./examples/agents \
  --default-stt openai/gpt-4o-mini-transcribe \
  --default-llm openai/gpt-4.1-mini \
  --default-tts openai/gpt-4o-mini-tts
These defaults are passed through to livekit-agents as raw strings. For provider-native plugin objects, configure them in Python with AgentPool instead of through the CLI flags.

​
list --resources (footprint)

With --resources, list adds:
  • Per-agent on-disk size of the discovered .py module when the path is known (see AgentConfig.source_path in the API docs).
  • Summary — total source bytes and a best-effort process memory metric from openrtc.observability.metrics (Linux: current VmRSS; macOS: peak ru_maxrss, not live RSS—see resident_set.description in --json output).
  • Savings estimate — a transparent estimate of the memory saved by one shared worker versus one worker per registered agent.
openrtc list --agents-dir ./examples/agents --resources
openrtc list --agents-dir ./examples/agents --resources --json

​
Notes

  • --agents-dir is required for every command.
  • list returns a non-zero exit code when no discoverable agents are found.
  • Worker commands discover agents before handing off to the LiveKit CLI.
  • The live dashboard, --metrics-json-file, and --metrics-jsonl reflect the running shared worker, unlike list --resources, which reflects the short-lived CLI discovery process.

​
Prove the shared-worker value locally

  1. Discover your agents: 
    openrtc list --agents-dir ./examples/agents --resources
  2. Start one shared worker with the dashboard and/or metrics output: 
    openrtc dev \
  --agents-dir ./examples/agents \
  --dashboard \
  --metrics-json-file ./runtime.json
    Or enable a streaming JSON Lines feed: 
    openrtc dev \
  --agents-dir ./examples/agents \
  --metrics-jsonl ./openrtc-metrics.jsonl
  3. Watch the dashboard (or tail the JSONL file with tail -f ./openrtc-metrics.jsonl, optionally piped through jq) for worker RSS, active sessions, routing, and errors.
  4. Use runtime.json or the JSONL stream for automation or scraping.
For production capacity planning, compare these snapshots with host or container telemetry from your deployment platform.