arc-relay 0.1.5

ARC Relay — WebSocket relay server for agent remote control by Axolotl AI

ARC — Agent Remote Control

Universal remote control for AI agent frameworks. Connect any agent to a browser-based interface for real-time monitoring and remote interaction.

Agent Support

Agent	Status	Viewer → Agent	Agent → Viewer	How
Hermes Agent	✅ Supported	✅ Messages	✅ Tool calls, responses, status	Native plugin + /remote-control
DeepAgent	🔜 Planned	—	—	@axolotlai/arc-adapter-deepagent
OpenClaw	🔜 Planned	—	—	@axolotlai/arc-adapter-openclaw
Claude Code	🔜 Planned	—	—	Skill file

Quick Start

# Install and configure
curl -fsSL https://raw.githubusercontent.com/axolotl-ai-cloud/arc/refs/heads/main/install.sh | sh
arc setup --hermes  # uses hosted beta relay — no server needed

# Start Hermes, then type /remote-control
hermes

The /remote-control skill starts a session, opens the viewer in your browser, and streams all tool calls and responses to the viewer in real-time.

Note: arc setup --hermes patches the hermes entrypoint to enable bidirectional viewer ↔ agent messaging. If hermes update overwrites the patch, run arc update to re-apply it. Once NousResearch/hermes-agent#3778 is merged, the patch is unnecessary.

How It Works

Agent (Hermes, DeepAgent, OpenClaw, any)
  │
  │  WebSocket
  ▼
Relay Server (arc-beta.axolotl.ai or self-hosted)
  │
  │  WebSocket
  ▼
Web Viewer (browser)
  • Live trace view (tool calls, messages, status)
  • Send messages to the agent
  • Approve/deny tool calls

Traces flow agent → relay → viewer. Commands flow viewer → relay → agent.

Self-Hosted Relay

For users who want to run their own relay server:

# Option 1: Docker
docker compose up relay

# Option 2: Python directly
pip install -r relay/requirements.txt
python -m relay

# Option 3: Use the CLI (handles everything)
arc setup --self-hosted

Environment Variables

# Required — set a fixed token, or one is auto-generated at startup
export AGENT_TOKEN="your-strong-random-token"

# Optional
export PORT=8600                        # Listen port (default: 8600)
export MAX_SESSIONS=100                 # Max concurrent sessions
export MAX_TRACE_LOG=2000               # Max trace events per session
export SESSION_TTL_HOURS=24             # Idle session cleanup
export ALLOWED_ORIGINS="*"              # CORS origins (comma-separated)
export REQUIRE_TLS=false                # Reject non-TLS connections
export AGENT_TOKEN_PREFIX=""            # Accept prefix-based tokens (beta mode)

Running with ARC_ENV=dev

For local development, use ARC_ENV=dev to default to self-hosted mode:

# arc setup will default to localhost relay, auto-start it
ARC_ENV=dev arc setup --hermes

# Or start the relay manually
AGENT_TOKEN=dev-token python -m relay

# Then configure the CLI to use it
ARC_ENV=dev arc setup --self-hosted

Verify the relay is running

curl http://localhost:8600/health
# {"status": "ok", "sessions": 0}

Open http://localhost:8600/viewer to access the web viewer (if web-client is built).

Supported Frameworks

Framework	Integration	How
Hermes Agent	Native Python plugin	/remote-control or arc_start_session tool
DeepAgent (LangChain)	TypeScript middleware	@axolotlai/arc-adapter-deepagent
OpenClaw	TypeScript plugin	@axolotlai/arc-adapter-openclaw
Claude Code	Skill file	/remote-control slash command

CLI Commands

arc setup [--hermes|--self-hosted]  # Configure relay URL, token, framework
arc config [--viewer|--relay|...]   # Get or set individual config values
arc update                         # Pull, rebuild, reinstall skills
arc install-skill                  # Install /remote-control skill
arc status                         # Show current configuration

Choosing a Viewer

By default, viewer links point to the viewer bundled into your relay server — the same built-in React app served at /viewer on whatever relay you're connected to. This is the simplest setup and requires no extra configuration.

There are two reasons you might want a different viewer URL:

Use the OSS viewer on GitHub Pages — useful if you want to verify the viewer source code independently of the relay operator. The viewer is deployed directly from this repo via GitHub Actions and can be inspected at any commit:

arc config --viewer https://axolotl-ai-cloud.github.io/arc
# When a session starts: https://axolotl-ai-cloud.github.io/arc?relay=https://arc-beta.axolotl.ai&session=...&s=...

Use a locally running viewer — useful when developing the web client or running fully air-gapped:

# In one terminal: start the dev viewer
npm run dev:viewer   # serves at http://localhost:5173

# In another terminal: point arc at it
arc config --viewer http://localhost:5173
# When a session starts: http://localhost:5173?relay=http://localhost:8600&session=...&s=...

Reset to the default relay-bundled viewer at any time:

arc config --viewer default

The viewer URL is stored in ~/.arc/config.json and can also be set via environment variable:

ARC_VIEWER_BASE=https://axolotl-ai-cloud.github.io/arc

Security Model

• Agent token — required to register sessions. Prevents unauthorized agents.
• Session secret — generated per-session, required for viewer subscription. Only share with trusted parties.
• Role enforcement — viewers can't send traces, agents can't send commands.
• Rate limiting — per-IP limits on HTTP (60/min) and WebSocket (120 msg/min).
• TLS — set REQUIRE_TLS=true in production.
• Prefix tokens — beta mode (AGENT_TOKEN_PREFIX) gives each agent its own session namespace.

Extensibility

The relay is extensible via protocol interfaces:

from relay import create_app, RelayConfig

app = create_app(RelayConfig(
    auth=MyAuthProvider(...),      # Authenticate agents, viewers, admins
    store=MySessionStore(...),     # Persist session state
    policy=MySessionPolicy(...),   # Enforce session limits
    hooks=MyLifecycleHooks(...),   # React to session events
))

Project Structure

├── relay/                     # OSS relay server (Python/FastAPI)
├── packages/
│   ├── protocol/              # Wire types, client, crypto
│   ├── cli/                   # `arc` CLI (setup, connect, update)
│   ├── web-client/            # React SPA viewer
│   ├── adapter-hermes/        # Hermes Agent TypeScript adapter
│   ├── adapter-deepagent/     # DeepAgent adapter
│   └── adapter-openclaw/      # OpenClaw adapter
├── hermes-plugin/             # Native Hermes Python plugin
├── tests/                     # Integration tests
├── install.sh                 # curl | sh installer
└── docker-compose.yml         # Local dev

Testing

npm run test:ts                # TypeScript tests
npm run test:py                # Python tests (needs pytest, pytest-asyncio, httpx)
npm test                       # All tests

Roadmap

• Per-session authentication
• CLI with arc setup / /remote-control
• Tool approval/deny from viewer
• Native Hermes Agent plugin with lifecycle hooks
• Beta relay with prefix-based tokens
• Web viewer auto-connect from URL params
• Viewer → agent message injection (pending Hermes PR)
• Mobile app (iOS/iPadOS)
• Multi-agent dashboard view
• Persistent trace storage

License

MIT. See LICENSE.