Agent Runtime

os-server doesn't run the conversation itself. It hands every turn (voice or chat) to an agent runtime — a separate process that owns the tool loop, memory, skills and channels. The Intern ships with OpenClaw; the Developer Edition can switch to any of these:

Runtime	Value	Service	Talks to the LLM as	Channels
OpenClaw (default)	openclaw	openclaw	OpenAI Chat Completions (custom URL) / catalog format (Autonomous)	Telegram, Slack, Discord, WhatsApp
Hermes	hermes	hermes-gateway	Anthropic Messages	Telegram, Slack, Discord, iMessage
Claude Code	claudecode	claudecode	Anthropic Messages	Telegram, Slack, Discord
Codex	codex	codex	OpenAI Responses	Telegram, Slack, Discord
OpenCode	opencode	opencode	OpenAI Responses	Telegram, Slack, Discord
PicoClaw	picoclaw	picoclaw	Anthropic Messages	Telegram
Remote Hermes	remote	— (Hermes on another machine on your LAN)	—	—

Switch runtime

Don't just edit agent_runtime in config.json. A switch installs the runtime if it's missing, generates its config, starts its service, waits until it's healthy, stops the old one, and only then saves the choice — and rolls back if anything fails. Hand-editing the key skips all of that.

Web UI: http://<device-ip>/setting?debug=true#runtime → Runtime → choose → Switch.

API:

curl -s -X POST $DEVICE/api/device/agent-runtime \
  -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{"runtime":"hermes"}'

The call returns immediately (409 if a switch is already running). os-server restarts at the end, so poll until the new runtime reports ready:

curl -s -H "Authorization: Bearer $TOKEN" $DEVICE/api/device/agent-runtime | jq .data
# → {"current":"hermes","options":[…],"ready":true, …}

The first switch to a runtime downloads and installs it from the internet, so it can take a few minutes. If a switch fails, the log is here:

sudo journalctl -u os-runtime-switch --no-pager -n 200

Remote Hermes

Run Hermes on a bigger machine on your LAN (e.g. your Mac) and let the Intern use it. On that machine:

curl -fsSL https://cdn.autonomous.ai/os/tools/setup-remote-hermes.sh | bash
# prints: Hermes URL : http://<mac-ip>:8642   and an API key

Then point the device at it:

curl -s -X POST $DEVICE/api/device/agent-runtime \
  -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{"runtime":"remote","url":"http://192.168.1.20:8642","token":"<hermes-api-token>"}'

See docs/agentic/remote-hermes.md for the host-side setup.

What carries over when you switch

• LLM settings stay in config.json and are regenerated into the new runtime's own config. Keep editing config.json / the API, not the runtime files (see below).
• Persona, channels and MCP tools are migrated into the new runtime.
• Channels the new runtime doesn't support are skipped and listed in channels_unsupported in config.json (e.g. WhatsApp after switching away from OpenClaw).

Where each runtime keeps its state

Runtime	Config / state on device	Generated from config.json
OpenClaw	/root/.openclaw/openclaw.json, workspace /root/.openclaw/workspace/	models.providers.autonomous.{baseUrl,apiKey,api,models}, agents.defaults.model.primary = "autonomous/<llm_model>"
Hermes	/root/.hermes/config.yaml, /root/.hermes/.env	custom_providers[0].base_url, AUTONOMOUS_API_KEY, model.default
PicoClaw	/root/.picoclaw/config.json, .security.yml, workspace /root/.picoclaw/workspace/	api_base (adds /v1), model, api_keys
Codex	/root/.codex/config.toml, .env, workspace /root/.codex/workspace/	[model_providers.autonomous] base_url, model, OPENAI_API_KEY
Claude Code	/root/.claudecode/.env, workspace /root/.claudecode/workspace/, /root/.claude/	ANTHROPIC_BASE_URL, ANTHROPIC_API_KEY, ANTHROPIC_MODEL
OpenCode	/root/.config/opencode/opencode.json, /root/.opencode/.env	provider campaign → baseURL, {env:LLM_API_KEY}, campaign/<llm_model>

These generated files are marked "managed — do not edit" and are rewritten on every boot and switch. Put your LLM settings in config.json.

Set a model when you leave the proxy

On the Autonomous proxy, the non-OpenClaw runtimes use the model alias Auto-AI, which only exists on the proxy. If you point llm_base_url anywhere else, you must set llm_model too, or every turn fails with a 400.

Subscription login instead of an API key

• Claude Code: set claude_code_oauth_token in config.json (or log in with claude in the Web CLI, which creates /root/.claude/.credentials.json). The runtime then uses your claude.ai subscription and drops the ANTHROPIC_* API settings.
• Codex: if /root/.codex/auth.json exists (from codex login), Codex uses your ChatGPT subscription with the built-in OpenAI provider.

Logs and restart

sudo journalctl -u openclaw -f          # or hermes-gateway, claudecode, codex, opencode, picoclaw
curl -s -X POST -H "Authorization: Bearer $TOKEN" $DEVICE/api/agent/restart

Adding your own runtime

A runtime is one folder under runtimes/<name>/ (with install.sh and presync.sh) plus a case in system/agent/factory.go. See docs/agentic/adding-agent-runtime.md and runtimes/README.md.