-
Notifications
You must be signed in to change notification settings - Fork 0
Run Modes
AgentDesk has explicit run modes so users can tell what will execute and what will remain simulated.
Demo mode is the safest path and the default onboarding mode.
Behavior:
- no API keys required
- no model calls
- no local command execution
- no MCP process execution
- deterministic trace evidence
- simulated costs and artifacts
Use Demo mode to evaluate the debugger experience, replay behavior, artifacts, validation, and export flows.
Ollama mode executes only model nodes with provider: "ollama".
Requirements:
- Ollama running locally
- default endpoint available at
http://127.0.0.1:11434 - selected model pulled locally
Example:
ollama pull llama3.2Behavior:
- Ollama model nodes execute against the local Ollama server.
- MCP nodes remain simulated unless Runtime mode is used.
- Local command nodes remain simulated unless Runtime mode is used.
- Cloud provider nodes remain simulated unless Cloud BYOK mode is used.
Cloud BYOK mode executes configured OpenAI and Anthropic model nodes from the browser session.
Behavior:
- API keys are kept in browser React state for the current session.
- API keys are not saved to localStorage.
- API keys are not included in replay-session exports.
- Prompts and model responses become trace/debug/artifact evidence.
- Provider CORS or organization policy may block browser-direct calls.
Use Cloud BYOK for local experimentation, not production secret handling. Production applications should route provider calls through a backend secret boundary.
Runtime mode uses the packaged loopback CLI. It is required for local command nodes and MCP discovery/tool calls.
Start it with:
npm run build
node ./bin/agentdesk.mjs --port 5173Or from npm:
npx --yes --package=@papaplus/agentdesk agentdesk --port 5173Behavior:
- Runtime API is loopback-only.
- Requests require explicit runtime headers.
- JSON request bodies are size-limited.
- stdout/stderr are capped and redacted.
- MCP discovery is user-triggered.
- Local command execution is explicit and bounded.
Use the least powerful mode that answers the question:
| Need | Mode |
|---|---|
| Understand the UI and debugger | Demo |
| Test local model prompts | Ollama |
| Test OpenAI/Anthropic model nodes with your own key | Cloud BYOK |
| Discover MCP tools or run approved local commands | Runtime |
Switch modes intentionally. A workflow can contain mixed node providers, but only nodes supported by the selected mode execute live.