Skip to content

LLM Configuration

karurikwao edited this page Jul 5, 2026 · 1 revision

LLM Configuration

AgentDesk can execute configured OpenAI and Anthropic model nodes in Cloud BYOK mode.

Providers

Supported cloud model providers:

  • OpenAI Responses
  • Anthropic Messages

The UI includes provider/model dropdowns plus editable base URL and model fields.

API Key Handling

Cloud BYOK keys are held in browser session state only.

They are not:

  • saved to localStorage
  • written into replay-session exports
  • written into trace bundle ZIP files
  • intentionally included in debug payloads

However, prompts and model responses are trace evidence. Do not put secrets into prompts, labels, tool payloads, stdout/stderr, screenshots, or model responses.

OpenAI Path

Typical flow:

  1. Pick a workflow with OpenAI model nodes, such as Local Research Agent.
  2. Open LLMs.
  3. Select OpenAI Responses.
  4. Choose a model preset or enter a model ID.
  5. Paste your API key.
  6. Click Use Cloud mode.
  7. Click Apply to nodes if you want model nodes updated.
  8. Click Run BYOK cloud.

Anthropic Path

Typical flow:

  1. Pick a workflow with Anthropic model nodes, such as Repo QA Swarm.
  2. Open LLMs.
  3. Select Anthropic Messages.
  4. Choose a model preset or enter a model ID.
  5. Paste your API key.
  6. Click Use Cloud mode.
  7. Click Apply to nodes.
  8. Click Run BYOK cloud.

Browser-Direct Caveats

Cloud BYOK mode currently calls providers directly from the browser tab.

This can fail because of:

  • provider CORS policy
  • organization security policy
  • browser extension interference
  • invalid API key
  • unsupported model ID
  • network restrictions

Production applications should use a backend proxy or server-side secret boundary. AgentDesk keeps BYOK mode intentionally local and explicit.

Evidence Handling

When a cloud node runs successfully, AgentDesk may capture:

  • prompt text
  • model response text
  • token usage
  • cost estimate
  • provider/model metadata
  • error payloads if the provider returns one

This is useful for debugging, but it also means model prompts and responses should be treated as evidence, not as private vault contents.

Clone this wiki locally