MyClaws Logo MyClaws

LLM & Credentials

Configure Anthropic / OpenAI / OpenRouter / custom providers; assign dedicated credentials for Codex / Claude Code tools; set up multimodal reading and local reclaude.

Updated Β· Mon Jun 22 2026 00:00:00 GMT+0000 (Coordinated Universal Time)

On this page 9
  1. When you need more than one Connection
  2. The minimum fields for a new Connection
  3. Provider options at a glance
  4. Default / fast / memory model roles
  5. Dedicated credentials for Claude Code / Codex tools
  6. Context window and auto-compaction threshold
  7. Multimodal reading and Web Search credentials
  8. Common authentication pitfalls
  9. Related

LLM & Credentials

In a nutshell: use "LLM Connection" to hold Provider + API Key + context window settings; assign dedicated credentials for Claude Code / Codex tools; configure a fast model for system tasks like title generation.

When you need more than one Connection

  • Use multiple LLM providers simultaneously and pick the right model for each task.
  • Claude Code / Codex subagents need independent credentials to avoid conflicts with the main conversation.
  • Reduce cost for lightweight system tasks like title generation.

The minimum fields for a new Connection

  1. Settings β†’ LLM Connection β†’ Add Connection.
  2. Select a Provider (Anthropic / OpenAI / Google / OpenRouter / Custom).
  3. Enter the API Key and Base URL (if applicable).
  4. Set the default / fast / memory model roles.
  5. Save β€” the connection can be switched instantly on the conversation page.

Provider options at a glance

  • Anthropic: choose "Anthropic API" or "Claude Code" connection type. The latter automatically appends a Claude Code emulation signal, useful for compatible gateways.
  • OpenAI: standard ChatCompletions / Responses compatible endpoint.
  • OpenRouter: enter the OpenRouter routing domain as the Base URL.
  • Google: Gemini series.
  • Custom: any OpenAI-compatible endpoint.

Default / fast / memory model roles

Each Connection can serve multiple roles simultaneously:

  • Default model: used by default in user conversations.
  • Fast model (isFast): reserved for system tasks like title generation.
  • Memory model: used by MemoryAgent.

Fallback order when memory model is unset: memory model β†’ fast model β†’ default model.

Dedicated credentials for Claude Code / Codex tools

The Claude Code and Codex tools can each be assigned a dedicated Connection; credentials are automatically inherited:

  • Claude Code: API Key / OAuth / local reclaude.
  • Codex: API Key or OAuth.

reclaude is a local pseudo-connection that reuses the Claude Code session credentials already logged in on the current host β€” no need to re-enter them.

Context window and auto-compaction threshold

Each Connection can independently configure its context window size. The system uses this to calculate token budgets and trigger auto-compaction. When the compaction threshold (e.g., 80%) is reached, a prompt appears above the input area; manual /compact is also supported.

Multimodal reading and Web Search credentials

Some multimodal models (Claude 3.5 / GPT-4o / Gemini) allow Agents to read images or PDF content directly via the read tool. No extra configuration is needed.

The Web Search tool can be configured per user with Serper / Google / Brave / Bing credentials. Without configuration it falls back to DuckDuckGo (free).

Common authentication pitfalls

Do not mix OAuth and API Key authentication

Having multiple authentication methods active at the same time can trigger 403 Request not allowed. If you use OAuth, clear ANTHROPIC_API_KEY from the relevant environment variables.

reclaude is only available on desktop

It depends on a Claude Code session already logged in on the host machine. Use API Key or OAuth for server / container environments.

Deleting a Connection disconnects all Agents that depend on it

Before deleting, check the Agent list for dependencies to avoid session interruptions.

Continue reading