Models
The engine is chosen per persona on Persona → Engine. The chat model is chosen on that same card when the engine is Single turn. The system LLM is shared and lives under Settings → System LLM.
Choosing a model
Section titled “Choosing a model”Pick the two models separately.
System LLM writes topic titles, the bulletin, summaries, and the other quiet jobs on Settings. It does not call tools. A cheap, fast model is enough. From the picker: gemini-2.5-flash, gemini-2.5-flash-lite, gpt-5-mini, claude-haiku-4-5-20251001, or deepseek-chat. A small local model is fine once Local model server is connected.
Agent engine runs the turn. Single turn, Classic, and Gemini ADK call tools: files, skills, schedules, and the rest of the tool list. Use a model that returns tool calls. claude-haiku-4-5-20251001 and gpt-5-mini can run the agent when you want a cheaper tool-calling model. They fit the system LLM better.
| Use | Models in the picker |
|---|---|
| Everyday agent | claude-sonnet-4-6, claude-sonnet-4-5-20250929, gpt-5.4, gemini-2.5-flash, grok-4.3, deepseek-chat |
| Harder turns | claude-opus-4-7, gpt-5.4, gemini-2.5-pro |
| Gemini ADK | gemini-3.8-flash or gemini-4-argon. |
| Local server | A tool-calling model, then Test tool calling on the engine card. Try qwen2.5-coder:7b (Ollama) or qwen2.5-coder-14b-instruct (llama.cpp). Classic cost routing stays off until that test passes. |
| Cursor (SDK) | A Cursor model. The sidecar default is composer-2.5. |
The live list under the picker can be newer than this page. Choose from that list.
Engines
Section titled “Engines”| Engine | Instruction |
|---|---|
| Single turn | One cloud model for the whole turn. Pick the provider and model on the card. |
| Classic · Cost routing | Local read-only discovery, then delegated sub-jobs. |
| Gemini ADK | Native multi-agent: a coordinator and specialists. |
| Cursor (SDK) | The turn runs through the Cursor sidecar. |
Connect Anthropic, OpenAI, Gemini, xAI, OpenRouter, or DeepSeek under Settings → Connections, or point Local model server at ollama, llama, or llamacpp. Local providers may omit an API key.
The chat model is not the background model. Quiet jobs use the system LLM. If that pair is missing, the job is skipped. The job list and what each one writes is on Settings.
Cursor
Section titled “Cursor”A persona on Cursor talks through the Cursor SDK. Advanced Cursor Configuration sets one Turn timeout for that chat, for schedules, for background jobs, and for an attached cursor_agent CLI run. The sidecar stops the run at the same time. Idle agent handles are closed when the turn ends, and unused ones are dropped after about ten minutes so they do not pile up.
Tmux detach mode is the switch on the cursor_agent CLI tool in that same panel. It does not change a Cursor chat turn.
When the switch is on, a cursor_agent call with detach: true starts cursor-agent in tmux and returns immediately. The reply names the session. Attach with tmux attach -t <session>. The default name starts with meido-cursor. A detached run is not cut by the Turn timeout. Skill creation asks for detach. An attached CLI call still stops at the Turn timeout.
Turn the switch off to refuse detach. Detach is also unavailable in Docker.
Thinking
Section titled “Thinking”Enable extended thinking and Show thinking in replies are on Settings → System LLM. They are shared across personas. A Google chat model sends this config. Other providers ignore it.