Choosing an Agent
Fletch doesn’t ship its own model — it drives the agent CLIs you already use and pay for, through one consistent interface. Part of working well in Fletch is picking the right agent for each task.
The agents
Section titled “The agents”Fletch supports Claude Code (the default), Codex, Cursor Agent, and OpenCode, with Pi and Antigravity as experimental options. You choose the agent and its model when you spawn a workspace; it stays fixed for that workspace’s life. Different agents have different strengths, so it’s common to reach for different ones — or to race several on the same task.
The model list is built live: Fletch asks each installed CLI which models it offers and enriches them with context-window and reasoning metadata from an online catalog, so newly released models show up in the picker without an app update.
Two ways to watch an agent
Section titled “Two ways to watch an agent”- Chat view — a normalized, readable transcript of the agent’s reasoning and tool calls. This is the default and works for every agent.
- Native view — the agent’s own raw terminal UI, streamed verbatim. It’s an experimental per-agent toggle, useful when you want the unfiltered TUI.
Tune how an agent behaves
Section titled “Tune how an agent behaves”- Thinking effort — a reasoning cap you set per message. The levels offered depend on the selected model: low / medium / high for most, with some models exposing higher tiers.
- Auto-edit — let the agent apply changes without approving each write.
- Custom agents — save a reusable preset: a name, a base agent and model, a reasoning budget, and standing instructions (a role), plus any Skills and MCP tools it should carry. Your presets show up in the composer alongside the built-in agents.