Backends, Models, and MCP

How Jean chooses a backend, routes model settings, and brings extra tools into a session.

Jean is not locked to a single agent backend.

You can work with Claude CLI, Codex CLI, OpenCode, Cursor Agent, PI, Command Code, Grok, and Kimi Code, then choose how much of that configuration is global, project-level, or session-specific.

Backends

Jean supports eight backends:

Claude CLI
Codex CLI
OpenCode
Cursor Agent
PI (Beta)
Command Code (Beta)
Grok (xAI headless / ACP)
Kimi Code (Beta)

You can set a global default backend and override it per project or per session.

Each backend has different capabilities:

Cursor Agent only supports plan and yolo execution modes
PI and Grok can steer follow-up prompts into a running turn
Command Code runs through the cmd CLI with a CLI-default model option
Grok defaults to managed CLI install and catalog models such as grok/grok-4.5
Kimi Code uses ACP-style sessions with kimi/* model choices

Models and reasoning

Model selection is session-aware.

Depending on the backend, Jean can also store and apply:

thinking level (Claude, Codex, OpenCode)
reasoning effort (Claude Opus adaptive, Grok)
PI provider model choices
Command Code model choices, including the CLI default
Grok model choices (grok/*)
Kimi Code model choices (kimi/*), including CLI default
execution mode

That lets you keep a lighter model for quick work and a heavier one for review, investigation, or tricky implementation.

PI model options come from the active PI provider and are stored as pi/* model choices. Command Code model options are stored as commandcode/* choices; commandcode/default leaves model selection to the cmd CLI. Grok and Kimi follow the same backend/model catalog pattern in the model picker.

Thinking levels

For Claude CLI, Jean exposes four levels with increasing token budgets:

Off - thinking disabled
Think - 4K token budget
Megathink - 10K token budget
Ultrathink - 32K token budget (default)

Higher levels let the model reason more before answering, at the cost of more tokens and latency. Off is appropriate for short generation tasks where reasoning is not useful.

Effort levels (Opus adaptive thinking)

Latest Claude Opus models on CLI 2.1.32+ replace the thinking-level dial with an adaptive effort dial:

Low - minimal thinking, skipped for simple tasks
Medium - moderate thinking
High - deep reasoning (default), almost always thinks
xHigh - extra-high effort (recommended default for coding and agentic work on Opus 4.7)
Max - no constraints on thinking depth

Jean picks the correct control for the selected model automatically.

Adaptive thinking per mode

Jean can apply different reasoning defaults depending on the execution mode.

Plan mode often benefits from heavier thinking. Build and Yolo can run with lighter reasoning to keep the loop fast. The Magic Prompts pane is where those per-workflow overrides live.

Providers

Jean supports custom Claude CLI provider profiles.

This is useful when you want Claude CLI to use a different endpoint or routing setup while keeping the rest of the Jean workflow unchanged.

The important distinction:

backend chooses the tool family
model chooses the model within that backend
provider changes how Claude CLI is routed

MCP servers

Jean discovers MCP servers from the backends you have installed.

It reads backend-specific config locations such as:

Claude config plus .mcp.json
Codex config plus .codex/config.toml
OpenCode config plus opencode.json

You can enable MCP servers:

globally
per project
per session

Newly discovered servers can be auto-enabled, but Jean keeps track of servers you explicitly turned off so they do not silently reappear.

Attachments and extra context

Beyond MCP, a session can include:

file mentions
attached text files
attached images
loaded GitHub context
loaded Linear context
saved context snapshots
linked project context

Jean treats these as part of the session state rather than a one-off paste.

Good defaults

For most teams, a good starting setup is:

one default backend you trust
project-level overrides only where they are truly useful
MCP enabled only for tools that add clear value
heavy models for investigation, lighter models for short generation tasks