Agent backends and models

What the Agent section connects

The Agent section selects the backend and model used for the next agent turn. 9Miho asks its server which backends are installed, signed in, and available, then shows the server’s exact status and recovery detail.

This is separate from Providers. Agent backends reason about the current production context or help author work. Provider profiles authorize image, video, and audio execution. Connecting ChatGPT, Claude, or another agent never grants permission to spend against a generation API.

The currently registered backend families are Claude, ChatGPT (Codex CLI), OpenCode, Cursor, and Ollama. The running server’s Agent list is the source of truth.

This page describes an Agent that 9Miho launches from inside the application. To start from Claude Code or Codex and operate a running 9Miho in the other direction, use Claude and Codex through MCP. The two connections share the same production contracts but are configured separately.

Install and sign in where the server runs

Run these commands on the same machine and under the same operating-system account that runs the 9Miho server. If the browser is connected to a remote server, installing a CLI on the browser machine will not make it available to 9Miho.

ChatGPT through Codex CLI

# macOS
brew install codex
# or, on macOS / Linux / Windows
npm install -g @openai/codex

codex login

On native Windows, run 9Miho from PowerShell under the same Windows account that completed codex login.

Claude

npm install -g @anthropic-ai/claude-code
claude
# then run /login inside Claude Code

On Windows, sign in first and restart the 9Miho server. The user-scoped OAuth environment may not be visible to a server process that was already running.

OpenCode

npm install -g opencode-ai
opencode providers login
# or run: opencode

Cursor

Install the CLI from Cursor’s official installation guide. Settings can show an OS-specific one-line installer, but verify it against the current official guide before executing a network-delivered script. Then verify the executable and use Cursor’s own login flow:

agent --version
agent login

Cursor documents the authentication flow separately in its CLI authentication reference.

Ollama

Install Ollama from ollama.com/download, then pull a model:

ollama pull qwen3-coder

Ollama must remain running where the 9Miho server runs. The model choices shown in Settings come from the server backend report.

Choose a backend and model

1. Open Settings → Agent.
2. Select a backend.
3. Read its availability and authentication detail.
4. Follow the OS-specific installation and login commands if needed.
5. Press Connect or Reconnect. Both actions re-read /api/agent/backends; login itself still happens in your terminal.
6. Choose one of the models reported for that backend.

The selection is remembered in this browser and applies to the next turn. It does not silently migrate another browser or another server installation.

Prompt assistance

Prompt assistance lets Storyteller and generation surfaces use the selected agent for supported authoring help. It does not let an agent approve Canon, authorize a paid Run, or bypass validation. Agent proposals remain reviewable inputs to the creator’s workflow.

Canvas also distinguishes the Agent panel from an Assistant node. The panel coauthors and inspects the Flow; an Assistant node is an executable node inside the graph. See Canvas overview and Direction and model Dialects.

If the status stays unavailable

• Confirm the CLI works in a fresh terminal under the server account.
• Confirm its executable is on that account’s PATH.
• Complete the backend’s own login flow; do not paste login tokens into Canvas.
• Restart the 9Miho server when a login changes user-scoped environment state.
• Press Connect or Reconnect again and read the new server detail.