Skip to main content
Codeg connects to and manages multiple local AI coding agents through a unified interface. Rather than switching between terminal windows for each agent, you run them all from one workspace — with shared conversation history, permission controls, and session management.

​
Supported agents

Codeg reads session data from each agent’s local data directory. The table below shows where each agent stores its data by default. If you set the corresponding environment variable, that path takes precedence.
AgentmacOS / Linux defaultWindows defaultEnvironment variable
Claude Code~/.claude/projects%USERPROFILE%\.claude\projects$CLAUDE_CONFIG_DIR/projects
Codex CLI~/.codex/sessions%USERPROFILE%\.codex\sessions$CODEX_HOME/sessions
OpenCode~/.local/share/opencode/opencode.db%USERPROFILE%\.local\share\opencode\opencode.db$XDG_DATA_HOME/opencode/opencode.db
Gemini CLI~/.gemini%USERPROFILE%\.gemini$GEMINI_CLI_HOME/.gemini
OpenClaw~/.openclaw/agents%USERPROFILE%\.openclaw\agents—
Cline~/.cline/data/tasks%USERPROFILE%\.cline\data\tasks$CLINE_DIR
Environment variables take precedence over the default paths listed above. If you store agent data in a non-standard location, set the corresponding variable before launching Codeg.

​
Connecting an agent

1

Open a folder

Open a project folder in Codeg. The folder appears in the sidebar with its conversation history.
2

Open the agent selector

Click the agent selector in the workspace toolbar. The list shows all supported agents with their current availability status and installed version.
3

Connect

Click Connect next to the agent you want to start. Codeg launches the agent process and establishes an ACP connection. The status indicator moves through connecting → connected once the handshake completes.

​
Session modes

Some agents support multiple session modes — for example, a plan mode where the agent outlines steps before acting, and an auto mode where it executes immediately. When modes are available, a mode selector appears in the toolbar after connecting. Switch modes at any point during a session; the change takes effect on the next turn.

​
Permission control

Agents request permission before executing tools (file writes, shell commands, web requests, and so on). Each request surfaces as a prompt in the workspace with the following options:
OptionBehavior
ApproveAllow this specific tool call and continue
DenyReject the tool call; the agent receives a denial signal
Always ApproveApprove this tool type for the rest of the session without prompting again
You can also respond to permission requests from a connected Chat Channel using /approve or /deny without opening a browser.

​
Session forking

Some agents support forking a session. Forking creates a new session branch starting from a chosen point in the conversation history, so you can explore an alternative approach without losing the original thread. When fork support is available, a Fork button appears in the conversation detail panel. Click it on any turn to branch from that point. The forked session opens as a new conversation alongside the original.

​
Agent configuration

Configure each agent’s environment variables, API keys, model settings, and native config files from Settings > Agents. Codeg reads the agent’s existing config file from disk and presents it in an editable form. Changes are written back to the same file paths the agent uses natively:
  • Claude Code — ~/.claude/settings.json
  • Codex CLI — ~/.codex/config.toml and ~/.codex/auth.json
  • OpenCode — ~/.config/opencode/opencode.json
  • Gemini CLI — ~/.gemini/settings.json
  • Cline — ~/.cline/data/globalState.json and ~/.cline/data/secrets.json
Environment variables set in Settings > Agents are injected into the agent process at connection time. They do not modify your shell environment permanently.

​
Consistent experience across agents

Codeg provides a uniform interface regardless of which agent you use. Every agent in Codeg supports:
  • Sending prompts and receiving streaming text output
  • Real-time tool call status and progress updates
  • Permission request handling with in-workspace approval
  • Session mode selection (where the agent supports it)
  • Session forking to branch from any point in history
This means you can switch between Claude Code, Codex, Gemini CLI, or any other supported agent without relearning the interface — the workspace behaves the same way for all of them.