Codeg

Codeg is a workspace for running several AI coding agents in one place and having them work together. It imports the session history of every supported agent CLI into one searchable timeline, so a conversation started in one agent can be resumed, read and referenced by another. It runs as a desktop application, a standalone server reachable from a browser, or a Docker container, with native iOS and Android clients that connect back to the machine holding your files and agent CLIs.

Fifteen agents are built in — Claude Code, Codex, Gemini, OpenClaw, OpenCode, Cline, Hermes, CodeBuddy, Kimi Code, Pi, Grok, Cursor, DeepSeek Harness, Qoder and Google Antigravity — and Codeg installs, pins and updates most of them. Any other agent that speaks the Agent Client Protocol can be registered from the public registry or a distribution JSON, after which it appears in the picker, accepts delegation and skills, and has its conversations recorded even if the agent keeps no history of its own.

Delegation happens in the composer: mention an agent, and Codeg launches it as its own session, hands over the task and streams the result back into the current thread. Mention two and they run side by side. When an agent spawns sub-agents of its own, each child gets a card that fills in as it works and can be opened to read the child's session.

Unattended work goes on a to-do board. A task records a title, a description and the agent to run it, and receives its own git worktree on its own branch, so several run in parallel without touching each other or the tree you are editing. A finished task moves to a review column instead of merging itself; on acceptance the agent pulls the base branch into its worktree and resolves conflicts there, and Codeg then checks git rather than the agent's report — a merge it cannot confirm returns to review.

The surrounding workspace is a full engineering loop: an editor with syntax highlighting, live diffs beside the conversation, a git client covering commit, pull, push, stash, branch, merge, rebase, reset and a three-pane conflict editor, an embedded terminal, and several folders linked into one workspace. The conversation area splits into as many tab groups as you want, with layouts and unsent drafts restored on restart.

Features

• Conversation aggregation: sessions from every installed agent import into one searchable workspace and resume where they stopped
• Multi-agent collaboration: mention an agent to delegate a task; sub-agents of different types run as parallel sessions inside one thread
• Custom agents: any Agent Client Protocol agent can be registered from the public registry or its distribution JSON and treated as built-in
• To-dos: queued tasks run unattended, each in its own worktree and branch, landing only after review and a verified merge
• Split view: the conversation area divides into tab groups, with tabs draggable between them and layouts restored on restart
• Git and worktrees: a full client with commit, remote account management, three-pane conflict resolution and parallel worktree flows
• Office documents: agents create and edit .docx, .xlsx and .pptx files through a bundled tool, previewed live in the same pane
• Token usage: trends, cache hit rate, an activity heatmap and breakdowns by folder, agent, model and session
• Chat channels: tasks, permission approvals and live updates over Telegram, Lark and WeChat
• Automations: a saved composer setup runs headlessly on a cron schedule or on demand, starting a session or filing a to-do
• Mobile clients: native iOS and Android apps connect to your desktop or server, with the access token held in the platform keystore
• Error handling: failed turns are classified by kind and offer the matching recovery action, with a reconnect that resumes the session