Codeg Multi-Agent Coding Workspace
Aggregate sessions from Claude Code, Codex, OpenCode, Pi, Grok Build, and more into one searchable workspace, and let them collaborate via `@` delegation.
Evidence shows: README claims local-first, network access only on user-triggered actions, token-based auth for web/server modes, and mobile tokens stored in system keychain. However, implementation details are lacking, such as permission model, user confirmation flows, data flow transparency, sensitive data handling, dependency security audits, external effect controls, rollback mechanisms, and source attribution. Thus each criterion scores 1 due to insufficient evidence or mere assertions without implementation support.
Evidence shows: CI tests exist (frontend lint, vitest, build, Rust tests and clippy) with detailed test configuration, indicating good internal consistency. However, dependency availability is not explicit, and failure messages only mention partial error categorization, not comprehensive. Therefore self_consistency scores 2, dependency_availability and failure_messages score 1.
Evidence shows: README describes multiple usage scenarios (desktop, server, Docker, mobile), supports multiple agents, has custom agent capability, and broad environment fit. However, trigger precision (e.g., @ mentions) is only briefly described without detailed boundaries. Thus audience_and_scenarios, capability_boundaries, environment_fit score 2, trigger_precision scores 1.
Evidence shows: README is well-structured, has installation instructions, multi-language docs, license (Apache-2.0), but lacks version changelog, known limitations, and maintenance responsibility. Naming stability is not explicit. Therefore information_architecture, install_notes, examples_and_faq, license score 2, others score 1.
Evidence shows: Output usability (e.g., Office documents, live preview) and marginal value (multi-agent collaboration, to-dos) are well described, but cost-benefit is not quantified, only sponsors mentioned. Thus output_usability and marginal_value score 2, cost_benefit scores 1.
Evidence shows: README claims have documentation links, but no specific evidence or test results provided. Cross-source corroboration is insufficient, and facts vs inferences are not clearly separated. Thus each criterion scores 1.
- Static review, no code execution, all conclusions based on file contents, low confidence.
- Publisher identity unverified, do not infer security based on brand.
- Security claims in README (e.g., local-first, token auth) lack implementation details, need further verification.
- Dependency security not audited, need to check for vulnerabilities.
- Version changelog and known limitations missing, may affect user decisions.
What does this agent do, and when should you use it?
Codeg (Code Generation) is a multi-agent coding workspace that aggregates sessions from all supported AI coding agents (e.g., Claude Code, Codex, OpenCode) into a unified, searchable workspace, and allows a main agent to delegate to sub-agents of other types within a single task. It runs as a desktop app, a standalone server, or a Docker container, with native iOS and Android clients for on-the-go access. Fourteen agents come built in, and you can register any other ACP-compatible agent. It includes a to-do board where each task runs in its own git worktree, pending your review before merging.
Codeg aggregates sessions from every installed agent, with one-click import of past history, searchable and resumable. Users delegate tasks via @ mentions; Codeg schedules each mentioned agent as its own session, running in parallel and streaming results. The to-dos feature creates a separate git worktree for each task on its own branch, running unattended, then moving to review. The built-in workspace provides a file tree, diff viewer, and full git client (commit, pull, merge with three-pane merge editor). Office document creation/editing (.pptx/.docx/.xlsx) is supported with live preview. MCP and skills management, automations, chat channels (Telegram, Lark, WeChat), and token usage reporting are included.
- A developer managing multiple AI coding agents simultaneously and wanting a single interface for their session histories.
- A team wanting to parallelize coding tasks: Claude Code drafts while Codex reviews, via
@delegation. - Automating background tasks like code refactoring or test generation, each in its own worktree, with human review before merging.
- A user needing to review or approve agent permissions from a mobile device while away from the desk.
- A professional generating or editing Office documents (e.g., decks, reports) with real-time preview.
- A developer working across multiple related repositories (shared library, docs) that should be treated as one workspace.
What are this agent's strengths and limitations?
- True multi-agent collaboration: delegate via
@to different agent types running in parallel within one conversation, with sub-agent sessions observed live. - Unified session aggregation: import and search historical sessions from all installed agents, resumable, and cross-agent context via
@mentions. - Advanced workflow tooling: to-dos with git worktrees for isolated parallel tasks, automations, and rich Office document editing with live preview.
- Flexible deployment: desktop app, self-hosted server, Docker, and open-source mobile clients (iOS, Android) with remote access.
- Complexity: a steep learning curve due to its many features, especially the workspace, to-dos, and advanced git operations.
- Dependency on external agent CLIs: each agent (e.g., Claude Code, Codex) must be installed and configured separately; Codeg only orchestrates them.
- Resource and network demands: self-hosting a server requires network access, and mobile usage depends on the Web Service being reachable.
- Documentation depth: some advanced features (e.g., scientific research skills) may require extra setup and have dependencies.
How do you install or deploy this agent?
Desktop: download the installer for macOS, Windows, or Linux from GitHub Releases and follow the installation docs. Server (Linux/macOS): curl -fsSL https://raw.githubusercontent.com/xintaofei/codeg/main/install.sh | bash, then CODEG_STATIC_DIR=/usr/local/share/codeg/web codeg-server. Windows (PowerShell): irm https://raw.githubusercontent.com/xintaofei/codeg/main/install.ps1 | iex, then set $env:CODEG_STATIC_DIR="$env:LOCALAPPDATA\codeg\web" and run codeg-server. Docker: docker run -d -p 3080:3080 -v codeg-data:/data ghcr.io/xintaofei/codeg:latest. Mobile: install the iOS app from the App Store or the Android APK from the repository, then point it to your desktop Web Service or your own codeg-server.
How do you use this agent?
After installation, first launch Codeg, which auto-detects installed agents (like Claude Code, Codex) or lets you register custom ACP-compatible agents from the settings. In a conversation, type @ and select the agent to delegate to, then enter the task; Codeg handles scheduling and streams results back. For to-dos, click 'Add Task', fill in title, description, and choose the agent; it automatically creates a git worktree and runs. Access via desktop app, server, or Docker; mobile clients require pairing with your server URL and access token. Detailed usage is in the official docs at docs.codeg.app.