Codexia
A lightweight agent workstation for Codex CLI, Claude Code, and any ACP agent, bundling a task scheduler, git worktree management, and remote control in one workspace.
- Source repo
- milisp/codexia
- Stars
- ★ 941
- Last updated
- today
- License
- MIT
- Primary language
- TypeScript
- FA score
- 57/100 · Major gaps
At a glance
- How it runs
- Works with
- Portable with changesCodex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Developers using Codex CLI who want a GUI to manage multi-project agent sessions and approvals instead of terminal-only interaction.
- Not a fit if
- Users without Codex CLI or Claude Code CLI installed
- Teams unwilling to secure the exposed remote web server themselves
- Users wanting out-of-the-box MCP/ACP with no configuration
- Source review
- 57/100 · Major gaps
What does this agent do, and when should you use it?
Codexia is an open-source desktop application (MIT licensed) that gives OpenAI Codex CLI, Claude Code, and any Agent Client Protocol (ACP) agent a unified working interface. The desktop backend is built on Tauri v2 + Rust, with a React + TypeScript + Zustand + shadcn/ui frontend. It manages session and turn lifecycles through Codex app-server JSON-RPC integration, and ships an Axum headless web server for remote control with real-time updates via a /ws WebSocket broadcast. The workspace includes git worktree management, a project file tree, an IDE-like editor, a prompt notepad, and local web preview, plus one-click PDF/XLSX/CSV previews, an MCP server marketplace, and an agent skills marketplace. All data stays local, agents run in isolated processes, and file and network access can be configured per agent.
Codexia reads a user-selected project directory, launches agent sessions via locally installed Codex CLI and Claude Code CLI, and tracks session/turn, model, and approval state through the Codex app-server JSON-RPC interface; tool calls render live in the thread and sessions persist across restarts. In headless mode, an Axum web server exposes a browser-accessible API: /health, /ws, /api/codex/thread/*, /api/codex/turn/*, /api/codex/model/*, /api/codex/approval/*, /api/automation/* (create/update/run/pause/delete for scheduled jobs), /api/filesystem/*, /api/git/*, /api/terminal/*, /api/cc/*, /api/notes/*, and more. It also manages git worktrees, previews PDF/XLSX/CSV files, provides MCP server and skills marketplaces, and offers a usage analytics dashboard.
- Developers using Codex CLI who want a GUI to manage multi-project agent sessions and approvals instead of terminal-only interaction.
- Engineers who need recurring agent jobs, using the Agent Task Scheduler for automation.
- Users working remotely, including from iOS, controlling agents through the headless web server in a browser.
- Developers doing parallel work who need git worktree management alongside an IDE-like editor.
- Teams using external ACP-compliant agents who want persisted sessions and live tool-call rendering in the UI.
- Privacy-conscious users relying on process isolation and per-agent permission control for local file and network access.
How do you install or deploy this agent?
Prerequisites: install Codex CLI (github.com/openai/codex) and Claude Code CLI (claude.ai/code); optionally any agent that speaks ACP. Install Codexia: on macOS via Homebrew: brew install --cask codexia; on Windows via Scoop: scoop bucket add milisp https://github.com/milisp/scoop-bucket && scoop install codexia; or grab prebuilt releases for macOS/Linux/Windows from GitHub Releases, including a portable Windows archive (codexia_<version>_x64_portable.zip) that runs without installing.
How do you use this agent?
- Launch Codexia. 2. Add your project directory. 3. Enter a prompt and start an agent session. 4. For recurring workflows, create an Agent Task Scheduler job. In headless/remote mode, the Axum web server lets browsers interact via the /ws WebSocket and /api/codex/* HTTP endpoints. Contributors can add API handlers under ./web/src/handlers/, register routes in ./web/src/router.rs, and add frontend calls in src/services/tauri/.
What are this agent's strengths and limitations?
- Unified support for Codex CLI, Claude Code, and any ACP agent in a single interface covering multiple agent runtimes.
- True remote control via the headless Axum server and /ws WebSocket, documented to work from browsers including iOS.
- Task scheduler for recurring jobs, plus git worktree management, an MCP marketplace, and a skills marketplace in one app.
- Clear security model: process-isolated agents, per-agent file/network permissions, all data stored locally.
- Core functionality depends on locally installed Codex CLI and Claude Code CLI; missing one CLI disables the corresponding agent features.
- Advanced features like MCP and ACP require users to understand and configure the relevant agents themselves; limited out-of-the-box scope.
- The remote web mode exposes filesystem, git, and terminal APIs; securing it on a public network is the user's responsibility.
- Deployment and authentication details for the headless server and scheduler live in docs/WEB_SERVER.md and other docs, not fully covered in the README.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Codexia This agent | 57 · Major gaps | Desktop appFree + model costs | ★ 941 | today | TypeScript | Codex · Claude Code |
| OpenTeams | 56 · Major gaps | Desktop appFree + model costs | ★ 580 | 9d ago | TypeScript | Codex · Claude Code |
| Delegatus Agent Workbench | 55 · Major gaps | CLI | ★ 33 | today | TypeScript | Codex · Claude Code |
| MetaBot Personal Agent Workspace | 51 · Major gaps | Self-hosted serviceFree + model costs | ★ 997 | 1mo ago | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
README claims process isolation, per-agent file/network permission configuration, and local-only data, but offers no permission-model detail, approval flow, or authentication story for the headless server; the API surface includes high-risk endpoints (/api/terminal/*, /api/git/*, /api/filesystem/*) with no evidence of auth or user confirmation, so least_privilege and user_confirmation get 1; no rollback mechanism described. Attribution to milisp is clear (2).
Versions are consistent (0.50.1 in both Cargo.toml and package.); dependencies are mainstream with one git dep pinned by tag, but codex-finder is unpinned and no lockfile evidence is shown; CI covers macOS/Linux builds. No user-facing failure-message documentation, so failure_messages gets 1.
Audience is well-defined (desktop workstation for Codex CLI + Claude Code + ACP users) with multi-OS install paths and requirements. However, capability boundaries (scheduler permissions, remote-control exposure) and trigger precision (scheduler behavior) are undocumented (1 each).
Information architecture is excellent: clear README, architecture docs, contributing guide, multi-platform install notes; MIT license complete. Missing: FAQ/examples, known-limitations section, CHANGELOG; naming stability and maintenance responsibility rest on a README 'Maintained by' statement (2).
Aggregating multiple CLI agents with scheduler/worktree/remote control gives clear marginal value and well-documented entry points; output usability depends on undocumented UI details. Cost-benefit is reasonable, but headless mode widens attack surface without matching security guidance.
README claims cross-check against code layout (crates/*, web/src/handlers, src/services/tauri) and named entry points, which is good; but the security claims (process isolation, permission control) lack supporting links or code excerpts in the evidence. Fact/inference separation is mostly maintained (2).
- Headless/web mode exposes terminal, git, and filesystem endpoints with no evidence of authentication in the provided files; do not enable remote control on untrusted networks.
- README security claims (process isolation, permission control) lack code-level corroboration in the evidence; audit independently before using with sensitive projects.
- Some Rust dependencies come from git repos by the same unverified author with unpinned revisions (codex-finder), a supply-chain risk.
- The automation scheduler runs agent jobs on a recurring basis; its permission scope and failure handling are undocumented — verify behavior before enabling.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
Features & Improvements - Codex: a new Agents tab in the right panel lists the sub-agents a thread spawned, with status, prompt, latest reply and a Stop action while they run. - Codex: stopping a turn also ends the commands it left running as background terminals; terminals you started yourself are left alone. - Sidebar: an Active section shows running sessions and sessions waiting for your input. - Review: apply or discard a thread's worktree from Git Actions. Fixes & Maintenance …View on GitHub ↗ -
Release notes excerpt
Features & Improvements - Claude Code overhaul: sessions run under Codexia's own session id and follow /clear, so the transcript, the live card and the session list stay in sync. - Claude Code: edit a past prompt to branch the conversation; the original session is kept intact and the prompt returns to the composer. - Claude Code: sub-agents nest under their Task call in the trace, with their prompt, steps and report. - Claude Code: a composer chip lists MCP servers and plugins that are not working, with the reason and a link to Connectors. …View on GitHub ↗ -
Release notes excerpt
Features & Improvements - Bots: a redesigned creation and settings flow; a new role opens in its chat instead of the settings dialog; a signed-in account is used by default, with sign-in offered in the chat; bots run without a project. - Pair a phone or remote browser with the server; the web package ships as static musl builds and is now the primary runtime. - ACP agent dropdown with restored presets and an animated Keke icon; the ACP composer matches the Codex and Claude composers. …View on GitHub ↗