Native desktop control center for AI coding tools — provider switching, local relay, MCP, skills, sessions, usage, and sync. Built with GPUI + axum.
Overview
One native control center for your AI coding tools. Connect providers, switch models, share MCP servers and skills, and route every supported client through one local gateway. Download OcHub · Why OcHub · Install · Build from source · Sponsors User documentation: docs.ochub.org. Its source and GPUI WebAssembly renderer are maintained in ochub-docs. AI coding tools are powerful, but each one keeps its providers, MCP servers, skills, sessions, and usage data in a different place. Changing an API endpoint or sharing the same setup across several clients quickly turns into repeated config-file editing. OcHub brings that work into one fast, native desktop app. Use it as a simple connection switcher, or place its local gateway between your tools and upstream APIs for centralized routing, compatibility conversion, failover, and usage tracking. Choose a tool, add an API connection, test it, and switch.
README
English · 简体中文
One native control center for your AI coding tools.
Connect providers, switch models, share MCP servers and skills, and route every supported client through one local gateway.
Download OcHub · Why OcHub · Install · Build from source · Sponsors
User documentation: docs.ochub.org. Its source and GPUI WebAssembly renderer are maintained in ochub-docs.
Why OcHub
AI coding tools are powerful, but each one keeps its providers, MCP servers, skills, sessions, and usage data in a different place. Changing an API endpoint or sharing the same setup across several clients quickly turns into repeated config-file editing.
OcHub brings that work into one fast, native desktop app. Use it as a simple connection switcher, or place its local gateway between your tools and upstream APIs for centralized routing, compatibility conversion, failover, and usage tracking.
OcHub currently manages Claude Code, Claude Desktop, Codex, Cherry Studio, Grok Build, Kimi Code, OpenCode, OpenClaw, and Hermes.
Everything in one place
| Capability | What it gives you | |
|---|---|---|
| 🔌 | Connection management | Discover, import, edit, test, and switch direct API connections without hand-editing client configs |
| ↗️ | Model providers | Give multiple clients one local endpoint, then control upstreams and routing centrally |
| 🔀 | Smart routing | Map model names and reasoning levels, retry compatible interfaces, switch to backup providers, and monitor health |
| 🧩 | MCP & skills | Keep reusable MCP servers and skills together, then distribute them to the tools that need them |
| 📊 | Sessions & usage | Browse local CLI sessions and understand tokens, cache, latency, requests, and estimated cost |
| 🔄 | Sync & backup | Protect your setup with snapshots and keep OcHub data in sync through supported remote storage |
| 🖥️ | Remote nodes | Switch existing providers on WSL, development machines, and headless servers securely over SSH |
| ⚡ | Native Codex modes | On macOS, launch Codex from OcHub and use Fast or Ultra directly in its native model picker with compatible models and upstreams |
| 🌙 | Kimi Code profiles | Manage native providers, model aliases, limits, capabilities, headers, and credential mappings in ~/.kimi-code/config.toml |
| 🍒 | Cherry Studio import | Save reusable provider connections in OcHub, then open Cherry Studio’s public Deep Link and confirm the import in-app |
Two ways to connect
Direct connection
Choose a tool, add an API connection, test it, and switch. OcHub writes the client’s native configuration, so the tool keeps working normally after OcHub is closed.
Best for: a straightforward endpoint change, a tool’s official login, or a small setup with independent connections.
Model provider mode
Point supported clients at OcHub’s loopback gateway. The gateway can translate between supported API dialects and apply model aliases, reasoning mappings, interface retries, health checks, and failover before a request reaches its upstream.
Best for: sharing an upstream across tools, normalizing model names, converting API formats, tracking usage centrally, or building a resilient multi-provider setup.
Native Codex Fast and Ultra launcher
On macOS, the Codex app page includes Launch Codex. It starts a separate Codex desktop instance and unlocks Fast and Ultra in the app’s native model picker for compatible GPT-5.6 models. OcHub intercepts the renderer script in memory over a loopback-only debugging connection; it does not modify or re-sign the installed Codex application.
This makes the native controls selectable and preserves the selected Fast service tier in Codex requests. The configured model provider must still support the requested service tier and reasoning effort. Launch Codex through OcHub each time you need the unlock. See the advanced Codex guide for supported models, security boundaries, and troubleshooting.
Built for the desktop
- Native UI powered by GPUI — no browser shell or webview
- Local-first storage in
~/.ochub/ - Cross-platform releases for macOS, Windows, and Linux
- In-app updates for DMG, Windows installer, and AppImage installations
- Open source under GPL-3.0-or-later
[!IMPORTANT] OcHub is under active pre-release development. Back up important tool configuration before trying an early build. OcHub only changes the live configuration of applications you explicitly manage.
Install
Homebrew
Homebrew automatically selects the correct Apple Silicon or Intel build:
brew install --cask ochub-team/tap/ochub
OcHub is signed with an Apple Developer ID. Notarization is not enabled yet, so if macOS blocks the first launch, open System Settings → Privacy & Security and choose Open Anyway.
Direct download
Download the latest release for your platform from the GitHub Releases page.
| Platform | Available packages |
|---|---|
| macOS | Apple Silicon and Intel .dmg; headless CLI .tar.gz |
| Windows 10/11 x64 | NSIS installer, portable GUI .zip, and headless CLI .zip |
| Linux x64 | AppImage, Debian .deb, and headless CLI .tar.gz |
Releases include SHA256SUMS and a GitHub artifact attestation. Packaging,
signature verification, and release details live in the
release guide.
The headless archives contain a single self-managing ochcli executable. See the included README
or run ochcli --help to configure and operate OcHub without the desktop GUI.
To control WSL or a development machine from OcHub Desktop, follow the
Remote Nodes guide.
Build from source
OcHub is a Rust workspace built with GPUI and axum. The repository pins its Rust toolchain, so rustup will select the correct version automatically.
git clone https://github.com/OcHub-team/OcHub.git
cd OcHub
cargo run -p ochub-app
Platform requirements:
- macOS: Xcode or Xcode Command Line Tools
- Windows: Visual Studio 2022 Build Tools with the Windows SDK
- Debian/Ubuntu:
./scripts/ci/install-linux-deps.sh
Common development commands:
just check
just test
just ci
just qa-app # macOS: builds /tmp/OCHUB-QA.app for acceptance testing
The workspace is split into these main components:
| Crate | Responsibility |
|---|---|
ochub-core |
Domain model, SQLite storage, client config, sync, MCP, skills, sessions, usage, and auth |
ochub-convert |
Request and response conversion between supported API dialects |
ochub-app |
Native GPUI desktop application |
ochcli |
Headless command-line interface |
Sponsors
OcHub’s development is supported by the API relay providers below. They are independent third-party services — you register and manage your own API key with each of them, and your requests go straight from your machine to theirs.
Learn more about our sponsors →
License
OcHub is licensed under the GNU General Public License v3.0 or later.
OcHub draws inspiration from
cc-switch and
new-api.
Special thanks to LINUX DO for providing a friendly, professional, and open community for AI developers.
Install
This server does not publish a one-line install command.
Open the repository installation guide New to MCP? Start with the mcpskill.org guide →