Cherry Studio: Free Open-Source AI Desktop Client (2026)

Verified against the official repository and documentation on 23 August 2026. Current stable release: v2.0.8.

Cherry Studio is one of the most-starred AI desktop clients on GitHub and one of the least-covered in English. It is not a model and it is not an inference engine — it is the client: a single native window that talks to every LLM API you already pay for, plus whatever you run locally, with a shared history, a document knowledge base, MCP tool calling and an agent runtime bolted on.

This guide covers what it actually is, whether it is genuinely free, how to install and configure it with real values, which features are real versus marketing, how it stacks up against LM Studio and Open WebUI, and the limitations worth knowing before you commit.

What is Cherry Studio?

Cherry Studio is a cross-platform desktop client for large language models, built in TypeScript on Electron and published by the CherryHQ project. You bring your own API keys — or point it at a local model server — and it gives you one interface for all of them instead of four browser tabs and three desktop apps.

The repository describes itself as an "AI productivity studio with smart chat, autonomous agents, and 300+ assistants," which is a fair summary of where the project landed after its 2.0 rewrite in August 2026. It started in May 2024 as a straightforward multi-provider chat client and has grown into something closer to a local AI workbench.

Cherry Studio at a glance

Two things in that table matter more than the rest. First, Cherry Studio does not run models itself — there is no bundled llama.cpp engine. Local inference means pointing it at a server you already run. Second, the license is AGPL-3.0, which is more restrictive than the BSD- and MIT-style licenses used by several competitors, and that has real consequences if you plan to embed it in a product.

Who is Cherry Studio actually for?

It fits a specific shape of user well and is overkill or wrong for others.

• Good fit: you already pay for API access to several providers. If you have OpenAI, Anthropic, Google and DeepSeek keys, Cherry Studio replaces four subscriptions' worth of chat UIs with one window and per-token billing you control.
• Good fit: you want local and cloud in the same app. Ollama running privately for sensitive notes, a frontier cloud model for hard problems, switchable mid-conversation.
• Good fit: you want RAG over your own documents without standing up infrastructure. The knowledge base is built in and needs no server.
• Poor fit: you only use one provider. If ChatGPT is your entire AI stack, the native ChatGPT app is simpler and you lose nothing.
• Poor fit: you need multi-user access. Cherry Studio is single-user desktop software. Teams that need shared accounts, permissions and a browser URL want Open WebUI instead.
• Poor fit: you want the simplest possible path to running a local model. You would still need Ollama or LM Studio underneath. Start there.

Is Cherry Studio free, and is it open source?

Yes to both, with one nuance worth understanding.

The Community Edition — the app almost everyone means by "Cherry Studio" — is licensed AGPL-3.0. The LICENSE file is the unmodified AGPL v3 text with no extra clauses appended. There is no free tier, paid tier, seat count or usage cap in the desktop app. Your only cost is whatever the providers you connect charge you, and connecting a local Ollama model costs nothing at all.

The nuance: AGPL-3.0 permits commercial use but carries the network-copyleft obligation — if you modify Cherry Studio and make it available over a network, you must publish your source. The project acknowledges this and sells an exemption. Its README states that if you "require a commercial license that provides an exemption from the AGPL-3.0 requirements," you contact bd@cherry-ai.com. There is also a separately sold, privately deployable Enterprise Edition with centralized LLM management, shared enterprise knowledge bases, role-based access control and an on-premises backend. Pricing is not published — it goes through the same sales contact.

For an individual or a team using the app as shipped, none of that applies. You download it and use it.

How do you install Cherry Studio?

Every release publishes signed installers for all three operating systems on the GitHub Releases page, and the official download page is cherryai.com.cn/download/v2. Pick the artifact matching your CPU architecture — arm64 builds now exist for all three platforms, which is a change from earlier versions.

Windows

1. Download Cherry-Studio-2.0.8-x64-setup.exe (or -arm64-setup.exe on a Snapdragon machine).
2. Run the installer. A per-user install avoids the admin prompt and is fine for normal use.
3. If you cannot install software or work from a locked-down machine, use Cherry-Studio-2.0.8-x64-portable.exe instead — it runs without installation.

A longer screenshot walkthrough lives in our Cherry Studio on Windows guide.

macOS

Homebrew is the cleanest path and handles the Gatekeeper quarantine for you:

brew install --cask cherry-studio

The cask currently tracks v2.0.8 and requires macOS 12 or later. If you prefer a direct download, take Cherry-Studio-2.0.8-arm64.dmg on Apple Silicon or the -x64.dmg on Intel and drag the app to /Applications. On first launch from a direct download, right-click the app and choose Open to clear the unverified-developer warning. See our Cherry Studio on Mac walkthrough for the full sequence.

Linux

• Debian / Ubuntu: sudo apt install ./Cherry-Studio-2.0.8-amd64.deb
• Fedora / RHEL: sudo dnf install ./Cherry-Studio-2.0.8-x86_64.rpm
• Anything else: chmod +x Cherry-Studio-2.0.8-x86_64.AppImage && ./Cherry-Studio-2.0.8-x86_64.AppImage

ARM builds (arm64.deb, aarch64.rpm, arm64.AppImage) are published in the same release, which makes Raspberry Pi-class and ARM server desktops viable. Our Linux and Ubuntu guide goes further into desktop integration.

How do you connect a model provider and send your first message?

This is the ten-minute path from a blank install to a working chat. The menu names below are the current 2.x labels.

1. Open Cherry Studio and click the gear icon in the left navigation bar to reach Settings.
2. Select the Model Services tab. You will see a long list of providers down the left side — OpenAI, Anthropic, Gemini, Vertex AI, DeepSeek, Mistral, Perplexity, GitHub Copilot, MiniMax, OpenRouter, SiliconFlow, Alibaba Cloud Bailian, Volcano Engine and dozens more.
3. Click the provider you have a key for and paste the key into the API Key field.
4. Toggle the switch on the far right of the provider row to enable it. A provider with a valid key but the toggle off will not appear in the model picker — this is the single most common setup mistake.
5. Enable only the specific models you intend to use. Providers expose dozens; leaving them all on makes the model picker unusable.
6. Close Settings, start a new chat, and choose your model from the picker at the top of the conversation.

If your provider is not in the list, use a custom provider. Cherry Studio groups everything by protocol: OpenAI-compatible (/v1/chat/completions), Anthropic-compatible (/v1/messages), Gemini, Bedrock and Azure OpenAI. Any endpoint speaking one of those shapes works — including proxies and gateways like OpenRouter, Helicone or a self-hosted LiteLLM. The Anthropic-compatible shape is the one Cherry's agent features are built around, so it matters if you plan to use agents.

How do you run local models in Cherry Studio?

Cherry Studio has no inference engine of its own. Local models mean a server running on your machine that Cherry connects to over HTTP. Three backends cover almost everyone. If you are choosing between them, our comparison of Ollama, LM Studio, vLLM, llama.cpp and MLX breaks down the trade-offs.

Ollama

1. Install Ollama and pull a model: ollama run llama3.2.
2. In Cherry Studio: Settings → Model Services → Ollama.
3. Set API Address to http://localhost:11434/. Leave API Key blank — Ollama does not use one.
4. Click + Add and type the model name exactly as ollama list reports it (for example llama3.2). Cherry does not always auto-discover; adding by name is the reliable path.
5. Toggle the provider on, then pick the model in a new chat.

There is also a keep-alive setting measured in minutes. If no new message arrives inside that window, Cherry disconnects from Ollama and lets it release the model from memory — useful on a laptop where a 7B model sitting resident costs you battery and RAM.

LM Studio

Start LM Studio's local server, load a model, and add it in Cherry Studio the same way — the endpoint is OpenAI-compatible, so the API address is LM Studio's server URL and the key can be left blank. This is a common pairing: LM Studio as the engine and model manager, Cherry Studio as the interface. Our LM Studio guide covers the server side.

vLLM

For a real GPU box, vLLM is the throughput option. Start it with its OpenAI-compatible server, which listens on port 8000 by default, then in Cherry Studio add a custom provider with API address http://localhost:8000/, a blank or arbitrary API key, and the model names you loaded. Cherry also documents GPUStack and OpenVINO Model Server for the same role. For the wider self-hosting picture, see our self-hosting LLMs guide.

Which Cherry Studio features actually differentiate it?

Plenty of clients offer multi-provider chat. These are the parts that are genuinely harder to get elsewhere.

Knowledge base (RAG over your own documents)

Open Knowledge Base in the left sidebar, click the add button, name it, and choose how retrieval should work. Cherry Studio 2.x offers three modes, and the first one is a meaningful change from older versions:

The "None" option matters: you can build a working knowledge base with zero embedding setup and no embedding costs. Cloud embedding buys better recall on loosely-worded questions at the price of sending your documents to that provider during indexing. Local embedding keeps everything offline. Cherry ships a retrieval test screen so you can check what a query actually pulls back before you trust it in a chat.

MCP tool calling

Model Context Protocol support lives under Settings → MCP, with a built-in MCP section offering common capabilities you can install or enable directly, plus custom servers. Both transports are supported: stdio for local command-based servers and SSE / streamable HTTP for remote ones. In v2.0.8 the composer surfaces MCP prompts and resources, not just tools.

The gotcha: a server has to be started in Settings before it can be bound to an agent, and newly-added tools only load on the next message you send, not the current one. If you are picking servers to wire up, our roundup of MCP servers worth installing is a reasonable starting list.

Agents and Code Mate

The 2.0 release moved the agent runtime inside the app, eliminating the separate dependency installation and environment setup earlier versions needed. Agents run through a local API gateway that does not need to be exposed to the network. Each agent has a working directory that scopes which files it can touch, persistent task records that survive across sessions, sub-task decomposition, and permission modes including step-by-step confirmation before actions execute.

The release notes state agents support "both OpenAI and Anthropic endpoints" and reuse "Claude and Codex subscriptions" — meaning you can drive agents from a plan you already pay for rather than opening separate API billing. Agents can also read and write knowledge bases, so a project's accumulated context compounds instead of resetting each session.

Code Mate is the coding half: one-click model-service configuration for Claude Code, Codex and others, with DeepSeek Harness integration added in v2.0.7. It is a configuration layer over existing coding agents, not a new agent of its own. For the wider landscape, see our AI coding agents guide.

Translation

A dedicated split-pane translation workspace — input left, output right, language and model controls on top. It handles pasted text, uploaded files, and images via OCR, and extracts text automatically from PDF, Word, PowerPoint and Excel. Limits are 5 MB for images and text files and 20 MB for documents. v2.0.8 added layout-preserving PDF translation through BabelDOC, which keeps the original page structure rather than dumping reflowed text. Options include markdown preview, synchronised scrolling, auto-copy on completion, a custom system prompt, and custom languages.

Backup and note-app integration

Data settings support WebDAV and S3-compatible backup targets, integration with Notion, Obsidian and Siyuan, and relocating the storage directory. There is no first-party cloud sync — deliberately, since the project's stance is that your data stays on your device — so backup is something you configure rather than something you get.

Cherry Studio vs LM Studio vs Open WebUI: how do they compare?

All the star counts and licenses below were checked on 23 August 2026.

Read the first column first. It splits the field cleanly: LM Studio and Jan ship their own inference engine, so they are a complete local-AI solution in one download. Cherry Studio, Open WebUI and Chatbox are clients — they need something else doing the inference, whether that is a cloud API or a local server. That single distinction resolves most "which should I use" questions.

Cherry Studio vs Open WebUI — which one should you pick?

These get compared constantly and they solve genuinely different problems.

Open WebUI is a self-hosted web application. You run it in Docker, it serves a browser UI, and it is built around multiple users: accounts, roles, permissions, shared model access, shared document collections, an admin panel. That is the whole point. It is also enormously more popular — 149,603 stars against Cherry Studio's 50,926 — largely because it became the default face for self-hosted Ollama deployments.

Cherry Studio is a native desktop app for one person. No Docker, no reverse proxy, no TLS certificate, no user administration. You install it and it works. It is faster to reach a working setup and better at the things a desktop app can do that a browser tab cannot: OS-level screen capture with annotation and OCR, direct filesystem access for agent working directories, native multi-window and split-view layouts, local file drag-and-drop into a knowledge base.

The decision rule:

• One person, one machine, want it working in ten minutes → Cherry Studio.
• A team, a shared server, accounts and permissions required → Open WebUI.
• You want to reach your AI from a phone or another machine → Open WebUI. Cherry Studio has no web surface.
• Licensing is a constraint → Open WebUI's BSD-3-Clause base is more permissive than AGPL-3.0, but note it added a branding clause in v0.6.6 requiring deployments above 50 users in a rolling 30-day window to keep Open WebUI branding unless they buy an enterprise license.

They are not mutually exclusive. Open WebUI on a home server for shared and mobile access, Cherry Studio on your laptop for desktop-native work, both pointed at the same Ollama instance, is a coherent setup. Our Open WebUI + Ollama setup guide covers the server half.

Cherry Studio vs LM Studio — what's the actual difference?

LM Studio is a model runner. It bundles its own inference engine, includes a model browser wired into Hugging Face, handles quantisation choices and GPU offload settings, and exposes an OpenAI-compatible server other tools can call. Its cloud-provider support is minimal because that is not what it is for. It is closed-source proprietary software, though it has been free for both personal and workplace use since July 2025, with no form to fill in.

Cherry Studio is a client. It has no engine, no model browser, no quantisation settings, no GPU configuration. What it has instead is 60+ cloud providers, a knowledge base, MCP, agents and translation.

Practical guidance:

• Only running local models, want one download → LM Studio, and skip Cherry Studio entirely.
• Only using cloud APIs → Cherry Studio; LM Studio would not help.
• Both → run them together. LM Studio (or Ollama) serves the models; Cherry Studio is the window you actually work in. This is the most common real-world configuration and the two tools do not conflict.
• Open source is a hard requirement → Cherry Studio is AGPL-3.0 and auditable. LM Studio's desktop app is closed source, though its MLX engine and lms CLI are open.

What are Cherry Studio's limitations?

Four worth weighing before you commit.

1. There was a real telemetry regression in 2026. A community audit found that disabling "send anonymous error reports and usage statistics" did not stop all outbound analytics. Two events — trackAppLaunch and trackAppUpdate — bypassed the enableDataCollection setting after a 22 March 2026 commit removed their opt-out guards, while only token-usage tracking still honoured it. The requests carried a device ID, OS version, CPU architecture and app version. The fix, PR #14390, merged on 20 April 2026 and restored the check on all three events. Chat content, files and API keys were never in that channel. The honest read: the outcome was fine, the process was not, and if telemetry matters to you, verify the toggle at Settings → Data Settings → Privacy Settings and confirm with your own network monitoring rather than trusting the switch.

2. AGPL-3.0 is the most restrictive license in its category. Fine for using the app. A problem if you want to fork it into a product — compare Jan (Apache-2.0), AnythingLLM (MIT) and Open WebUI (BSD-3 with a branding clause).

3. Documentation is Chinese-first. English pages exist and are good, but they lag and some sections are only complete in Chinese. The project is led from China and the primary community is Chinese-speaking. This is the main reason a 50k-star project stays under-covered in English.

4. It is single-user desktop software. No web access, no mobile app, no accounts, no sharing. If you want your AI setup reachable from a phone, this is the wrong tool.

One smaller note: at 1,366 open issues the project moves fast and does not close everything. Release cadence is genuinely rapid — v2.0.0, v2.0.5, v2.0.7 and v2.0.8 all shipped within two weeks of each other in August 2026 — which is good for features and occasionally rough for stability. Staying one point release behind is a defensible strategy.

What's new in Cherry Studio 2.0?

Version 2.0 shipped on 5 August 2026 and is the largest release in the project's history — it is what turned a chat client into a workbench. The current stable is v2.0.8 (19 August 2026).

• Built-in agent runtime. Agents run inside the app with no separate dependency installation or environment setup, support OpenAI and Anthropic endpoints, and reuse Claude and Codex subscriptions.
• Agents that maintain knowledge bases. Agents read and write knowledge bases so project context accumulates like a wiki.
• Code Mate. Code tools consolidated with one-click model-service configuration for Claude Code, Codex and more; DeepSeek Harness added in v2.0.7.
• Rebuilt interface. Chats, roles, projects and tasks reorganised, with multi-window and split-view layouts.
• Performance rework. A rebuilt data layer that stays responsive with long histories, smoother scrolling in very long conversations, and large-folder imports that no longer block the UI.
• Interactive screen capture with annotation and OCR (v2.0.8).
• Layout-preserving PDF translation via BabelDOC, with history (v2.0.8).
• MCP prompts and resources surfaced directly in the composer, plus global enablement controls for skills (v2.0.8).
• Durable agent messaging that survives across sessions, and a floating capsule showing task progress (v2.0.8).
• Artifact spreadsheet preview and configurable send / line-break / steer shortcuts (v2.0.7).

To update: brew upgrade --cask cherry-studio on macOS, or the in-app updater on Windows and Linux. Migration paths between v1 and v2 are documented if you are coming from an older install.

FAQ

Is Cherry Studio free?

Yes. The Community Edition is free with no tiers, seat limits or usage caps, licensed AGPL-3.0. Your only costs are the API providers you connect — and if you use a local model through Ollama, LM Studio or vLLM, there is no cost at all. A separately sold Enterprise Edition exists for organisations needing centralized management, SSO-style access control and on-premises deployment; its pricing is not published.

Is Cherry Studio open source?

Yes, under GNU AGPL-3.0. The full source is at github.com/CherryHQ/cherry-studio, where the repository had 50,926 stars and 4,847 forks as of 23 August 2026. The LICENSE file is unmodified AGPL v3 with no extra clauses. AGPL permits commercial use if you comply with it; the project sells an exemption license for those who cannot.

Cherry Studio vs LM Studio — what's the difference?

LM Studio ships its own inference engine, so it runs models directly and is the better single-download choice for pure local AI. Cherry Studio has no engine — it is a client for 60+ cloud providers plus whatever local server you point it at, with a knowledge base, MCP and agents on top. Many people run both: LM Studio as the engine, Cherry Studio as the interface.

Does Cherry Studio run models locally?

Not by itself. It has no bundled inference engine. It connects to a local model server you run separately — Ollama on http://localhost:11434/, LM Studio's server, vLLM on port 8000, GPUStack or OpenVINO Model Server. Once connected, you can work fully offline, including RAG if you pick a local embedding model or the keyword-only retrieval mode.

Does Cherry Studio support MCP?

Yes. Configure servers under Settings → MCP, using either stdio for local command-based servers or SSE / streamable HTTP for remote ones. A built-in MCP section offers common capabilities you can install directly. As of v2.0.8, MCP prompts and resources appear in the composer alongside tools. Servers must be started before they can be bound to an agent.

Is Cherry Studio safe to use?

The architecture is sound: API keys, chat history and knowledge bases stay on your device, and requests go straight from the app to each provider with no Cherry-operated relay. One caveat: an April 2026 audit found the analytics opt-out toggle was being bypassed for app-launch and update events, sending a device ID and system info. It was fixed on 20 April 2026 in PR #14390. Chat content and API keys were not involved.

What platforms does Cherry Studio support?

Windows 10 or later (x64 and arm64, installer or portable), macOS 12 or later (Intel and Apple Silicon builds, or via brew install --cask cherry-studio), and Linux as .deb, .rpm and .AppImage for both x86_64 and arm64. There is no web version, no mobile app and no browser access.

How many models and providers does Cherry Studio support?

Its documentation lists 60+ model service providers, and the app ships 300+ pre-built assistants. Beyond the built-in list, any endpoint speaking an OpenAI-compatible, Anthropic-compatible, Gemini, Bedrock or Azure OpenAI protocol can be added as a custom provider, which effectively covers everything including gateways like OpenRouter and self-hosted proxies.

Can Cherry Studio do RAG over my own documents?

Yes. Create a knowledge base from the sidebar, then pick one of three retrieval modes: no embedding model at all (BM25 keyword search), a cloud embedding model (BM25 plus vector hybrid), or a local embedding model (BM25 plus local vector search, fully offline). Drag in documents, use the built-in retrieval test to check what a query returns, then attach the knowledge base to a chat or an agent.

What is the difference between Cherry Studio and Cherry AI?

Same project. Cherry Studio is the product name and CherryHQ is the GitHub organisation; the marketing site sits at cherryai.com.cn, with the older cherry-ai.com domain redirecting there. CherryAI also appears inside the app as one of the built-in model gateway options.

Should you install Cherry Studio?

A simple decision rule. If you use exactly one AI provider, keep using its native app — Cherry Studio adds complexity you will not be repaid for. If you need shared access for a team or want to reach your AI from a phone, run Open WebUI instead. If you want the shortest path to running a local model and nothing else, install LM Studio.

But if you juggle several providers, want cloud and local in the same window, and want document RAG and MCP tools without standing up infrastructure, Cherry Studio is the strongest free option available in 2026 — and the AGPL license, the auditable source and the local-only data model mean you are not betting on a vendor staying generous. Set aside an afternoon, connect two providers and one Ollama model, and you will know within a day whether it replaces your current setup.

If you are working through the wider question of which models to run and where, our open-source LLM landscape guide covers the model side of the same decision.