Cherry Studio is a desktop AI workspace for Windows, macOS and Linux. It brings cloud services such as OpenAI, Anthropic, Gemini and DeepSeek together with local endpoints such as Ollama and LM Studio, then adds multi-model conversations, preconfigured and custom assistants, document processing, knowledge bases, translation, drawing, WebDAV backup, global search, Mermaid rendering and MCP tool connections.
It is a client and orchestration surface, not a model. Answer quality, privacy, cost, retention and regional processing still depend on the exact model provider, endpoint, local runtime, embedding service, reranker, backup destination and MCP server selected for each workflow.
What Cherry Studio replaces—and what it does not
| Job | Cherry Studio contribution | Still owned elsewhere |
|---|---|---|
| Multi-model chat | One UI, assistant presets and simultaneous model conversations | Provider accuracy, safety, availability, rate limits and billing |
| Local models | Connects Ollama, LM Studio and compatible endpoints | Model download, hardware, serving security and license |
| Document Q&A | Imports common text, image, Office and PDF formats into knowledge workflows | OCR/extraction fidelity, embedding quality and source authorization |
| Tools | Runs configured MCP servers and mini-program integrations | Tool provenance, permissions, sandboxing and action approval |
| Backup | Can connect WebDAV and S3-compatible storage | Cloud encryption, retention, access control and restore testing |
| Team governance | Community desktop app provides individual configuration | Central accounts, access control and shared resources belong to Enterprise or external controls |
Community and Enterprise are different products
The Community Edition is the open-source desktop client. The repository describes an Enterprise Edition with a privately deployable backend, centralized provider/model management, employee accounts, shared knowledge bases, role-based access, backup and recovery. The published comparison says the Enterprise Edition is only partially released to customers and is sold by subscription or buyout.
Do not infer enterprise controls from the popularity of the desktop client. A business rollout of independently configured Community installations can create unmanaged keys, inconsistent providers, local knowledge copies and fragmented audit evidence. If central administration is required, evaluate the Enterprise contract or place Cherry Studio behind an approved gateway and device-management program.
License: AGPL-3.0 needs an actual decision
The current repository states that Cherry Studio Community Edition is licensed under AGPL-3.0 and that commercial use is permitted when its conditions are followed. A commercial license that exempts a customer from AGPL requirements is available from the vendor. This is materially different from describing the app only as “free and open source.”
| Use | Question | Evidence to keep |
|---|---|---|
| Unmodified internal desktop use | Does ordinary use trigger any distribution or network-source obligation in your setup? | Legal interpretation, exact binary/source revision and license notices |
| Modified desktop distribution | Are corresponding source and notices offered correctly? | Build source, patches, installation information and recipient offer |
| Modified network service | Does remote interaction invoke AGPL network-source provisions? | Deployed source, offer mechanism and dependency license inventory |
| Bundled enterprise product | Would reciprocal terms conflict with your distribution model? | Commercial license or compliant product architecture |
| Models and extensions | What separate terms cover providers, local models, assistants, themes and MCP servers? | Per-component SBOM/license register and acceptable-use terms |
This is a practical issue for counsel and engineering, not a reason to avoid the project automatically. Decide before modifying, redistributing or exposing it as part of a network service.
Where data goes
Cherry Studio's privacy policy, effective May 31, 2026, says API keys, conversation history and preferences are stored locally and are not uploaded to Cherry Studio servers. Model requests connect directly from the device to the configured third-party provider. That means “local history” and “private inference” are separate claims: a conversation stored on the laptop can still be sent to OpenAI, Anthropic or another configured endpoint for generation.
The policy describes anonymized basic operational and product-improvement data, including version/aggregated feature usage and crash/error information. Both categories can be disabled under Data Settings → Privacy Settings. It also warns that the relevant toggles are reset to their default enabled state after the data-collection architecture update, so users who previously opted out should revisit settings after upgrading.
| Data | Primary location/path | Control to verify |
|---|---|---|
| API keys | Local application storage according to policy | OS account isolation, disk encryption, backups and secret rotation |
| Prompts and responses | Local history plus direct request to selected model provider | Provider retention/training settings and local deletion |
| Knowledge documents/vectors | Local workspace and any selected embedding/reranking endpoint | Confirm whether chunks leave the device during indexing/query |
| Telemetry/crash data | Cherry Studio operational collection, described as anonymized | Recheck both privacy toggles after every upgrade |
| WebDAV/S3 backup | User-configured third-party storage | Encryption, credentials, bucket ACL, retention and restore |
| MCP context/actions | Configured MCP process or remote service | Command, file, network and account scope per server |
API-key safety
- Create separate keys for Cherry Studio instead of reusing production service credentials.
- Set provider-side spend limits, rate limits and alerts; the client cannot enforce every provider's billing control.
- Prefer restricted project keys and approved gateway endpoints over personal high-privilege keys.
- Use full-disk encryption and a locked OS account because local storage is accessible to malware or another local user.
- Do not paste keys into chats, assistants, knowledge documents, screenshots or MCP configuration shared publicly.
- Rotate keys after device loss, suspicious extensions, support-log sharing or untrusted backup exposure.
Evaluating the knowledge base
A successful upload is not evidence of reliable RAG. Build a test set from the documents with answerable questions, unanswerable questions, tables, scanned pages, conflicting versions, exact numbers and access-sensitive passages. For each answer, inspect the retrieved chunks and citations before judging fluency.
| Layer | Test | Failure signal |
|---|---|---|
| Parsing/OCR | Compare extracted text, tables, reading order and page references | Missing columns, headers or scanned content |
| Chunking | Check whether definitions, qualifications and table rows stay together | Answer cites a fragment without its exception |
| Embeddings | Test language, synonyms, acronyms and domain vocabulary | Relevant passages rank below generic text |
| Retrieval | Measure recall@k on a labeled set | Required evidence never reaches the model |
| Generation | Require citation-linked answers and abstention | Confident synthesis without supporting chunk |
| Updates/deletion | Replace a source and remove one sensitive document | Old vectors or answers remain searchable |
MCP turns a chat client into a privileged application
An MCP server can expose files, databases, browsers, shell commands, issue trackers or messaging systems. Tool descriptions and returned content can be malicious or compromised. Install servers from reviewed sources, pin versions, inspect their commands and dependencies, and run them with the least OS/network/account privileges possible.
Keep read-only research servers separate from mutating servers. Require explicit confirmation for deletion, payment, publishing, external messages, code execution and account changes. A friendly assistant preset is not a security boundary; the operating system, credentials and server implementation determine real authority.
A staged rollout
- Define approved jobs. Start with low-risk writing, comparison and public-document research.
- Install a verified release. Download from the official site or GitHub Releases and check signatures/hashes where provided.
- Review privacy settings. Decide telemetry policy and repeat the check after upgrades.
- Connect one restricted provider. Test spend, retention, model availability and error handling.
- Benchmark multi-model comparison. Use identical prompts and record quality after correction, latency and cost.
- Add a non-sensitive knowledge base. Validate parsing, retrieval, citation and deletion.
- Add one read-only MCP server. Observe process, file and network access before granting writes.
- Test backup and restore. Encrypt remote storage and verify a clean-device recovery.
- Choose team architecture. Decide whether unmanaged clients are acceptable or central Enterprise/gateway controls are needed.
Alternatives
| Option | Best fit | Tradeoff versus Cherry Studio |
|---|---|---|
| Cherry Studio | Feature-rich cross-platform desktop hub with many providers and MCP | Large trust surface and AGPL/team-governance decisions |
| Open WebUI | Browser-based shared or local Ollama-first deployments | Server operation rather than native desktop focus |
| LibreChat | Self-hosted multi-user web chat and provider routing | More infrastructure, stronger shared deployment model |
| LM Studio | GUI-first downloading and running of local models | Less broad cloud productivity/knowledge/MCP surface |
| Chatbox | Simpler cross-platform multi-provider chat | Fewer integrated workspace features may reduce complexity |
| TypingMind | Polished hosted/browser multi-model client | Closed-source/commercial model and different local-data posture |
| Dify | Building and deploying governed AI workflows/apps | Heavier platform; not primarily a personal desktop client |
Frequently asked questions
Does Cherry Studio provide AI models?
It primarily connects models and services. Availability, quality and pricing come from the provider or local runtime you configure.
Is it completely offline?
Only with local models, embeddings, rerankers, tools and backup. Cloud provider requests and optional integrations send data to their respective services.
Where are chats and API keys stored?
The current privacy policy says they are stored locally and not uploaded to Cherry Studio servers. Protect the device and verify provider-side processing separately.
Can telemetry be disabled?
The policy says both basic operational and product-improvement collection can be disabled in Privacy Settings. It warns that toggles may reset to enabled after the named architecture update.
Is the Community Edition commercially usable?
The repository says yes under AGPL-3.0 compliance. Modified distribution or network use deserves legal review; a commercial exemption is offered by the vendor.
Is Community Edition enough for a company?
It can serve individual power users, but centralized accounts, shared knowledge, role permissions and private backend management are described as Enterprise capabilities.
Primary sources
- Official repository, feature matrix and edition comparison
- Official privacy policy effective May 31, 2026
- Official security and supported-version policy
- Community Edition AGPL-3.0 license
- Official MCP setup documentation
- Official knowledge-base documentation
- MCP security best practices
Last reviewed July 25, 2026. Features, policies and editions change quickly; check the exact installed release, provider terms and current license before rollout.
