Cherry Studio

Comparing models before choosing a provider

Add two or three providers in the model settings, sync their model lists and run Model Check on each. Then open one conversation, select several models and send the same prompts, including a real document or code sample from your work. Compare answer quality, latency and the token counts the app shows, and read the provider price list alongside them, because Cherry Studio itself does not bill model usage. Keep the winning setup as an assistant preset so colleagues can repeat the test.

Building a private document knowledge base

Create a knowledge base, add a folder or a set of files and wait until each source reports Ready. Start with BM25 only, run the Recall Test with questions whose answers you know, and add an embedding and rerank model if phrasing mismatches cause misses. Remember that an embedding model hosted in the cloud receives your chunks, so pick a local model through Ollama or LM Studio when the material must stay on the machine. Re-index after changing chunking rules, as described in the knowledge base troubleshooting guide.

Delegating file work to a local agent

Create an agent with a dedicated working directory and the Ask Before Acting permission mode, then give it a bounded task such as turning a spreadsheet and notes into a slide deck. Watch the status panel for sub-agents and background commands, review generated files in the file pane and use file comments to request corrections on specific lines or slides. Only move to Auto-accept Edits once the agent behaves predictably on that folder, and keep the folder under version control or backup.

Routing coding CLIs through one model configuration

Open Code Mate, choose a working directory and a detected terminal, and pick a CLI such as Claude Code, Codex or OpenCode. Either select a compatible provider or use the unified gateway so the CLI can use any model configured in Cherry Studio. Verify with a read-only command first and confirm that the account or model shown by the CLI matches what you intended, following the Code Mate guide.

Exposing models to local scripts

Start the API Gateway, copy its URL and bearer key and point a local script or tool at the OpenAI- or Anthropic-compatible endpoint. Test a model list request and one short completion before wiring it into anything larger. Keep the key out of repositories and screenshots, regenerate it if it leaks, and leave the gateway bound to the local machine unless a firewall and access control are in place.

Running scheduled reports into a team channel

Connect a Slack or Telegram channel under Settings, Channels and send a test message. Create a scheduled task that runs a verified agent with a fixed workspace, run it manually once and read the log, then enable the schedule. Set the channel entry point to inherit agent settings or a stricter mode, and allow only the chat IDs that should reach the agent. A message that was not delivered does not prove the task did not run, so check the run log after failures.