Rewrite work messages with local AI on Windows, macOS, and Linux

Soften the tone without losing the request

Across Windows, macOS, and Linux, ClipsX can dispatch saved text clips directly to a locally configured Ollama chat model via the Rewrite extension. The example below shows an unedited prompt and the actual response produced by gemma3:4b using the Business preset.

Configure a text generation model, not an embedding model

Rewriting text requires an instruction-tuned chat model rather than an embedding model (which only calculates vector coordinates for Meaning Search). You will need ClipsX, the first-party Rewrite extension, a running Ollama instance, and a pulled LLM. While downloading a model requires an initial internet connection, all subsequent inference runs completely offline on your local machine.

1. Install and start Ollama on your system (Windows, macOS, or Linux). Pull a chat model tailored to your hardware (for example, gemma3:4b or llama3.2).
2. In ClipsX, navigate to Intelligence → Models. Verify your Ollama endpoint, detect installed models, and configure a chat preset.
3. Assign that preset to "Extension text generation". Ensure this slot is populated even if Meaning Search or Recall is already running.
4. Under Extensions, install and enable the first-party Rewrite package, granting it the text generation capability. Always start with manual invocation.

Generate a suggestion and compare it against the original

With ClipsX active, copy a draft message to capture it into your history. Open the clip details, choose Transform → Rewrite, and pick a preset such as Business or Casual. Use the built-in Compare view to inspect exact differences side-by-side before committing to Copy or Paste.

Combine model output with quick human edits

Often, the most effective message is a hybrid. For example: “Could you please send the agenda by Thursday at 3:00 PM? I need it before I can finish the review.” This human revision retains the polite opening generated by the model while cutting the passive, overly bureaucratic ending. AI suggestions serve as a fast sounding board, not an absolute authority.

ClipsX preserves your original clipboard content alongside transformed versions. Generated rewrites remain attached as child variations of the root clip, while "Save as new clip" creates an independent entry. Automated transform rules can be added later, but starting with manual invocation gives you full visibility into model behavior.

Select the right boundary for text processing

A localhost endpoint like localhost:11434 does not inherently guarantee zero data egress if the upstream server proxies requests elsewhere. Always verify that your Ollama installation is running genuine, locally stored weights (such as gemma3:4b). When drafting internal memos, credentials, or sensitive customer communications, strict on-device execution guarantees privacy boundaries that cloud endpoints cannot match.

Troubleshooting: performance, latency, and quality issues

When testing a local AI pipeline for the first time, start with a simple plain-text snippet. Remember that local inference latency and memory footprints scale with model parameter count and context window length; on-disk file size does not equal active runtime VRAM/RAM consumption. Text rewriting does not require embedding models or vector index configuration.

• Rewrite menu item missing: Verify that the extension is installed and enabled, text generation permissions are granted, and the clip contains editable text.
• Generation fails or times out: Ensure the Ollama service is active, the targeted model is downloaded, and your chat preset is correctly assigned to "Extension text generation".
• Output hallucinated or truncated: Revert to your original clip, make manual edits, or experiment with a different temperature or model size. A 200 OK inference response does not guarantee semantic accuracy.