What each route does with your code and prompts
Catalase can run a model on your machine, or send your prompt to a provider you choose. This page says what each route is documented to do and where independent verification is still pending.
Network traffic for each route has not been independently audited. This page states what Catalase documents about itself and identifies where independent verification is still pending.
Four ways to run it
Each route sends something different off your machine. The Evidence column says what has been confirmed on the installed build.
| Route | Where the model runs | What is sent off your machine | Evidence |
|---|---|---|---|
| Local model through Ollama | On your machine, through Ollama. | Local turns are not documented as sent to any cloud provider. Pulling a model downloads it to your machine. Network traffic for this route has not been independently audited. | Confirmed: a local-model turn ends at $0.0000 in the cost meter. Full network audit is pending. |
| API key | At the provider whose key you saved: Anthropic, OpenAI, Google, OpenRouter, Zhipu or Zhipu Coding Plan. | Your prompt and the context you attach go to that provider. Keys show masked on screen. Test Key contacts the provider. | Verified: saved keys show masked in the UI. Routed calls with paid keys have not been tested. |
| Subscription CLI | At the provider. Catalase drives the CLI you installed and signed in to: Claude Code, OpenAI Codex, GitHub Copilot, Gemini CLI, Qwen Code, Amazon Q, opencode or GLM. | Catalase documents that it never sees your credentials. Your prompt and context reach the provider through that CLI. What the CLI sends is governed by that tool and its provider. | Present in this release; network traffic has not been independently measured. |
| Local-only mode (Settings → Privacy) | On your machine. A cloud model you pick is replaced by a local one. | Sends nothing to cloud models. If you pick a cloud model, Catalase substitutes a local one and tells you in the chat. | Present in this release; not yet verified end to end. Whether it also stops background requests is unverified (see below). |
What stays in your project
Catalase keeps its project data in a hidden .catalase folder inside the project. The Explorer does not show it. The folder holds:
- Snapshots in
.catalase\context, one per day. Snapshots are experimental. They are taken when you close the app, every 30 minutes (skipped when nothing has changed), and when you use/remember. - Project memory, written by
/remember. - Custom commands in
.catalase/commands. - MCP server settings in
.catalase/mcp_servers.json.
When you resume a snapshot, it goes to the model in your first message. A vision-capable model receives the image, and other models receive the text summary. The API path has been verified. The vision path has not been tested. The model you resume with may be a cloud model.
Network activity we know about
These are the requests Catalase is documented to make. The list is not a complete traffic capture.
- Model catalogue. Catalase discovers GGUF models on Hugging Face alongside a curated list. It refreshes when you press Refresh catalog, and in the background every 24 hours.
- Ollama pulls. Installing a model downloads it from Ollama.
- Push and Create PR. Push sends the current branch to origin. Create PR opens a pull request on GitHub through the GitHub CLI (
gh). It works with GitHub only. - Test Key. Checks a saved key with its provider.
Not yet verified: whether Local-only mode also stops the catalogue refresh and other background requests.
Cost controls
- The cost meter in the chat panel shows this session's spend against its cap. The default cap is $5.00. You can set your own in Settings → Budget.
- Catalase warns you at 80% of the cap and alerts you at the cap. The cap behaviour has not been independently verified against live paid spend.
- A local-model turn shows $0.0000.
- The context meter shows how much of the model's context window the conversation uses. Start a new conversation with
/clearwhen it fills.
The cap behaviour has not been independently verified against live paid spend. Set a low cap and monitor it on first use.
Pending independent verification
How we plan to verify
This is a plan, not a result.
We will capture each route's network requests through an intercepting proxy on a clean Windows 11 install. We will publish the results as the table on this page, and update the Evidence column for each route as its traffic is measured. Local-only mode and the budget cap will be tested end to end on the same install.
This website: the download form
The paragraphs above are about the Catalase application. The website is a separate thing, and one part of it does collect personal data:
- What we collect. When you click a download button, we ask for your first name, last name and email address, and which platform you want. We also record the page you were on and how you arrived at the site.
- Why. To email you a personal download link, and to understand how much interest there is in each platform.
- Tracking. The email we send contains a tracking pixel, so we know when it is opened, and the download link records when it is clicked. That is the full extent of it: open, click, download.
- What we don't do. No advertising cookies, no third-party analytics on this site, no selling or sharing of your details. The free tools on this site run entirely in your browser and send nothing anywhere.
- The application is different. Nothing in this paragraph is about the Catalase editor. The app has no account, no login and no telemetry — the rest of this page is about the app.
How long we keep download requests, and which email provider delivers them, will be published before the download goes live.