Every plan, key and model it can drive
Catalase does not sell you model access. It connects to what you already have: a coding plan you subscribe to, an API key you hold, or a model on your own disk. This page lists all three routes and explains exactly how each one authenticates.
Coding plans, driven through the vendor's own CLI
These are flat-rate subscriptions, not metered API access. For each one you install the vendor's own command-line tool and sign in to it yourself; Catalase then starts that tool as a subprocess and streams the response into the editor.
No credential passes through Catalase, and none is stored by it. The CLI keeps its own session in its own credential store, exactly as it would if you ran it directly in a terminal.
Product names and marks belong to their respective owners, downloaded from each vendor's own brand resources. Their presence indicates compatibility only — not endorsement, affiliation or partnership.
| Plan | How Catalase reaches it | You install |
|---|---|---|
| Claude Pro / Max | Vendor CLI, your own login | npm install -g @anthropic-ai/claude-code |
| GitHub Copilot | Vendor CLI, your own login | npm install -g @github/copilot |
| ChatGPT Plus / Pro | Vendor CLI, your own login | npm install -g @openai/codex |
| Google AI Pro / Ultra | Vendor CLI, your own login | npm install -g @google/gemini-cli |
| Alibaba Coding Plan | Vendor CLI, your own login | npm install -g @qwen-code/qwen-code |
| Amazon Q Developer | Vendor CLI, your own login | AWS installer |
| opencode | Vendor CLI, your own login | npm install -g opencode-ai |
| Z.ai GLM Coding Plan | API key against the coding-plan host | Nothing — paste the key in Settings |
Two further CLI providers, Antigravity and Kilo Code, are wired in the same way.
Why a flat-rate plan shows $0.00. Every provider above is already paid for by a monthly fee, so the cost meter prices a turn at $0.00 (subscription). That is not a claim the usage is free — it is a statement that Catalase has no basis to meter it, because billing happens entirely on the vendor's side. Your plan's own usage limits still apply, and Catalase cannot see or raise them.
Check your own plan's terms. Catalase does not reimplement, proxy or impersonate any vendor's sign-in — it launches a program you installed, under a session you created. Whether a particular vendor permits driving their plan from a third-party tool is between you and that vendor, and some terms restrict automated or non-human access. If that matters for your situation, read your plan's terms, or use an API key or a local model instead.
Why Cursor and Windsurf are not on the list
Their subscriptions are tied to their own editors and there is no command-line tool a third-party application can drive. Listing them would promise access that does not work. This is a deliberate omission, not an oversight.
API keys, billed per token
If you hold a normal metered API key, Catalase uses it directly. Test verifies a key against the real API before you rely on it, so a bad key fails in Settings rather than halfway through a task.
| Provider | Notes |
|---|---|
| Anthropic | Metered per token |
| OpenAI | Metered per token |
| Metered per token | |
| OpenRouter | One key, many models behind it |
| Z.ai / Zhipu | Coding-plan keys are served from a different host than pay-as-you-go keys. Catalase defaults to the coding-plan host; set ZHIPU_API_BASE if you hold a metered key instead |
Keys are encrypted at rest with your operating system's own keychain, through Electron's safeStorage, and are never written to a config file. They are masked on screen. A key only ever goes to the provider it belongs to.
Local models, free
Catalase runs models on your own machine through Ollama, which is free and open source (MIT) for Windows and Linux. A turn on a local model costs nothing and needs no network connection.
- The Marketplace ranks models against the RAM and VRAM actually free on your machine, so you do not install one that will not run.
- The default is
qwen2.5-coder:1.5b— about 1 GB, and it runs on a laptop with no GPU. - Inline completion has no cloud network client at all — structurally, not as a setting. Ghost-text suggestions are always local.
- Privacy mode forces every request to a local model, and is enforced on the server side rather than only in the UI.
Full detail on hardware fit, quantization and expected latency is on the local models page.
What Catalase never does
- It does not sell model access or resell anyone's capacity.
- It does not see, store or transmit your subscription credentials.
- It does not reimplement or impersonate a vendor's login flow. That approach was considered and rejected.
- It does not send telemetry. The error log stays on your machine.
- It does not require a Catalase account, because there isn't one.