Because GLM-5.3-Flash — the model that ran as Ox Alpha during its stealth preview — is served through OpenRouter’s OpenAI-compatible API, any coding tool that lets you set a custom base URL and model id can use it: Cursor, Cline, Roo Code, Continue, Aider, OpenCode, Zed and most others. No dedicated plugin is needed. This guide gives the three values every tool asks for, the generic setup, and the settings that actually matter for agentic coding. If your tool is still configured with stealth/ox-alpha, it is now broken — that ID was retired on August 26, 2026 with no redirect.
The three values you need
| Provider / API type | OpenAI-compatible (or “OpenRouter” if the tool lists it) |
| Base URL | https://openrouter.ai/api/v1 |
| API key | Your OpenRouter key (sk-or-…) — how to get one free |
| Model id | z-ai/glm-5.3-flash (formerly stealth/ox-alpha, retired) |
| Context window (if the tool asks) | 1048576 |
| Max output (if the tool asks) | 131072 |
Generic setup (works for most tools)
- Open the tool’s model or provider settings.
- Choose OpenRouter if it is listed; otherwise choose OpenAI-compatible / custom and set the base URL above.
- Paste your OpenRouter key.
- Enter
z-ai/glm-5.3-flashas the model id (type it — it will not always appear in a dropdown). - If there are fields for context length and max tokens, use 1048576 and 131072.
- Send a small test prompt. A reply means it works; a 401 means the key is wrong; a 404 on the model means the id is no longer listed.
Tool-specific notes
- Cursor: add the model id under the models list and override the OpenAI base URL with the OpenRouter one; Cursor routes through its own servers for some features, so check which modes honour a custom endpoint.
- Cline / Roo Code: both have a native OpenRouter provider — pick it, paste the key and type the model id.
- Continue: add an entry to the config with provider
openrouter(oropenaiwithapiBaseset), modelz-ai/glm-5.3-flash. - Aider: set
OPENROUTER_API_KEYand run with--model openrouter/z-ai/glm-5.3-flash. - OpenCode: OpenRouter is a built-in provider; select the model by id. (The listing also appears via OpenCode’s own Zen catalogue.)
- Any other tool: if it can talk to an OpenAI-style endpoint, the generic setup above applies.
Menus and option names change between releases; if a step does not match your version, the three values in the table are what matter.
Settings that matter for agentic coding
- Reasoning is always on. Expect a pause before the first token. If a tool has a “reasoning effort” option,
lowspeeds up small edits; leavemaxfor debugging and refactors. - Tool calling is supported (
tools,tool_choice), so agent modes that edit files and run commands work as designed. - Use the context. Tools that let you attach whole folders can put far more in a single request than usual — see the context-window guide.
- Watch for 429s. Agent loops make many calls; no rate limits are published, and heavy use is the most likely way to be throttled.
Migrating a tool from Ox Alpha
If you configured a coding tool during the free preview, two things need attention. Change the model ID to z-ai/glm-5.3-flash — there is no redirect, so the old one returns a 404 on the model rather than silently rerouting. Then make sure the OpenRouter account behind your key has credit, because agent loops that cost nothing during the preview now bill per token. Everything else in your configuration stays as it is.
Things worth knowing
- Your code leaves your machine. It now goes to an identified provider — Z.ai — under standard commercial terms rather than the Stealth Program’s data-for-access arrangement. Apply the same diligence you would to any third-party API before pointing it at proprietary code.
- Agent loops have a bill attached now. It is a cheap model, but a coding agent left running against a large repository makes a lot of calls. Watch the first day’s spend before trusting a loop unattended.
- Caching rewards stable context. Cached input costs one fifth of normal input, so tools that keep a consistent repository prefix across turns get materially cheaper. See pricing.
- You can self-host. The weights are MIT-licensed on Hugging Face, which is a genuine option if your code cannot leave your infrastructure.
- Benchmarks: Z.ai’s own figures put it at 63.4 on DeepSWE v1.1 and 84.3 on Terminal Bench 2.1 — vendor claims, not independently reproduced. See benchmarks.
Related: API and migration reference · GLM-5.3-Flash specs · Pricing · Try it in the browser first.
Sources and verification
Last substantively reviewed: August 27, 2026. Reviewed by: OxAlpha.chat Editorial Team.
How we verified this page
Tool menus change frequently, so the stable values on this page are checked against each tool’s official provider documentation. A tool is described as compatible only through the route documented by its maintainer; this page does not claim that every proprietary feature uses a custom model.
Primary sources
- Z.ai tool support: GLM Coding Plan tool and endpoint guide — documents supported coding tools and compatible endpoints.
- Cursor documentation: Cursor API-key settings — explains the limits of external model keys in Cursor.
- Cline documentation: Cline OpenRouter provider setup — documents Cline’s native OpenRouter configuration.
- Continue documentation: Continue OpenRouter provider configuration — documents the Continue model entry and provider fields.
- Aider documentation: Aider with OpenRouter — documents the environment variable and model-name format used by Aider.
- API baseline: OpenRouter API quickstart — documents the shared endpoint and authentication used by compatible tools.
Found a factual error or a changed price? Send us the page URL and supporting source; corrections follow our editorial policy.