If you use more than one coding agent, you already know the tax: a separate model account for each, a separate card on file, a separate place where the balance runs out. Explyt 5.19 removes that for a specific set of tools. The subscription that used to power only the Explyt plugin in JetBrains IDEs now gives you an OpenAI-compatible URL and an API key, and the key works in Claude Code, Claude Desktop, Cursor, Codex, Cline, OpenCode and Kilo Code.
This post is the practical version: where the values are, three configs you can paste, which models answer, and the boundary that the release notes put in bold.
Step 0: get the two values
Open your Explyt account. With an active subscription or a positive Flex minutes balance you will see a block called Connect to Explyt.
It holds two things:
- The URL
https://api.explyt.ai/openai/api. Same for everyone. - Your API key. Copy key copies it; the eye icon reveals it.
If the block is not there, the account has neither a subscription nor Flex minutes, and there is nothing to connect yet.
Keep the key like a password. It is one key for every tool; Reissue in the account kills the old one instantly, and you then update every tool you configured.
Step 1: Claude Code
From the Claude Code page. Open or create ~/.claude/settings.json:
{
"env": {
"ANTHROPIC_BASE_URL": "https://api.explyt.ai/openai/api",
"ANTHROPIC_API_KEY": "<YOUR_API_KEY>",
"CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": 1
}
}
Start Claude Code. When it asks Do you want to use this API key?, answer yes. If it misbehaves later, /logout in the chat and start again.
Step 2: Codex CLI
From the Codex CLI page. Open or create ~/.codex/config.toml:
model = "auto"
model_provider = "explyt"
model_catalog_json = "./model-catalog-full.json"
[model_providers.explyt]
name = "explyt"
base_url = "https://api.explyt.ai/openai/api"
wire_api = "responses"
requires_openai_auth = true
Two more files sit next to it, both described on the docs page: model-catalog-full.json with the list of Explyt models (a ready-made file is linked there) and ~/.codex/auth.json with "auth_mode": "apikey" and your key. If you were signed in through a ChatGPT account before, run codex logout first and then create auth.json. One caveat from the same page: Codex CLI and the Codex app in ChatGPT share config.toml, and the ChatGPT app can only use OpenAI models.
Step 3: Cursor
From the Cursor page. Settings → Models → API Keys:
- Turn on OpenAI API Key and paste the Explyt key.
- Turn on Override OpenAI Base URL and paste
https://api.explyt.ai/openai/api.
Cursor allows custom keys only on a Pro plan or higher; the Cursor docs page spells this out, and the restriction comes from Cursor.
The one rule that applies to every tool
Do not also sign in to the tool's own account through ChatGPT or Anthropic. That sign-in sends requests around the Explyt gateway, and they get rejected.
Everything else that connects
| Tool | Type | Docs |
|---|---|---|
| Claude Code | CLI | page |
| Codex CLI | CLI | page |
| OpenCode | CLI | page |
| Claude Desktop | Desktop app | page |
| Codex in ChatGPT | Desktop app | page |
| Cursor | Editor | page |
| Claude Code for VS Code | Extension | page |
| Kilo Code | Extension | page |
| Cline | Extension | page |
| Explyt plugin (Community mode) | JetBrains IDE plugin | page |
Claude Desktop uses a dedicated gateway dialog: Configure third-party inference → Connection → Gateway, URL into Gateway base URL, key into Gateway API key, scheme x-api-key.
Anything not in the table that accepts an OpenAI-compatible base URL plus key will probably work; we have not verified it. Errors and their meanings are collected on the error messages page.
Which models answer
OpenAI and Anthropic models, plus auto, which picks a model for you and has the highest availability. Many tools fetch the list from the gateway on their own. If yours does not:
curl https://api.explyt.ai/openai/api/models \
-H "Authorization: Bearer <YOUR_API_KEY>"
The only restriction: the Codex app in ChatGPT works with OpenAI models and auto. For Anthropic models in a Codex workflow use Codex CLI.
What the key does not carry
This is the part to read before you tell your team "we have Explyt in Cursor now".
The key gives access to models and nothing else. In an external tool you will not get:
- IDE tools: project navigation, run configurations, the debugger, IDE-powered refactoring;
- skills, modes, rules and checks of the Explyt agent;
- subscription data inside the third-party tool's interface. Remaining minutes and the current chat's spend live in your personal account.
Paying for tokens through Explyt does not make Claude Code aware of your IntelliJ project's run configurations or let it stop on a breakpoint. Those capabilities live in the JetBrains IDE. The route for them is different and older: the Explyt MCP server from 5.17, which exposes selected tool groups to an external agent, does no inference of its own, and needs no subscription.
The key and the MCP server solve different problems and can work together: model through the gateway, IDE tools through MCP, both in one Claude Code session.
If you also use the plugin
The Explyt plugin in a JetBrains IDE has two ways to use the subscription. Personal signs in from the plugin and knows about the subscription: remaining minutes, current chat spend, a live model list with parameters. Community uses the same URL and key as any other tool and sees only an OpenAI-compatible provider. Minutes cost the same either way. Stay on Personal unless you have a reason not to.
Two fixes in the same release
Context indicator. The chat's fill bar used to understate history size by about a third, because agent chats are mostly code and tool output, which tokenize more densely than the prose the estimate assumed. A chat at the model's limit looked half empty, auto-compression did not start, and the request failed at the provider. Now Explyt reconciles the estimate with the input token count the provider reports after each response, and keeps the correction with the chat across IDE restarts. Before the first response you still see the local estimate.
MCP tokens. Servers that authenticate through a browser hand out tokens that live about an hour. The first call after expiry failed. Explyt now renews ahead of expiry, one token per server. After the update, sign in once more for each such server. Manual-header and local servers are untouched.
Also: fewer file-system checks for paths the agent mentioned but that do not exist in the project, and forks and subagents inherit the already compressed history rather than paying for the uncompressed one.
For teams
One Explyt subscription per person covers every client that person uses. In Explyt Enterprise the gateway sits inside the company perimeter; the administrator issues the URL and key, and traffic from Cursor, Claude Code or Codex to the models goes through the corporate platform, not through personal vendor accounts. See the pricing page or contact sales.
Try it on Monday
- Open your Explyt account, copy the key.
- Pick the one tool you open most and paste the config above.
- Send one request, then open the account and confirm the minutes moved.
- If the same tool also needs your IDE, wire the MCP server next; it is a separate setup.
Which tool did you connect first, and did it pick up the model list on its own? Tell us in the comments.
Full release notes: Explyt 5.19 on explyt.ai.
Top comments (0)