Coding agents
Coding agents are where inference spend adds up fastest — long sessions, large context, many turns. Every agent below reaches Cloakray through a base URL and a key, with no change to how you use the tool.
Two things decide which setup a given agent needs:
| The agent speaks | Base URL | Credential |
|---|---|---|
| OpenAI Chat Completions | https://api.cloakray.com/v1 | Authorization: Bearer |
| OpenAI Responses | https://api.cloakray.com/v1 | Authorization: Bearer |
| Anthropic Messages | https://api.cloakray.com | x-api-key or Authorization: Bearer |
Pick your agent
- Claude Code — two environment variables, Anthropic format
- Codex — a provider block in
config.toml, Responses format - opencode — a provider block in
opencode.json - Cline — VS Code, OpenAI Compatible provider
- Roo Code — VS Code, OpenAI Compatible provider
- Aider — two environment variables and a model prefix
- Continue — an
apiBaseinconfig.yaml - Cursor — override the OpenAI base URL in Cursor's settings
- Kilo Code — VS Code, OpenAI Compatible provider
- GitHub Copilot — a custom OpenAI-compatible model provider
- Zed — an
openai_compatibleprovider in settings - Crush — a provider block in
crush.json - OpenClaw — a provider block in
openclaw.json - Hermes Agent — a custom endpoint in
config.yaml - Pi — a provider block in
models.json
Any other agent that accepts an OpenAI-compatible base URL works the same way. To let an agent call Cloakray models as tools instead, see MCP.
Model pickers
Cloakray answers GET /v1/models with the catalogue, so agents that populate a dropdown from the API will find one. Agents that don't ask — opencode and Continue declare their models in config — take the id from the models page instead. Ids look like claude-opus-4-8 or gpt-5.5, and every text model works on all three text endpoints (image models answer on /v1/images/generations instead).
Keeping costs down
- Pick a model with a low cached-input price: agents resend most of their context every turn, so prompt caching does most of the saving.
- Use a reasoning effort that fits the task; the highest setting can multiply output tokens.
Keep the key out of your repo
Agent config files are easy to commit by accident. Prefer an environment variable and reference it from the config where the agent supports that, and add any file holding a literal key to .gitignore.