Use a capped, OpenAI-compatible coding model in Codex CLI, Cline and Aider

dev.to2026-10-10 22:080 views

Disclosure: I build ElderAI, the service used in this tutorial. Drafted with AI help and checked by me.

Coding agents run in loops. Loops retry. On a pay-as-you-go API that can quietly turn into a surprise bill overnight. The fix I wanted was boring: a hard cap at the API gateway. When the credits are gone, requests stop with a clear error. No overage.

This post shows how to point Codex CLI, Cline and Aider at an OpenAI-compatible endpoint that works that way (ElderAI's atlas-code-1-preview). The same steps work for any OpenAI-compatible API; just swap the base URL and model id.

1. Get a key

export ATLAS_API_KEY=elk_...

2. Smoke test with curl

curl https://elderai.cloud/v1/chat/completions \ -H "Authorization: Bearer $ATLAS_API_KEY" -H "Content-Type: application/json" \ -d '{"model":"atlas-code-1-preview","messages":[{"role":"user","content":"Write a pytest test for a function add(a, b)."}]}'

If the GPU was idle, the first request can take about a minute while it wakes up. Raise client timeouts (300 s is safe).

3. Codex CLI

Codex talks the Responses API, which the endpoint supports. In ~/.codex/config.toml:

model_provider = "atlas" model = "atlas-code-1-preview" [model_providers.atlas] name = "ElderAI" base_url = "https://elderai.cloud/v1" env_key = "ATLAS_API_KEY" wire_api = "responses"

Without the top-level model_provider line, Codex ignores the provider block and calls its default endpoint, which is the most common setup mistake.

4. Cline

Settings → API Provider: OpenAI Compatible

Base URL: https://elderai.cloud/v1
API key: your key
Model ID: atlas-code-1-preview
Context window: 24576, max output: 8192, images off

5. Aider

export OPENAI_API_BASE=https://elderai.cloud/v1 export OPENAI_API_KEY=$ATLAS_API_KEY aider --model openai/atlas-code-1-preview --timeout 300 --no-show-release-notes

6. What to use it for (honestly)

It's a small (8B) preview model. In my own tests it's useful for tests, boilerplate, renames and small refactors. It is not the model for multi-file architecture changes, and it does make mistakes, so run your tests. Good pattern: route the cheap, repetitive agent steps to a capped small model and keep a bigger model for the hard calls.