Skip to main content
Aider is a terminal pair-programming CLI. It runs in your repository, it configures more than one model, and its repository map needs no embeddings, which makes it one of the harnesses least likely to trip over the missing endpoint. Read coding harnesses first.

​
The configuration surface

Aider takes the base URL either as a flag or from the environment:
The flag form is --openai-api-base, and its own environment variable is AIDER_OPENAI_API_BASE. Note the variable name: Aider’s documentation names OPENAI_API_BASE, and does not mention OPENAI_BASE_URL, which is the name the official OpenAI SDKs read. They are different strings and only the first is documented here, so set that one. Include the version segment in the value. Aider’s own documentation gives no example value for a compatible endpoint, so this comes from the mechanism rather than from a quoted example: the endpoint path is appended to what you give it. The project carries a report of a version where the segment was dropped from the outgoing path and requests answered 404 as a result, so if every request comes back 404 with endpoint_not_supported, print the path Aider is calling before assuming the key or the model is wrong.

​
The three models

This is the edit people miss on Aider specifically. It configures up to three models separately, and each one needs its own pinned id: --weak-model and --editor-model fall back to defaults rather than to your main model in some configurations, and a default is a bare vendor name, which this API refuses with model_id_not_pinned. Set all three explicitly if you use architect mode or let Aider write your commit messages.

​
The model id, and the prefix collision

Aider’s documented way of naming a model on a compatible endpoint is to prefix it with openai/. That prefix is a provider selector for the library underneath Aider rather than part of the model name, and it is consumed to choose the adapter, with the rest forwarded. Our ids already begin with the provider that serves them, and for models served by OpenAI that provider is also openai. So a pinned id and Aider’s prefix convention produce a doubled first segment:
We did not verify what reaches the wire in that case. The library’s own issue tracker shows doubled and ambiguous prefixes being handled inconsistently, so we will not tell you which of the two segments survives. What we can tell you is how to find out in one request: whatever you configure, a 400 with model_id_not_pinned means only one segment arrived, and a 404 with model_not_found means both did. The refusal names which, and neither costs anything. An id whose provider is not openai does not have this problem, because the prefix and the provider are then different strings.

​
The repository map does not use embeddings

Aider builds its repository map with a parser and a graph ranking over your source files. It calls no embeddings endpoint, and the project has never shipped an embeddings-based index. So the feature that would break first on other harnesses is not a concern here, and Aider needs nothing configured for it.

​
Two things that cost money quietly

Prompt caching is provider-specific, and the keepalive is extra requests. Aider’s --cache-prompts is off by default and its documentation names specific vendor families it works on, so whether it does anything for the model you pinned is a question about that model rather than about this API. --cache-keepalive-pings is the one to think about before enabling: it keeps a cache warm by sending periodic requests, and each of those is a request here, billed and recorded in Activity like any other. Aider does not enforce token limits. Its documentation is explicit that it reports the provider’s limit errors rather than pre-empting them, and that for a model it does not recognise it assumes an unknown context window and no cost, warning you once. Ours is a model it does not recognise, so expect that warning and expect Aider’s own cost display to be empty or wrong. What a request cost is in your usage record. You can supply the missing figures with a model metadata file, .aider.model.metadata.json, or --model-metadata-file. Take the context window and maximum output from GET /v1/models, which publishes both per model.

​
Keep the key out of the repository

Aider reads .aider.conf.yml and .env from your home directory, your repository root and the current directory, in that order, with the later files overriding. Two of those three locations are inside the repository. Aider can add its own files to .gitignore for you, and the safe habit is the home-directory copy plus the environment variables above.

​
What we did not verify

  • That an Aider session against this API succeeds. Nothing was run.
  • What model string reaches the wire for a doubled provider prefix. Described above with the way to find out.
  • Whether --list-models queries this API or matches against the library’s own local model database. Its wording suggests the local database, so do not expect it to show our catalog; ask GET /v1/models for that.
  • Whether Aider honours OPENAI_BASE_URL as well as OPENAI_API_BASE. Only the second is documented, so only the second is written above.
  • Whether the openai adapter path can be made to call anything other than chat completions. We found no documentation either way, and this API serves only chat completions among the inference endpoints, so a configuration that reached another one would answer 404.