Aider

Aider is an open-source AI pair programmer that runs in your terminal, edits files in your git repo, and commits each change with a message it writes itself. Every model call goes through LiteLLM, so Aider already speaks OpenAI-compatible Chat Completions — pointing it at Melious is two environment variables and an openai/ prefix on the model name. Add two small files and Aider edits with the same diff format and cost tracking it gives the models it ships metadata for.

Setup

python -m pip install aider-install
aider-install

uv tool install --force --python python3.12 --with pip aider-chat@latest

export OPENAI_API_BASE=https://api.melious.ai/v1
export OPENAI_API_KEY=sk-mel-<YOUR_API_KEY>

$env:OPENAI_API_BASE = "https://api.melious.ai/v1"
$env:OPENAI_API_KEY = "sk-mel-<YOUR_API_KEY>"

OPENAI_API_BASE=https://api.melious.ai/v1
OPENAI_API_KEY=sk-mel-<YOUR_API_KEY>

model: openai/glm-5.2
openai-api-base: https://api.melious.ai/v1
openai-api-key: sk-mel-<YOUR_API_KEY>

aider --model openai/glm-5.2

Warning for openai/glm-5.2: Unknown context window size and costs, using sane defaults.
Did you mean one of these?
- openai/gpt-5.2
- zai/glm-5.2
You can skip this check with --no-show-model-warnings

Aider v0.86.2
Model: openai/glm-5.2 with whole edit format
Git repo: .git with 1 files
Repo-map: using 1024 tokens, auto refresh

- name: openai/glm-5.2
  edit_format: diff
  use_repo_map: true
  accepts_settings: ["reasoning_effort"]

{
  "openai/glm-5.2": {
    "max_input_tokens": 1000000,
    "input_cost_per_token": 0.000001,
    "output_cost_per_token": 0.000004,
    "litellm_provider": "openai",
    "mode": "chat"
  }
}

Aider v0.86.2
Model: openai/glm-5.2 with diff edit format
Git repo: .git with 1 files
Repo-map: using 4096 tokens, auto refresh

export MELIOUS_API_KEY=sk-mel-<YOUR_API_KEY>
aider --model melious

Picking a model

Aider doesn't use tool calls. It asks the model to write edits as text blocks and applies them itself, so what matters is a model that follows a strict output format across a long conversation. Browse melious.ai/hub/models, or GET /v1/models?include_meta=true for context length and pricing.

Images work on the models that take them: add "supports_vision": true to the model's metadata entry and pass the file like any other. Without that flag Aider quietly leaves the image out of the request.

Name a model every time. With no --model, Aider picks a default from the keys it finds, sees OPENAI_API_KEY, and announces Using gpt-4o model with API key from environment — a model we don't run. Pin it with model: in .aider.conf.yml if you'd rather not pass the flag.

aider --list-models <name> won't help you browse our catalog: it searches LiteLLM's built-in table plus whatever you put in your metadata file, so it returns other providers' copies of a model name. Use the models hub or GET /v1/models.

Aider also uses a weak model for commit messages and chat summaries, and falls back to your main model when you don't name one. It sends that model the diff plus the chat history, so pick something cheap — and non-reasoning:

aider --model openai/glm-5.2 --weak-model openai/mistral-small-3.2-24b-instruct

A reasoning model is the wrong fit here. Aider waits for the whole reply, and a small hybrid model can spend thousands of reasoning tokens on a one-line commit message. Check _meta.reasoning_type for non_reasoning.

Architect mode splits the work between two models — one plans the change, the other writes the edits:

aider --architect --model openai/glm-5.2 --editor-model openai/qwen3-coder-next

Append a flavor suffix to bias provider selection — openai/glm-5.2:speed, :price, or :eco. See Routing for the decision table.

What's different

• OPENAI_API_BASE is global to the transport. In a shell where it's exported, every openai/ model goes to Melious — including OpenAI's own, if you switch to one mid-project. Keep the pair in a project .env when you use both.
• Settings match the exact model name. openai/glm-5.2:speed is a different name from openai/glm-5.2, so a flavored model falls back to whole edits until the settings file has an entry for it, and loses cost reporting until the metadata file does too.
• Reasoning shows up in the transcript. Hybrid and reasoning models stream their thinking, and Aider prints it under a THINKING heading before the ANSWER. Those tokens bill as output.
• Costs print in dollars. Aider always labels cost with $, but the number comes from the rates in your metadata file. Enter our EUR rates and read the figure as EUR.
• No Melious CLI support. melious tools install covers Claude Code, Codex, OpenCode, Mistral Vibe, and Pi. Aider you wire by hand.
• Caching happens, but Aider can't see it. We cache prefixes automatically — a repeated 6.2k-token prompt comes back with cached_tokens: 6144 — and bill reads at the cheaper rate. Aider looks for DeepSeek's prompt_cache_hit_tokens or Anthropic's cache_read_input_tokens, not OpenAI's cached_tokens, so it never prints a cache-hit line and the cost it shows is the uncached one. --cache-prompts changes nothing here. See Models.
• Extra response fields are ignored. Non-streaming responses carry environment_impact and billing_cost; streamed chunks don't. Aider doesn't read either; your usage dashboard aggregates them.

When it breaks

• Using gpt-4o model with API key from environment, then Model not found: gpt-4o — you didn't pass --model, so Aider picked a default for the OpenAI key it found. Name a Melious model.
• LLM Provider NOT provided. Pass in the LLM provider you are trying to call. You passed model=glm-5.2 — the openai/ prefix is missing. LiteLLM can't tell which transport to use without it.
• NotFoundError: OpenAIException - The requested resource was not found. — the model ID is fine, but OPENAI_API_BASE is missing /v1.
• APIConnectionError: OpenAIException - Model not found: <MODEL_ID> — a model ID we don't run. We report it as an error event inside the stream, which LiteLLM surfaces as a dropped connection rather than a clean 404, so the failure looks like a network problem instead of a bad model ID. That's on us. Check the ID with GET /v1/models.
• AuthenticationError: OpenAIException - Invalid API key format — the key is truncated or mistyped. Copy it again from melious.ai/account/api/keys.
• OPENAI_API_KEY: Not set — Aider warns before the first prompt, and the first request then fails with The api_key client option must be set. Export the key in the shell that launches Aider, or add it to .env.
• Unknown context window size and costs, using sane defaults — Aider has no metadata for the model name you passed. Add it to .aider.model.metadata.json, spelled exactly as on the command line.
• Aider applies the edit, then sits silent — it's waiting on the commit message from a reasoning weak model. Switch --weak-model to a non-reasoning model.
• 429 mid-session — per-plan token caps, and Aider resends the repository map and chat history on every turn. Rate limits covers which plan lifts which limit.

Errors and retry patterns: Errors.