⌁ Guides

Aider

Two environment variables and the openai/ model prefix put Aider on the full Venice catalog, paid per request in USDC on Base.

Create an API key

Sign in with your Base wallet at /keys and mint a Gondola API key. Keys start with gnd_. It authenticates every request and each call is charged in USDC against your Gondola balance.

Top up USDC on Base once at /wallet. The same balance powers the browser chat and every tool you connect.

Point Aider at Gondola

Aider treats Gondola as an OpenAI-compatible provider. Export the base URL and key, then start Aider with the openai/ prefix on the model id:

shell

export OPENAI_API_BASE=https://api.gondola-ai.com/v1
export OPENAI_API_KEY=gnd_xxxx

aider --model openai/claude-sonnet-5

The environment variable names differ across Aider versions, so the command-line flags are the most reliable path if the exports do not take:

shell (flags)

aider \
  --openai-api-base https://api.gondola-ai.com/v1 \
  --openai-api-key gnd_xxxx \
  --model openai/claude-sonnet-5

Base URL gotcha

Set the base to api.gondola-ai.com/v1 only. Do not append /chat/completions; Aider adds the path itself.

Pick a model

Gondola serves the full Venice catalog. Model ids must be typed exactly as the catalog lists them (keep the openai/ prefix in front of the id). Pull the live list:

shell

curl https://api.gondola-ai.com/v1/models

Each entry includes context_length and max_completion_tokens (the limits the gateway actually enforces, so copying them is always safe). Example ids today include claude-sonnet-5 and qwen3-coder-480b-a35b-instruct-turbo, but always confirm against the live list, since the catalog changes.

Prompt privacy

Gondola never stores prompts or completions: requests are metered, forwarded, and gone when the response finishes (the one exception, a video source clip over 18 MiB, is held privately for the length of its job; see /privacy), and Venice runs a no-logging model underneath. For zero retention on Venice-controlled infrastructure too, pick a model whose privacy field reads private on GET /v1/models. Details on /privacy.

Notes and troubleshooting

Silence the unknown-model warning

Aider warns that it has no metadata for a model it does not recognize. Give it the context window by adding an .aider.model.metadata.json file (use the real numbers from the models list):

.aider.model.metadata.json

{
  "openai/claude-sonnet-5": {
    "max_input_tokens": 1000000,
    "max_output_tokens": 32768,
    "litellm_provider": "openai",
    "mode": "chat"
  }
}

Cap the price on unattended runs

Aider's architect mode and scripted runs can fire off many requests unwatched. If your Aider config lets you add a custom request header, set x-gondola-max-factor to a value from 0 to 1 to refuse any request priced above that ceiling instead of paying more; omit it and a request with no cheaper capacity can fall through to Gondola's own backup key at Venice list price. It also works when you call the endpoint directly for scripted runs. For a hard dollar limit rather than a relative one, send x-gondola-max-charge-usd (for example 0.50) and any request whose estimated charge would exceed it is refused with a 402 before anything is billed.

The openai/ prefix is required. Without it, Aider does not route the request through the OpenAI-compatible provider and the call fails. If a request returns insufficient balance, top up at /wallet.

Loading page