Continue

Continue is an open-source coding assistant for VS Code and JetBrains, with a terminal agent called cn. Both read the same config.yaml and accept any OpenAI-compatible endpoint.

Verified against the cn CLI 1.5.47. The config below is re-run against that release every week, including a tool call round trip. Upstream has made the Continue repository read-only; 1.5.47 and the VS Code extension 2.1.0 are its last releases.

Configure

Add LowRouter as a model in ~/.continue/config.yaml:

name: LowRouter
version: 1.0.0
schema: v1
models:
  - name: LowRouter Mistral Large
    provider: openai
    model: auto/mistralai/mistral-large-2512
    apiBase: https://api.lowrouter.ai/v1
    apiKey: ${{ secrets.LOWROUTER_API_KEY }}
    roles: [chat, edit, apply]
    capabilities: [tool_use]

${{ secrets.LOWROUTER_API_KEY }} is read from the environment or from ~/.continue/.env, so the key stays out of the config:

export LOWROUTER_API_KEY=sk-lr-...

The editor extensions pick the model up from the same file. From a terminal, cn runs one task headless:

cn -p --config ~/.continue/config.yaml "Summarise what this repository does"

Details of this config:

• provider: openai is Continue’s OpenAI-compatible client; apiBase ends in /v1.
• capabilities: [tool_use] is required. Continue cannot tell that a LowRouter model supports tools, and without it agent mode runs with no tools at all.
• Continue sends model names that look like OpenAI’s o-series or gpt-5 to the Responses API. LowRouter serves it, so those models work either way; useResponsesApi: false on the model entry turns that off.
• cn -p only runs tools that need no approval. Add --auto to allow all tools, which is what an unattended task usually needs.

Picking a model

Agent mode acts through function calling, so pick a model with the function-calling tag. The model browser filtered to them lists every one. Each model is another entry under models:; the name is what the model picker shows.

The example uses auto/mistralai/mistral-large-2512, served from the EU, which has been through cn’s tool loop in our runs.

Recommended setup

• Use a dedicated key, as with any agent, so its spend is attributed and it can be revoked on its own.
• Set defaultCompletionOptions (contextLength, maxTokens) on a model entry when the model’s limits differ from Continue’s defaults; the model browser lists both.

Troubleshooting

• 404 on every request: apiBase is missing /v1.
• Agent mode never reads or edits files: capabilities: [tool_use] is missing from the model entry.
• 401: LOWROUTER_API_KEY is not set in the environment that started the editor or cn, and not in ~/.continue/.env.