How to Set Up Continue.dev with an OpenAI-Compatible API (AtmoRouter)

Continue.dev talks to models through providers. If you want Claude, GPT-class, GLM, or DeepSeek models inside VS Code or JetBrains without wiring a custom plugin for each vendor, the OpenAI-compatible path is the shortest one: set provider: openai, override apiBase, and use exact catalogue model IDs. This guide wires Continue to AtmoRouter that way.

Prices and IDs below were read from AtmoRouter's public /v1/models response on November 7, 2026 (Asia/Jakarta). Rates refresh — treat the models catalogue as authoritative. This is an AtmoRouter-authored setup guide, not an independent review of Continue.

What you need

• Continue for VS Code or JetBrains (config lives in ~/.continue/config.yaml on macOS/Linux, or %USERPROFILE%\.continue\config.yaml on Windows).
• An AtmoRouter API key from the API Keys page.
• A model ID copied from the live catalogue — family prefix included (for example cc/claude-sonnet-4-6, not a bare marketing name).

Chat Completions base URL: https://atmorouter.dev/v1. Do not paste /chat/completions into apiBase; Continue appends the path.

Minimal config.yaml

Continue's OpenAI provider documents apiBase for compatible gateways. A working AtmoRouter block looks like this:

yamlname: AtmoRouter Continue
version: 0.0.1
schema: v1

models:
  - name: AtmoRouter Claude Sonnet 4.6
    provider: openai
    model: cc/claude-sonnet-4-6
    apiBase: https://atmorouter.dev/v1
    apiKey: ${{ secrets.ATMOROUTER_API_KEY }}
    roles:
      - chat
      - edit
      - apply
    capabilities:
      - tool_use
    defaultCompletionOptions:
      maxTokens: 8192

  - name: AtmoRouter GLM 5.3
    provider: openai
    model: ali/glm-5.3
    apiBase: https://atmorouter.dev/v1
    apiKey: ${{ secrets.ATMOROUTER_API_KEY }}
    roles:
      - chat
      - edit
      - apply
    capabilities:
      - tool_use

Put the key in .continue/.env or a workspace .env as ATMOROUTER_API_KEY=ar-xxxxxxxxxxxx so you do not commit it. Save config.yaml; Continue reloads on save.

For Agent mode, set capabilities: [tool_use] explicitly. Continue only auto-detects native tools for a short list of well-known OpenAI model names; a custom ID such as ali/glm-5.3 otherwise falls back to a weaker prompt-based tool path.

If a GPT-class ID fails with Responses API or organization-verification errors, add useResponsesApi: false on that model entry so Continue stays on /chat/completions.

Pick models with live rates (snapshot)

AtmoRouter exposes 150+ models on one OpenAI-compatible endpoint. Lowest listed tiers on November 7, 2026 for a few Continue-friendly IDs (USD per million tokens):

| Model ID | Context | From input / output | Cached input | |---|---|---|---| | cc/claude-sonnet-4-6 | 1M | $0.525 / $2.625 | $0.0525 | | cc/claude-haiku-4-5 | 200K | $0.175 / $0.875 | $0.0175 | | cb/gpt-5.3-codex | 1M | $0.153125 / $1.225 | $0.0153125 | | ali/glm-5.3 | 1M | $0.028 / $0.088 | $0.0028 | | ali/deepseek-v4-pro-0813 | 1M | $0.00165 / $0.00495 | $0.000165 |

Higher tiers exist when the network routes elsewhere. Cached input on AtmoRouter is billed at 10% of the input rate where the model supports cache. Prefer a cheap chat/edit model for most turns and reserve Sonnet-class IDs for harder refactors.

Billing notes that stay true regardless of the model: prepaid balance, no seat fees, and failed requests are not billed. Check the pricing page when you top up.

Verify outside Continue first

Before debugging the IDE extension, prove the key, base URL, and model ID with curl:

bashcurl https://atmorouter.dev/v1/chat/completions \
  -H "Authorization: Bearer $ATMOROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ali/glm-5.3",
    "messages": [{"role": "user", "content": "Reply with only: continue-ok"}],
    "max_tokens": 20
  }'

If curl fails, fix the key, model ID, or network path before touching Continue settings. If curl succeeds but Continue fails, the usual culprits are a typo in apiBase (missing /v1, or a full /chat/completions URL), a stale secret name, or Agent mode without tool_use.

A practical Continue layout

Many teams run two or three AtmoRouter entries in one config:

1. 1.Daily chat/edit — ali/glm-5.3 or ali/deepseek-v4-pro-0813 for volume.
2. 2.Harder edits — cc/claude-sonnet-4-6 or cb/gpt-5.3-codex when the cheap model stalls.
3. 3.Fast/cheap autocomplete (optional) — a flash-tier ID with roles: [autocomplete] only if you actually use Continue autocomplete; otherwise leave that role off so you do not burn tokens by accident.

Switching models is a dropdown change in Continue, not a new integration. That is the point of a shared OpenAI-compatible gateway: one key, many IDs.

FAQ

Does Continue need a special AtmoRouter provider?

No. Use provider: openai with apiBase: https://atmorouter.dev/v1. Continue documents this pattern for OpenAI-compatible gateways.

Can I use Anthropic-format Claude Code settings here?

This guide is for Continue's OpenAI provider. Claude Code and the Anthropic SDK use a different base URL shape (https://atmorouter.dev without /v1). See the Anthropic-compatible endpoint and Cline setup posts if that is your stack.

Why does Agent mode ignore my tools?

Custom model IDs often need capabilities: [tool_use]. Without it, Continue may not call native function calling even when the gateway supports tools.

Are the prices above guaranteed?

No. They are a live snapshot. Always re-check Models before you budget a project.

Next step

Create a key, paste the YAML above, run the curl check, then send one low-risk chat from Continue. When you want adjacent setups, the OpenCode guide and prompt caching notes cover the same gateway from other agents.