Docs · Providers
Volcengine Ark Coding Plan in Claude Code and Codex
Volcengine Ark (火山方舟) is sold three ways: the Coding Plan, the Agent Plan and pay as you go. Each plan has endpoints of its own, and its quota is spent only there; a request to pay-as-you-go's /api/v3 is billed apart. Add the key in magpie with its plan, and Claude Code, Codex, Gemini CLI, OpenCode and other tools use Ark's Doubao, GLM, DeepSeek and Kimi models through magpie's local gateway on 127.0.0.1:3425.
Add Volcengine Ark
- Get a keyCreate one on Ark's API key page (ark.volcengine.com). A plan's quota is spent only once the plan is subscribed on its own page.
- Add the provider and pick the planOn Providers, click ＋ Add provider, pick Volcengine Ark, set the key's plan to Coding Plan, Agent Plan or Pay as you go, and save.
- Pick it for a toolOn Agents, pick a model such as ark-code-latest or doubao-seed-2.1-pro on a tool's row.
magpie claude volcengine/ark-code-latest magpie codex volcengine/doubao-seed-2.1-pro magpie opencode volcengine/glm-5.3
Plans and endpoints
| Plan | Base URL (Chat Completions and Responses) | Anthropic Messages |
|---|---|---|
| Coding Plan | https://ark.cn-beijing.volces.com/api/coding/v3 | https://ark.cn-beijing.volces.com/api/coding |
| Agent Plan | https://ark.cn-beijing.volces.com/api/plan/v3 | https://ark.cn-beijing.volces.com/api/plan |
| Pay as you go | https://ark.cn-beijing.volces.com/api/v3 | — (no Anthropic endpoint in the preset) |
The plans' paths carry their version and are used as written; no /v1 goes on them. Each key in the provider names its own plan, so one provider can hold Coding Plan, Agent Plan and pay-as-you-go keys, and each key's requests go to its own plan's endpoint.
The plans' models
Both plans list the same models. Their /models doesn't give them (the Coding Plan's lists Ark's general catalog, the Agent Plan's answers 404), so magpie uses the text models Ark's plan docs list, ark-code-latest first: ark-code-latest, doubao-seed-evolving, doubao-seed-2.1-pro, doubao-seed-2.1-lite, doubao-seed-2.0-mini, glm-5.3, deepseek-v4-pro, deepseek-v4.1-flash, kimi-k3 and kimi-k3.1-pro, among others. Which of them take images is the docs' word too: glm-5.3-flash and deepseek-v4.1-flash take an image as it is. A pay-as-you-go key reads Ark's own model list.
In each agent
- Claude Code speaks Anthropic Messages. The plans have Anthropic endpoints of their own (
/api/coding,/api/plan), so requests go as they are and spend the plan. See Claude Code with other models. - Codex speaks only OpenAI Responses. Ark's plan endpoints serve Responses themselves, so Codex's requests go as they are, and the models show in Codex's
/model. See Codex with other models. - Gemini CLI speaks Gemini's API; magpie points it at the local gateway and translates. See Gemini CLI with other models.
- OpenCode: magpie writes the models into opencode.json. See OpenCode with other models.
Web search
Ark's Responses API runs the web_search tool itself, on pay as you go and both plans; its Chat API doesn't. So Codex's web search goes to Ark to run instead of being taken out of the request. include values Ark refuses are left out, so the request isn't turned away whole.
Images and video
The Agent Plan also has Seedream image models (doubao-seedream-5-0-pro, doubao-seedream-5.0-flash) and Seedance video models (doubao-seedance-2.0, 2.0-fast, 2.0-mini, 2.5). Pick them in Settings → Images to use them through magpie's image and video tools. Ark has no OpenAI-shaped videos API, so magpie starts and polls Ark's asynchronous generation tasks.
Usage windows
The Coding Plan and Agent Plan have 5-hour, weekly and monthly windows. Ark tells them only to the account's control plane (open.volcengineapi.com), never to the inference key, so the provider's editor asks for the account's AccessKey ID and Secret. With them, the row shows what is left of the plan; pay as you go has no windows.
FAQ
How do I use the Volcengine Ark Coding Plan in Claude Code?
Add Volcengine Ark in magpie with the key's plan set to Coding Plan, then pick a model such as ark-code-latest on Claude Code's row, or run magpie claude volcengine/ark-code-latest. Requests go to the plan's Anthropic endpoint, https://ark.cn-beijing.volces.com/api/coding, and spend the plan.
Why is a Coding Plan key billed as pay as you go?
A plan's quota is spent only at its own endpoints: /api/coding/v3 and /api/coding for the Coding Plan, /api/plan/v3 and /api/plan for the Agent Plan. /api/v3 is pay as you go. Set the key's plan in magpie and its requests go to the plan's endpoint.
Does Volcengine Ark work in Codex?
Yes. The plan endpoints serve OpenAI Responses, so magpie passes Codex's requests through, and Ark runs web search itself. Run magpie codex volcengine/doubao-seed-2.1-pro, or pick it on Codex's row.
How do I see what is left of an Ark plan?
Edit the Volcengine Ark provider and add the account's AccessKey ID and Secret. Ark tells the windows only through its control plane; the row then shows the 5-hour, weekly and monthly windows.
The same for other providers: GLM Coding Plan, Kimi, DeepSeek, Qwen. Per-agent guides: Claude Code with other models, Codex with other models. Hit an error? See troubleshooting, or ask on Discord.