Home → Guides → Cline with a custom API: provider, base URL and which model handles its tools
Cline needs three things from an API: a base URL, a key and a model ID. The part that decides whether an agent run works is less obvious — whether the model calls tools with exactly the arguments Cline defines. We tested that on six models before writing this.
Open the Cline panel in VS Code and click the gear icon. Then:
| API Provider | Base URL | Model ID |
|---|---|---|
| OpenAI Compatible | https://apiclan.us/v1 | gpt-6.1-sol, deepseek-v4-pro, claude-opus-5-5 |
| Anthropic (custom base URL) | https://apiclan.us | claude-opus-5-5, claude-fable-5-1 |
The two base URLs differ on purpose. The OpenAI Compatible provider appends /chat/completions, so it needs /v1; the Anthropic provider appends /v1/messages itself, so it must not have it. A /v1 in the wrong place is the most common cause of a 404.
Every API key on this gateway belongs to one model group, and it can only call that group’s models. To use Claude in Cline, create the key in a Claude group; for GPT, in a GPT group; for DeepSeek, in the DeepSeek group. If you switch model families in Cline, switch the key with them.
With OpenAI Compatible, Cline cannot look the model up, so it shows Model Configuration fields: context window, maximum output tokens and image support. Set the context window to the model’s real limit. Too low and Cline trims the conversation early; too high and long tasks fail with a context-length error instead of being compacted.
Cline’s agent works through tools such as read_file. We gave each model a read_file tool whose single argument is named path, asked it to open src/app.ts, and checked two things: that it called the tool with exactly the argument name the schema declares, and how long the first response took (median, streaming on). A renamed argument is the usual reason an agent step fails without a clear error.
| Model | Correct tool calls | First response |
|---|---|---|
claude-opus-5-5 | 13 / 13 | 3.3 s |
claude-fable-5-1 | 13 / 13 | 2.8 s |
gpt-6.1-sol | 3 / 3 | 2.6 s |
gpt-5.6-sol | 3 / 3 | 3.3 s |
deepseek-v4-pro | 3 / 3 | 2.6 s |
deepseek-v4.1-flash | 3 / 3 | 1.5 s |
Every model called the tool correctly every time. The samples are small — 13 runs each for the Claude models, 3 for the others — and one gpt-5.6-sol run took 40 seconds on a cold start; the table shows the median. Tested 12 October 2026.
claude-opus-5-5 through the Anthropic provider — Claude’s native protocol, and the one on which Cline sends prompt-cache markers.gpt-6.1-sol or claude-fable-5-1.deepseek-v4.1-flash, the quickest to answer in our test./v1, Anthropic must not have it.is not supported by any configured account in this group) → the model ID is misspelled, or your key belongs to a group without that model.https://api.apiclan.us/v1 (or https://api.apiclan.us for Anthropic), which avoids the 100-second proxy limit.Setting up other tools? See Claude Code base URL and Codex with a custom API.
No subscription, no monthly minimum, no sales call. Top up with USDT and spend what you use — 1 USDT gives you 2 credits of API balance.
Read the 30-second quickstartLanguage: English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Deutsch · Français · Español · Türkçe · Português · Italiano · Bahasa Indonesia · Tiếng Việt · Bahasa Melayu · Polski · Čeština · فارسی · Русский
Prices on this page are filled in automatically from live billing data. Third-party terms are quoted from that party's own documentation.