Compatibility notes
Tools people ask about that are not in the verified list, and why.
Not currently supported
Cursor. Cursor’s agent mode with a custom API endpoint currently has known limitations upstream,
so we do not list it as supported. Cursor’s non-agent features that speak plain chat completions can
still be pointed at https://api.devotel.com/v1 — see
Using with coding tools — but the autonomous agent flow is not something we can
recommend today.
OpenAI Codex CLI. Codex CLI talks to OpenAI’s Responses API (/v1/responses). We expose
Chat Completions (/v1/chat/completions) and Models (/v1/models), not Responses, so the CLI cannot
connect regardless of base URL.
Claude Code. Claude Code speaks Anthropic’s Messages API (/v1/messages), a different wire
format from the OpenAI one we serve. Support for it is on our roadmap; until then, pointing Claude
Code at api.devotel.com will not work.
Unverified
GitHub Copilot / VS Code BYOK. Recent VS Code and Copilot builds allow a “bring your own key”
custom provider, and the OpenAI Compatible option there looks like it should accept
https://api.devotel.com/v1 with kimi-k3. We have not verified it, so treat it as unknown rather
than supported. If you try it, tell your Devotel administrator how it went — this page will be
updated.
For a VS Code experience we have actually verified, use Cline or Roo Code.
What makes a tool compatible
A tool works with Devotel AI when it:
- lets you set the base URL (
https://api.devotel.com/v1), - lets you set the API key (your
dvt-key), and - speaks Chat Completions with tool calling.
That is the whole list. Anything requiring a provider-specific endpoint — Responses, Messages, Assistants, embeddings — will not work until we expose that endpoint. See Using OpenAI SDKs for the current surface.
Reasoning output and agent harnesses
kimi-k3 is a reasoning model: alongside the answer in content, responses carry a
reasoning_content field with the model’s thinking, and streamed responses deliver
delta.reasoning_content before the answer begins.
Every harness on the verified list handles this automatically — they render or discard the
reasoning and act on content. You only need to think about it if you are writing your own
client, in which case see Chat Completions.
Two practical effects when running agents:
- Reasoning tokens are billed as output tokens. They are counted inside
completion_tokens, so an agent’s spend is driven by how hard it thinks, not by the size of the diff it produces. - There is a pause before visible output. The model reasons first. A harness with an aggressive timeout may give up before the answer starts; raise the timeout rather than assuming a failure.
A tool that concatenates every string field it finds in the response may show you the model’s
thinking as though it were part of the answer. That is a bug in the tool, not in the response —
content and reasoning_content are separate fields for exactly this reason.