dev.kurlo/kurlo

## Kurlo 2.5.0

### Added

- **AI requests.** A new request type sends a prompt to a language model: OpenAI, Anthropic, OpenRouter, Groq, Ollama, LM Studio, or any server with an OpenAI-compatible API. Choose the provider and the model next to the URL (**Load models** reads the provider's list), and write a system prompt and a prompt with `{{variables}}`. The answer streams into a conversation that shows the model's thinking, each tool call with its arguments and result, the stop reason, the time to the first token, the speed and the tokens in and out. **Reply** continues the conversation, **Raw** shows each stream as it arrived, and an answer cut off by the token limit or an error in the middle of a stream is called out where it happened. API keys come from `{{openaiApiKey}}`, `{{anthropicApiKey}}` and the like, so they stay out of the YAML.
- **MCP servers as tools for the model.** The **Tools** tab of an AI request lists the MCP requests in the workspace. Tick one, and when the model calls one of its tools Kurlo runs the call with that request's URL, auth and headers and gives the result back, step by step, until the model answers or the step limit is reached.
- **Markdown docs for every request.** The **Docs** tab has **Edit** with a live preview and **Preview**, which renders headings, lists, tables, code, links and images and adds a reference built from the request: the method and URL, the auth type, and the path variables, query params, headers and body with each row's description.
- **A status bar button for the response layout.** It puts the response beside or below the request; before, this was only in the command palette. Choosing the side-by-side layout yourself keeps it on editors down to 760 px wide.
- **`pm.response.llm()` in tests.** Test scripts of an AI request get the answer in one shape for every provider: the text, the tool calls with their parsed input, the stop reason, the tokens and the timings. Tests run once, after the last step.

### Changed

- **The AI conversation reads like a chat.** Answers are rendered as Markdown, the conversation uses the whole width of the panel, and the message field sends on `Enter` (`Shift+Enter` for a new line) without the system's spelling suggestions. **New chat** moved to the status line, and the Prompt tab says which text goes with every message and which only starts the chat.
- **Raw stays quick in long conversations.** Each step shows its URL and request body, steps are folded except the last, long bodies show their first 20,000 characters with **Show all**, and earlier sends no longer keep their raw bodies in memory.
- **The prompt field fits its text.** It starts at about six lines and grows as you write, up to the free height of the tab, instead of filling the whole tab.
- **The full address under the base URL.** An AI request shows where it sends the prompt, such as `POST https://api.deepseek.com/v1/chat/completions`, with the path Kurlo adds highlighted, so it is clear that the field takes the base URL.
- **Clearer errors when the model list does not load.** The provider and model box says what went wrong in a short title, shows the status and the address Kurlo asked (without its query string, where keys can hide), and suggests what to do next, such as checking the base URL or typing the model name.
- **Length limits for prompts.** The prompt and each chat message take up to 200,000 characters, the system prompt up to 50,000. A paste that does not fit is cut at the limit, a counter appears next to the label as you get close, and a longer prompt from an edited YAML file is refused when you send.
- **The method menu is shorter.** HTTP methods and the other protocols sit side by side.

### Fixed

- Clicking the provider and model box while it was open closed it and opened it again at once in the app; it now closes. The gRPC method picker had the same problem.
- With the response beside the request, the request side could be dragged so narrow that settings fields w