AI requests

An AI request sends a prompt to a language model and shows its answer as it streams in. You pick the provider and model, write a system prompt and a user prompt with {{variables}}, and can give the model tools from your saved MCP servers, which Cloak Proxy runs for it in a loop. Use it to try prompts against your API’s data, to compare models, to test an MCP server the way an assistant would use it, or to check model output in the collection runner.

Send a prompt

1. Store your provider’s API key in the vault. New AI requests look for a vault entry named anthropic_api_key. Name keys for other providers as you like, for example openai_api_key.

2. In the API client, click New › Other Protocols › AI, or right-click a collection and choose Add Other Request › AI. Search Everything (⇧ ⌘ A) also has New AI Request.

3. Choose a provider and a model at the top of the request.

4. Open Settings and check the API key field references your vault entry, for example {{vault:openai_api_key}}.

5. Write your prompt in Prompt, and optionally a system prompt in System.

6. Press ⌘ ↩ or click Send. The answer streams into Response.

1234567

A new AI request. The answer streams into Response; the log on the right shows the request and any tool calls.

1. Provider
2. Model from the suggestions, or Default
3. Any model id
4. Send (⌘↩)
5. MCP Tools: give the model tools
6. New Conversation clears the history
7. Messages log

×

Providers and models

Provider	API	Default base URL	API key sent as
Anthropic (default)	Messages API	https://api.anthropic.com	x-api-key header
OpenAI	Chat Completions	https://api.openai.com/v1	Authorization: Bearer header
Google Gemini	generateContent	https://generativelanguage.googleapis.com/v1beta	x-goog-api-key header
OpenAI-compatible	Chat Completions at a base URL you choose	http://localhost:11434/v1 (Ollama)	Authorization: Bearer header, only when a key is set

OpenAI-compatible covers any server that speaks the Chat Completions format, such as Ollama, LM Studio, vLLM and OpenRouter. Local servers usually need no key.

The model menu offers Default (the first model below) and a list of suggestions for the provider:

Provider	Suggested models
Anthropic	claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5, claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5
OpenAI	gpt-4.1, gpt-4.1-mini, gpt-4o, gpt-4o-mini, o4-mini, o3
Google Gemini	gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite
OpenAI-compatible	llama3.2, qwen2.5, mistral, openai/gpt-4o-mini

To use any other model, type its id in Model id (any). Whatever you type there is the model the request uses.

Typical setups:

Anthropic

Field	Value
Provider	Anthropic
API key	{{vault:anthropic_api_key}}
Base URL	Empty

OpenAI

Field	Value
Provider	OpenAI
API key	{{vault:openai_api_key}}
Base URL	Empty

Ollama

Field	Value
Provider	OpenAI-compatible
Model	The model you pulled, such as llama3.2
API key	Empty
Base URL	Empty, or http://localhost:11434/v1

Request tabs

Tab	What it holds
Prompt	The user message. {{variables}} are filled in when you send.
System	An optional system prompt, also with {{variables}}. A dot on the tab means it’s set.
MCP Tools	MCP servers whose tools the model may call (see below). The badge counts the selected servers.
Settings	The key, base URL and generation settings.

Settings

Setting	Default	What it does
API key	{{vault:anthropic_api_key}}	The provider key. Type a reference such as {{vault:openai_api_key}} rather than the key itself. If the reference can’t be resolved, the request is sent without a key and the provider answers with an authentication error.
Base URL (optional; Ollama: http://localhost:11434/v1)	Empty	Overrides the provider’s default base URL. For OpenAI-compatible servers, /chat/completions is added to the base URL; a full URL that already ends in /chat/completions (optionally with a query string) is used as is. {{variables}} work.
Temperature	1.0	Randomness, from 0 to 2.
Max tokens	1024	The longest answer allowed, from 16 to 64,000 in steps of 256.
JSON output	Off	Asks for a single JSON object: OpenAI-style JSON mode for OpenAI and compatible servers, a JSON response type for Gemini, and an instruction added to the system prompt for Anthropic.

Conversations and streaming

In the editor, answers stream token by token into Response, with You and Assistant labels. Each Send adds your current prompt to the same conversation, so you can ask follow-up questions; click New Conversation to start over. Stop cancels an answer in progress.

Next to the Response heading, Cloak Proxy shows the latency, the time to the first token, and the input and output token counts the provider reported.

Without MCP tools, the Messages log on the right shows the exact HTTP request sent to the provider as a cURL command, with the API key redacted. Errors appear there as Request failed, with the provider’s message.

Give the model MCP tools

An AI request can let the model call tools from MCP servers you’ve saved as MCP requests. Cloak Proxy connects to those servers, offers their tools to the model, runs each tool call the model makes, and sends the results back, repeating until the model gives a final answer. This is the same loop an AI assistant runs.

1. Save at least one MCP request in a collection. Requests in the MCP section count.

2. In the AI request, open MCP Tools. Every MCP request in your workspace is listed as name · command or URL.

3. Check the servers whose tools the model may use.

4. Set Max steps, from 1 to 30 (default 8). A step is one call to the model; the loop stops when the model answers without calling a tool, or after this many steps.

5. Send the prompt.

While the loop runs, the log shows how many tools are available, each tool call as server · tool with its arguments, and each result as tool · 140 ms (in red when the tool reported an error). The answer area shows the model’s text from every step.

How the loop works:

• Tools from all selected servers are offered together. When two servers have a tool with the same name, the second is renamed to server__tool. Names are made safe for the model’s API and kept to 64 characters.
• All tool calls the model makes in one step run at the same time; their results go back in the order the model asked.
• Arguments are checked against the tool’s input schema first. Invalid arguments, unknown tools and failed calls are sent back to the model as error results, so it can correct itself.
• A server that fails to list its tools is logged and skipped; the others still work.
• The MCP servers are connected when you send, and closed when the answer is complete.

The model calls tools without asking

Cloak Proxy runs every tool call the model makes, without asking you first. Only attach MCP servers whose tools are safe to run unattended, and keep Max steps low while you try a new server. An STDIO server is a program running on your Mac with your permissions.

Run AI requests in the collection runner

The collection runner and cloakctl send a saved AI request as a single, non-streaming call with the Prompt as the only user message. With MCP servers selected, they run the tool loop to the end. The run waits at least 60 seconds for the model to start answering, or the request timeout if that’s longer.

Response part	Value
Status	200 OK when the model answered. Provider errors, such as a bad key or an unknown model, are reported as an error with no response.
Content-Type	text/plain; charset=utf-8, or application/json with JSON output on.
x-ai-provider	openAI, anthropic, gemini or openAICompatible.
x-ai-model	The model that answered.
x-ai-input-tokens, x-ai-output-tokens	Token usage, summed over every step of a tool loop, when the provider reports it.
x-ai-stop-reason	The provider’s stop reason, or for a tool loop finalAnswer or maxStepsReached.
Body	The model’s final text. With JSON output on, a surrounding Markdown code fence is removed so pm.response.json() works.

AI requests have no Scripts tab of their own. Put tests in the collection’s or a folder’s post-response script, and check pm.info.requestName for tests meant for one request:

Collection post-response script

if (pm.info.requestName === "Classify ticket") {
  pm.test("model returned a known category", () => {
    const answer = pm.response.json();
    pm.expect(["bug", "question", "feature"]).to.include(answer.category);
  });
  pm.test("stayed within budget", () => {
    pm.expect(Number(pm.response.headers.get("x-ai-output-tokens"))).to.be.below(300);
  });
}

Model output varies from run to run, even with a low temperature. Test for structure and ranges rather than exact text.

Keep API keys in the vault

The API key field is saved with the request in your workspace files, so a key typed there ends up in Git and in exports. Store keys in the vault, which keeps them in your macOS Keychain, and reference them as {{vault:name}}. Keys are used only for the call and are redacted in the log. AI requests go straight to the provider, not through Cloak Proxy, and calls to paid providers are billed to your account.

Related

MCP clientConnect to the MCP servers whose tools you give the model.

Variables and environmentsStore API keys in the vault.

Agent modeLet an assistant work in the API client for you.

Collection runnerRun AI requests with tests.