kurloStarDownloadAgent mode · New in Kurlo 3.0
Ask in plain words. The agent runs your requests, finds out why one failed and writes the missing test. Bring any model, Claude Code or Codex, and approve every change.
Approved 0
LiveNot a screenshot
Pick who answers
One chat list for all three. Switch per chat.
Cancel order keeps failing. Fix the test if the API is right.
DELETEOrder 41 has already shipped, so the API answers 409 Conflict. That’s by design; the test expects 204. I’ll test for the conflict instead.
save_requestrequests/Cancel-order.yml
testScriptJs: | - pm.test("status is 204", () => pm.response.to.have.status(204)) + pm.test("refuses shipped orders", () => { + pm.response.to.have.status(409) + pm.expect(pm.response.json().error).to.eql("order_shipped") + })
save_requestDone. Cancel order now checks for 409 and passes.
Rejected. Nothing was written.
Ask to run, explain or change something
Three agents, one chat list
Pick who answers in the message box. Kurlo has no credits of its own, so you use the model or the plan you already have.
/usage, which uses no tokensWhat to ask
Chips in the message box pick the collection and environment, and Templates fill in the common jobs. The rest is a sentence.
Why did my last request fail?
Reads the history entry, compares it with the saved request and the environment, and names the cause.
Run the Orders folder and explain what failed.
Runs the collection with its auth, scripts and tests, then goes through each failed test.
Write tests for Create order.
Sends it, reads the response and proposes pm.test checks for status, shape and values.
Cover POST /orders/{id}/refund.
Adds a request to the collection with a body, variables and tests, ready to apply.
Explain the Login request.
Walks through its auth, headers, scripts and which variables it sets for the next call.
Point Staging at the new host.
Proposes the variable change in the environment file, so you see the exact line first.
Approvals
The same rules for all three agents. Reading is free; anything that changes a server or your files waits for you.
| The agent wants to | Details | What happens |
|---|---|---|
| Read requests, environments, history | Listing, reading and diagnosing a failure. | At once |
| Send GET, HEAD, OPTIONS | Runs, collection runs and one-off calls that only read. | At once |
| Send POST, PUT, PATCH, DELETE | The card shows the call before it goes out. | Asks first |
| Change workspace files | The exact YAML lines, file by file: Apply, Reject or Allow for this chat. | Asks first |
| A file you edited meanwhile | The agent’s change is dropped and it is told to try again. | Yours wins |
| Secrets | Reach your API. The model gets [secret] and never writes a value into YAML. | Masked |
| Protected environments | Listed, but no agent can send with them. You run those yourself. | Refused |
Privacy
No. 01
Kept in your local profile, never in the workspace, so they never reach Git. 90 days after the last message.
No. 02
Every tool call, by the agent or by an assistant on kurlo mcp, with its arguments, result and the files before and after. Encrypted, 30 days.
No. 03
There is no Kurlo server between you and the model. Requests go to your provider or to the CLI on this computer.
No. 04
Stop ends the turn even before the first reply. Unanswered approvals end it after 30 minutes, and a turn ends after 50 steps.
The other way round
Agent mode works on your requests, not your code. When the fix is in the server, give the coding assistant you already use the same tools with kurlo mcp: it runs the collection, changes the code and runs it again.
FAQ
The Kurlo agent with any model you connect, and the Claude Code and Codex CLIs installed on this computer. All three share one chat list, and each chat keeps the agent it started with. Read more
Nothing from us. The Kurlo agent runs free on a local model or on your own key, and Claude Code and Codex use the plan you already pay for.
No. The Kurlo agent works with OpenRouter, an Anthropic, OpenAI or compatible key, or a local Ollama or LM Studio model. Read more
Not without your Apply. Every change is shown line by line first, nothing is written while the card waits, and the files are YAML in Git, so any change can be reverted.
No. Secrets reach your API, but the model gets [secret], and protected environments are off limits to every agent.
Same tools, other side. Agent mode runs the agent inside Kurlo with approval cards. kurlo mcp hands the tools to the assistant in your terminal or editor, which can also read and fix your code. Read more
Agent mode · Kurlo 3.0 · Free
Your requests, your model, your Apply. Nothing changes until you say so.
macOS 12+ · Windows 10+ · Ubuntu 22.04+ · x64 and Arm64