Privacy & security
Kurlo is local-first. What stays on your machine, what leaves it, and how your data is protected.
Kurlo is a local-first desktop application. There is no Kurlo backend, account system, cloud sync, usage analytics, crash reporting, or telemetry. Network access occurs only for product features that require it, such as requests, Git operations, OAuth flows, schema introspection, and update checks.
This page explains what the Kurlo app stores, when it connects to the network, and how kurlo.dev measures visits.
None of it has to be taken on trust: Kurlo is open source under the MIT license, so every claim below can be checked against the code at stormhop/kurlo, and the binaries can be reproduced from that source.

What Kurlo stores on your machine
Section titled “What Kurlo stores on your machine”Workspace state, requests, environments, request history, cookies, secrets, and preferences are written to local disk. Request history records what was sent and what came back, including the response body, which is stored in its own file under history/ and encrypted with the same key as the rest of the local profile. A response over 2 MB keeps only its head, and a binary response keeps no body at all. Kurlo is a single-tenant, file-based application:
| Platform | Location |
|---|---|
| macOS | ~/Library/Application Support/Kurlo/ |
| Windows | %AppData%\Kurlo\ |
| Linux | ~/.config/Kurlo/ or $XDG_CONFIG_HOME/Kurlo/ |
Inside that directory you may find:
requests.json— an AES-256-GCM encrypted JSON envelope containing local profile state, history, workspace cookies, Git-local secret values, and storage metadata.request-store.key— a base64-encoded recovery copy of the encryption key, written with owner-only0600permissions.request-store.key.dpapi— an additional Windows DPAPI-protected key copy.preferences.json— the configured default workspace location, cookie sync settings, and AI connections with their keys encrypted by the request-store key. Other UI preferences are stored by the desktop WebView.agent-chats/— agent chats and copies of sent attachments, encrypted on disk. Kurlo keeps up to 200 unarchived chats across the local profile for up to 90 days of inactivity. Archived chats and their attachments are kept until you delete them. Markdown exports are unencrypted files saved at a location you choose.agent-activity/— the encrypted log of tool calls made by the agent and by assistants connected throughkurlo mcp, with the files they changed, kept for 30 days.workspaces/— the default folder-backed YAML workspace location.
If you select another workspace folder or Git repository, its shared YAML files live at that selected path. Those files are intentionally human-readable and are not encrypted as a unit.
You can move, back up, or delete this directory at any time. Deleting it returns Kurlo to a clean state on next launch.
Encryption at rest
Section titled “Encryption at rest”The request store is encrypted with AES-256-GCM:
- A fresh 256-bit key is generated locally the first time Kurlo starts.
- The key is additionally stored in the OS credential store wherever possible:
- macOS — Keychain (the same store used by Safari / login items)
- Windows — DPAPI (current user; non-exportable to another user account)
- Linux — libsecret via
secret-tool(GNOME Keyring, KWallet, etc.)
- Kurlo also keeps
request-store.keyas a recovery copy even when the credential-store write succeeds. This avoids losing encrypted data if the credential-store entry disappears. - The recovery file is owner-readable only, but it is not protected by another secret. Anyone who can read both
requests.jsonandrequest-store.keycan decrypt the local profile. - Every encrypted payload uses a unique random nonce. There is no nonce reuse across writes.
For folder/Git workspaces, Kurlo replaces sensitive YAML values with {{kurloSecret:...}} placeholders and stores the real values inside the encrypted local profile.
What leaves your machine
Section titled “What leaves your machine”A few bounded categories of network calls happen on your behalf:
-
Requests you send. This is the whole point — when you press Send, Kurlo opens a connection to the host you typed in the URL bar. The request is built on your machine and sent directly unless you configured a system, global, or per-request proxy.
-
Auto-update checks. On launch (and when you press Check for updates), Kurlo fetches
latest.jsonfrom the public release repository (github.com/stormhop/kurlo) over HTTPS. Release builds verify the downloaded binary by SHA-256 and minisign signature. No usage data is included in the request. -
OAuth 2.0 authorization and token calls. Kurlo can open the provider’s authorization page and call the authorization/token endpoints you configured.
-
GraphQL schema introspection. When you load a GraphQL schema URL in the Schema tab, Kurlo sends a standard introspection query to that URL.
-
Git remotes. Clone, fetch, pull, push, remote tests, and remote branch actions contact the Git server you configured.
-
An OpenAPI spec you import from a URL. Import collection -> OpenAPI / Swagger from URL fetches the link you paste, through the same sender your requests use, so it honours your proxy settings and the response cap. Cookies are not attached to it.
-
AI connections. When you test an AI connection or load its models, Kurlo calls that provider’s model list (and, for OpenRouter, its key check) with your key. Connecting OpenRouter opens openrouter.ai in your browser; after you approve, OpenRouter sends the browser to
kurlo.dev/connect/openrouter/, a static page that forwards it to Kurlo on127.0.0.1, and Kurlo exchanges the code for a key atopenrouter.ai/api/v1/auth/keys. The code passes through that page’s URL but cannot be exchanged without a secret that never leaves Kurlo, and it expires after ten minutes. If kurlo.dev cannot be reached, OpenRouter returns tolocalhostdirectly. Find local models only contacts127.0.0.1ports 11434 and 1234. These calls follow the system proxy environment variables, not the proxy configured in Kurlo.
Proxy settings can route these requests through your selected system or custom proxy. There is no separate telemetry channel or first-launch ping.
Ports Kurlo opens on your machine
Section titled “Ports Kurlo opens on your machine”Four features bind a local listener. All four are loopback-only (127.0.0.1, and ::1 for the OpenRouter sign-in), so nothing on your network can reach them, and none of them is open unless you start it.
| Listener | When it exists | What it serves |
|---|---|---|
| OAuth 2.0 redirect | Only while an Authorization Code sign-in is in flight, on an ephemeral port | The single /callback that receives the authorization code, then closes |
| OpenRouter sign-in | Only while Connect OpenRouter is waiting (at most five minutes), on an ephemeral port | The single /callback that receives OpenRouter’s code, then closes |
| Cookie sync bridge | While cookie sync is on (ports 3199-3203) | Pairing and the WebSocket a paired browser pushes cookies over. Extension origins only; a web page probing the port gets a 403 |
| Mock server | While you have it running, on the port you choose | The saved examples of the collection you pointed it at — status, headers and body as recorded. Any program on your machine that can reach the port can read them, so treat a mock made from real responses accordingly. In a browser, only a page served from this machine is answered; a site on another host gets a 403 |
Scripting safety
Section titled “Scripting safety”Pre-request and test scripts run in a sandboxed JavaScript VM by default, with a legacy Tengo engine for existing requests:
- Imports,
require, process access, filesystem access, and direct network sockets are disabled. - Execution is capped at 2 seconds per script.
- Scripts have access only to the explicit
pm.*API (request/response/variables/environment/log). They cannot reach the file system, spawn processes, or open network sockets directly. Any HTTP call goes through the regular request engine.
Cookies
Section titled “Cookies”By default Kurlo maintains a cookie jar across requests, just like a browser. Cookies are stored in encrypted requests.json. You can:
- Inspect and edit cookies per domain in the Cookies modal (the icon next to the URL bar).
- Clear all cookies from that modal.
- Disable the cookie jar for an individual request in Per-request settings.
Cookies are never transmitted anywhere except in the request to the domain that set them, following standard browser rules.
Cookie sync (Cookies window -> Sync) is off unless you turn it on. When on, Kurlo opens a listener bound to 127.0.0.1 that a browser extension can find and ask to connect to; you approve the request in Kurlo by matching a six-digit code, and only then does Kurlo hand over a token. The extension then holds a WebSocket to that port and pushes cookies for the domains you allowlist there, and only those. The token and the allowlist are stored in preferences.json. Cookies travel browser -> Kurlo over loopback only, and a snapshot replaces what the jar held for those domains. Disconnect mints a new token and drops the browser immediately.
Data export and deletion
Section titled “Data export and deletion”- Export - Settings -> General -> Data -> Export all data writes workspace data, history, cookies, secrets, and selected preferences to a plaintext JSON file you choose. External Git repositories, the global proxy password, and some local UI preferences are not bundled. See Backup & recovery.
- Import — replaces the current profile with a previous export. This is destructive; Kurlo shows a confirmation dialog before running.
- Delete — quit Kurlo and remove the directory listed under What Kurlo stores on your machine.
Documentation feedback
Section titled “Documentation feedback”When you rate a documentation page, kurlo.dev sends the page path, your vote, and any reason you select to its GoatCounter instance. The feedback form has no free-text field and sends no account details.
Reporting a security issue
Section titled “Reporting a security issue”If you find a vulnerability, please follow the process in our security policy. Please do not file public GitHub issues for security problems.
Local coding agents
Section titled “Local coding agents”Connecting Codex starts the installed CLI on this computer. The CLI owns authentication and sends conversation content and tool results to its configured provider. ChatGPT sign-in uses the official Codex flow; Kurlo does not collect subscription access or refresh tokens. Kurlo saves the executable path and default model in its preferences and associates each local chat with a native Codex session ID.
Kurlo’s chat copy is encrypted in its profile. Codex separately stores native sessions and credentials according to its own settings. Deleting a Kurlo chat or removing the connection does not delete the native Codex copy or sign you out. Workspace changes still pass through Kurlo’s approval controls.
Claude Code runs through the installed CLI. Its configured provider receives prompts and tool results. Kurlo associates its encrypted chat with the native Claude Code session ID. The CLI manages its credentials and native transcript separately from Kurlo.