Hycli turns any website into typed CLI and MCP tools for agents
Hycli is a Rust binary that inspects a website's docs, frontend code and network traffic to generate a typed CLI and an MCP server, so an AI agent can read and write on that site using the user's own session without the model ever seeing a password.
Image: GitHub
Why it mattersLetting an agent act on real logged-in websites is the step most useful work needs, and the thing most agent tools skip because credentials are hard to hand over without leaking.
An AI agent that can browse is one thing. An AI agent that can log in to your account and do real work on a site is harder, because the credential has to live somewhere the model never sees. Hycli, an open-source project under the Apache-2.0 licence from a developer who goes by Hybirdss, takes the second step. It is a single Rust binary that inspects a website's documentation, frontend code and network traffic, then generates a typed command-line interface for that site plus an MCP server an agent can call. The repository was created on 9 October 2026 and has reached 416 stars in the thirteen days since.
The one-sentence pitch on the README is "Turn any website into a typed CLI and MCP server for your AI agent." The install is a single binary with no Rust or Node runtime, with native installers for Windows 10 and 11, macOS 13 and later, and Linux x64 and ARM64.
What it does in practice
Hycli builds the CLI in three worker passes. The first worker reads whatever API documentation the site publishes. The second reads the site's frontend code to find the data shapes behind each page. The third watches real network calls as a user clicks around, which fills in anything the first two missed. The output is a set of CLI subcommands grouped by resource, each one with its arguments typed and its description generated from the site itself.
Credentials stay on the user's own machine, in the operating system keyring. The agent calls the CLI or the MCP server, which signs the request using those credentials, and the model sees only the response. Reads happen immediately against an authenticated site. Writes, which Hycli defines as any create, edit or delete, need an explicit per-site confirmation the first time they are used, and no automatic retry on failure.
The limits as the author states them
Four limits sit in the README and are worth reading before adoption. The first is that native browser session discovery, where Hycli finds and reuses an existing logged-in session, is opt-in and defaults to off on macOS. The second is that some browser profiles that are encrypted or sit in private containers need a one-time manual sign-in through the companion extension. The third is that there is no automatic write retry, which the author gives as a design choice to avoid accidental duplicate records. The fourth is that all testing to date has run against synthetic local websites, so a real site's behaviour under load, rate limits, or anti-bot measures is still unmeasured.
The project is version 0.1.1 and under active work, which the version number says out loud. The traction is the 416 stars in thirteen days and 101 forks, both plausible on an agent-tooling project that offers to connect real sites to Claude Code, Codex or any other MCP-aware agent without handing over passwords.
A short note on scope. Hycli is the sort of tool a developer, product manager or ops person could put between an agent and a work tool they already use, which is why it lands here rather than in a repository list for infrastructure builders. The reader who clicks through gets a binary that writes its own typed interface for one website at a time.
Source
Hybirdss/Hycli on GitHub.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.