Hybirdss/Hycli

Hycli – Turn any website into a typed CLI for AI agents

What it is

• A native desktop tool (single binary) that inspects a website you own or have logged into, then automatically generates a typed command‑line interface (CLI) and an MCP (Multi‑Channel Protocol) server exposing the site’s read and write capabilities to an AI assistant.
• The generated CLI lets you run actions such as search, create drafts, edit, or delete content directly from the terminal, while the MCP server lets coding agents (Claude, Cursor, Codex, etc.) call the same actions via JSON messages.

How it works

1. Preparation – hycli prepare <url> --intent "…" launches three internal “AI workers” that analyse the site’s documentation, OpenAPI/GraphQL specs, static JS bundles, and live network traffic captured from a real browser session. They infer the available endpoints, request/response schemas and produce a SiteSpec describing typed commands.
2. Zero‑mutation safety – During preparation no write request is ever sent. Write actions are discovered but stay disabled (writes: false). The first time an AI wants to perform a write, Hycli asks you once per site to enable writes (hycli writes <site> allow).
3. Local vault – Browser cookies, tokens, and other secrets are imported into an encrypted local vault (OS keyring) and never appear in LLM prompts. When a command runs, Hycli injects the secret locally and scrubs it from logs.
4. Execution – Commands can be run directly (hycli <site> search-posts --query "…") or via the MCP server (hycli mcp). Reads execute immediately; writes run only after you have granted permission.
5. Resilience – If a request hits a bot‑challenge (e.g., Cloudflare) Hycli switches to a stealth Chromium engine or in‑page fetch with the imported session, and it rate‑limits traffic to avoid bans.

Key features

• Typed CLI: every endpoint becomes a sub‑command with auto‑generated --help showing required parameters and JSON/YAML schemas.
• Multi‑language UI: dashboard and generated docs are available in 20 languages.
• Cross‑platform: native installers for Windows, macOS, Linux (no Rust/Node runtime needed for end users).
• AI‑first integration: provides an MCP server that coding agents can call, and a simple JSON connection recipe for agents like Claude, Cursor, or Codex.
• Safety controls: explicit one‑time write enable, no automatic retries, strict origin‑bound cookies, and clear separation of preparation (no writes) vs. execution.
• Local dashboard: http://127.0.0.1:4318 shows discovered actions, results, and lets you manage accounts and AI connections.

Typical workflow (30‑second quickstart)

# 1. Discover a site and what you want to do
hycli prepare https://example.com --intent "Search posts, save drafts, publish"

# 2. Inspect the generated commands
hycli describe example.com

# 3. Run a read immediately
hycli example.com search-posts --query "release notes"

# 4. When you need to create/edit, enable writes once
hycli writes example.com allow
hycli example.com create-draft --title "Hello" --body "From Hycli"

AI connection example

{
  "mcpServers": {
    "hycli": { "command": "hycli", "args": ["mcp"] }
  }
}

An AI agent can then issue hycli_prepare, hycli_run, etc., to the MCP server.

Security & privacy

• Secrets stay in a local encrypted vault; only symbolic identifiers are sent to the LLM.
• No credential leakage to the model, no automatic retries, and strict rate‑limiting.
• Optional import of browser sessions via a companion extension; cookies are scoped to the target domain only.

Who might use it

• Developers building autonomous agents that need to interact with existing SaaS tools (Notion, GitHub, Slack, etc.) without writing custom API wrappers.
• Power users who want to script personal workflows on sites that only expose a web UI.
• Researchers exploring safe LLM‑driven web automation.

Getting started

• Download the appropriate installer from the v0.1.0 release page.
• Run hycli open to launch the dashboard, connect an LLM provider, import your browser session, and start preparing sites.

License – Apache‑2.0, with a clear disclaimer to use only on sites you are authorized to access.