Contexts are the reason to run a gateway instead of pointing every editor at every server: one application gets the documentation plugin, another gets something else, and neither carries the other's tool definitions in its context. Plugins declare their network and file permissions in the manifest, so what one can reach is readable before you install it.
A cross-platform Rust binary that acts as an MCP server for your AI applications and gets its capability from plugins. Each plugin declares its tools, resources and prompts in a TOML manifest and asks only for the permissions it needs.
- One stdio endpoint that Claude Desktop, Claude Code, Cursor, VS Code, Zed, Codex CLI and Opencode all connect to
- Plugins installed, created, configured and inspected — jilebi plugins add, create, setup and log
- Application contexts that limit which tools each application sees, so one editor gets a documentation plugin and another does not
- Plugin manifests in TOML declaring tools with input schemas, read-only resources with MIME types, and prompt templates with interpolated arguments
- A permission model per plugin: allowed hosts, specific URLs, readable and writable files, readable and writable directories, with user_defined prompting at setup time
- Environment variables and secrets declared separately, with secrets allowed no defaults
- Persistent state for plugins across runs
Windows, Linux or macOS. Install with the script from mcp.jilebi.ai, or download the platform archive. On macOS the binary needs its quarantine attribute removed and the xz dependency installed. Then add jilebi with the stdio argument to your client configuration.
One command — curl -fsSL https://mcp.jilebi.ai/install.sh | bash