1.1 Introduction

Casbin Gateway is a local gateway for the AI coding agents installed on your machine. It finds them, shows what they carry, puts one endpoint in front of the model vendors they talk to, and keeps the record of what they did and what it cost.

It runs on the machine whose agents you want to watch. Nothing is sent anywhere: the database is a file next to the executable, and the UI binds to 127.0.0.1 unless you say otherwise.

What it is forCopy Anchor Link

How it fits togetherCopy Anchor Link

   agents on this machine           Casbin Gateway               model vendors
+------------------------+     +----------------------+     +------------------+
| Claude Code            |     | permissions (Casbin) |     | OpenAI           |
| Codex CLI              |---->| translation          |---->| Anthropic        |
| Gemini CLI, Cursor ... |     | records + usage      |     | DeepSeek, Kimi   |
+------------------------+     | provider keys        |     +------------------+
            ^                  +----------------------+
            |  reads their config, transcripts, skills and MCP servers
            +----------------- same machine -----------------

Two independent halves, and either is useful without the other:

• The proxy. An agent pointed at Gateway has its requests authenticated, checked against its permissions, translated into whatever API the provider speaks, relayed, and recorded. Gateway holds the vendor key, so the agent never has it.
• The reader. Agents are discovered by reading the home directories, install paths and transcripts of this machine. That works from the first start, with no provider configured and nothing routed — which is why Usage can account for requests that never went through Gateway at all.

Because the reader half works on files this machine owns, Gateway has to run on that machine. In a container it sees the container's own filesystem and finds nothing — see 5.4 Docker and Podman.

What each agent getsCopy Anchor Link

Around thirty agents are recognised, and how far Gateway goes with each one differs. The table in the README is the current list, column by column:

• Monitoring — audit-only records of prompts, tool calls and permission prompts. Nothing an agent does waits on Gateway, and no answer changes because monitoring is on.
• Provider — Gateway writes the agent's own configuration file to point it at a bound provider, in the format that agent's client speaks. An agent without this still reaches Gateway through environment variables.
• MCP · Skills · Prompt — read, compare and copy MCP servers, skills and instruction files between agents.
• Sessions — conversations and token usage read from the agent's own transcripts.
• Install — how Gateway installs, upgrades and removes it: npm, brew, winget, the Microsoft Store, the agent's own updater, or the vendor's own install script.

Built withCopy Anchor Link

Next: 1.2 Installation.