Gateway
Chat with your accounts' models through a Gateway such as CLIProxyAPI, on the app's own agent.
A Gateway is a server you run, such as CLIProxyAPI, that offers the models of your accounts behind one Anthropic- and OpenAI-compatible address and key. The app knows at most one Gateway and offers it as a harness of its own: Gateway chats run on the app's own agent, so they work with no harness installed.
The app does not start, bundle or manage the Gateway. Start it with its own tools.
Connect it
Start the Gateway
Run your Gateway and sign in to your accounts there, with its own tools.
Add the address
Open Settings › Gateway and enter the address under Connection, for example http://127.0.0.1:8317. A loopback address may use plain HTTP; any other address must use HTTPS.
Save a key
Enter a key the Gateway accepts and save it. The key is kept in the system keyring; without one, as under WSL, it goes in the config folder, readable only by you.
Pick the models
Once the page says Connected, choose the Gateway models the chat's model picker offers. If the Gateway does not answer, the page says so and offers Check again.
The app's own agent
A Gateway chat runs an agent loop inside the app. The model's reply streams in, the agent runs the tools it asked for, hands their output back and continues. Anything you send in the middle is added at the next round, and older turns are summarised when the context is full. For each model, the agent uses the API the model's provider speaks natively (Anthropic Messages, OpenAI Responses or Chat Completions), based on the owner listed for it by the Gateway's /v1/models.
Its tools cover files, search, shell, web fetch and web search, plans, checklists, subagents and the tools of your MCP servers. Every call is judged against the chat's permission mode:
- known read-only commands are allowed;
- dangerous commands are always asked;
- edits outside the group's folder are never remembered.
For web search, the agent asks the chat's Gateway model in a request of its own, where the provider's built-in search is the only tool, and passes on what it found together with the sources. So every search adds one model request to your accounts' usage.
MCP servers
Gateway chats use MCP servers from Settings › MCP and from a project's .mcp.json. The app asks once per folder whether to trust or ignore a .mcp.json; Settings › MCP lists those choices under Project servers, where Forget asks again. Servers that are on start with every Gateway chat, and each of their tool calls goes through the chat's permission mode like a shell command. Secret values and headers are kept in the system keyring.
The engine speaks MCP over stdio and streamable HTTP, and implements tools only: not resources, prompts or sampling.
Limits
- Gateway chats are chats only; they have no terminal view.
- A Gateway chat cannot be started as a Worker.
- The agent's commands are not sandboxed by the operating system yet. What protects you is the permission mode and the app's check of each command.
- Translation between providers is the Gateway's job. If a model rejects the history, the error shows in the chat.
Each Gateway chat keeps one session file in the gateway-sessions folder of the config folder. See Chat pipeline for how the engine fits beside the other harnesses.