CLIProxyAPI is the endpoint. What do you point at it?

What CLIProxyAPI actually is

CLIProxyAPI is an MIT-licensed Go gateway. It logs into the CLI agents you already have, Gemini CLI, ChatGPT Codex, Claude Code, and a couple of others, through their normal OAuth flows, then re-serves them behind a single local HTTP server that speaks the OpenAI, Gemini, Claude, and Codex API formats. The motivation is simple: a Claude Pro or Max plan, or a ChatGPT subscription, is a flat monthly fee, while raw API access is metered per token. The proxy lets any API client borrow the flat-fee subscription as if it were a metered endpoint.

That solves the supply side. It does nothing for the demand side. Once the proxy is running you still need a client, and the examples in the wild are almost all command-line: curl, a coding CLI configured with a custom base URL, or an OpenAI SDK script. If what you wanted was a desktop chat app with real session management, you are on your own. The Anthropic SDK that Claude Code is built on reads one environment variable, ANTHROPIC_BASE_URL, and any app that sets it correctly becomes a valid CLIProxyAPI client. Fazm exposes that variable as a single text field.

Where the request goes

With the endpoint set, a Claude request you type (or speak) into Fazm travels through the proxy before it ever reaches a model provider. Your subscription credentials live only in CLIProxyAPI; Fazm never holds them.

Connecting Fazm to your proxy

Four steps. The only one people miss is the last one, and it is the reason an endpoint can look connected while receiving no traffic.

What that single field actually does

This is the part you will not find on any CLIProxyAPI page, because it lives in Fazm’s source rather than the proxy’s. When you set a valid endpoint, Fazm rewrites the environment of the agent process it spawns. The relevant code is makeBridgeEnvironment in Desktop/Sources/Chat/ACPBridge.swift (around line 2425).

Three deliberate behaviors are worth calling out, because each one is a bug you would otherwise hit by hand:

• The value is validated, not trusted. validCustomAPIEndpoint (ACPBridge.swift line 301) requires a real http or https URL with a host. A bare localhost:8317 with no scheme is rejected. This matters: a malformed value landing in ANTHROPIC_BASE_URL would make the Anthropic SDK throw “Invalid URL” on every turn and silently brick chat. On a bad value, Fazm logs it and falls back to the default endpoint instead.
• The bundled key is removed and a placeholder is substituted. Fazm sets ANTHROPIC_API_KEY to sk-fazm-custom-endpoint. The reason is subtle: with no key at all, the SDK would try a Claude OAuth handshake against your proxy; with a real key, you would be leaking Fazm’s credential to a third-party server. A harmless placeholder keeps an Anthropic-compatible gateway on the simple API-key path. If your CLIProxyAPI config enforces an api-keys allowlist, add that exact string to it.
• The override is Claude-only, by design. It rewrites ANTHROPIC_BASE_URL and nothing else, so Gemini and Codex models route through their own backends untouched. The settings card watches your selected model and renders a warning the moment it is not a Claude model, with a button to switch (SettingsPage.swift around line 1086). That warning exists specifically because new installs default to Gemini, and without it the endpoint would look configured while receiving zero requests.

What the GUI adds that the endpoint cannot

A proxied endpoint answers “where do the tokens come from.” It says nothing about how you work day to day. That is the half Fazm fills, and it is the same agent loop (real Claude Code, wrapped via ACP) with the rough edges fixed:

• Sessions survive a restart. Quit, reboot, reopen, and every window comes back with its full conversation intact. The proxy is stateless; persistence is on the client side.
• Forking is one click. Any chat opens into a new window carrying the full prior context, and the original is left untouched, so you can branch an investigation without a session-id dance.
• No auto-compacting. The full chat history stays live in context for the lifetime of the window. Long sessions through your proxy do not quietly drop earlier decisions.
• Voice-first input. Hold a hotkey and talk; the same agent runs with no typing.
• Reach beyond the terminal. Through macOS accessibility APIs the agent can drive your real browser and native Mac apps, not just files in a working directory.

Fazm is open source and runs locally, which is the right pairing for a setup whose whole premise is keeping your credentials and traffic on your own machine.

Questions people actually ask