Start a coding agent on HIPS

00000000000000000000000000000000000000001hi <agent> [hi's options] [--] [the agent's own arguments]

hi <agent> starts that coding agent on HIPS for one session: it finds the agent on this computer, checks its version, makes sure you are signed in, chooses the model, and starts the agent with HIPS's address, your sign-in and the model in that session's configuration. The agent then keeps its own commands, terminal, signals and exit code. hi never changes the agent's own settings, never puts a credential in its environment, files or arguments, and never reads another vendor's stored sign-in.

Each agent talks to HIPS in its own protocol, which the gateway serves: Anthropic's Messages API for Claude Code, the Responses API for Codex, chat completions for the rest.

After the agent's name, hi reads only its own options. The first argument that is not one of them, or a literal --, ends them, and everything after goes to the agent unchanged: hi claude --model <id> -p "..." sends -p "..." to Claude Code. An option the agent shares by name, such as Claude Code's own --model, is reached after --.

An unknown model id is refused before the agent starts, with the nearest listed ids; hi never corrects or substitutes one, and a model without tool calls is refused. Status and models has hi models.

HIPS tests each new release of every agent automatically, on each of Linux, macOS and Windows where the agent's vendor offers it, and publishes the versions that pass, per system, as a signed list, which hi reads at most once a day; a list without HIPS's signature is never used. A newer release than the list names is refused until it is tested, so a session never runs on a version HIPS has not seen. The agent's own auth, update and similar commands run as they are, with no sign-in.

For a session on HIPS, hi turns off the agent's own telemetry and its nonessential traffic to its vendor (for example CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC for Claude Code), and says once at launch what it turned off. --keep-agent-telemetry leaves it as you set it. Your sessions of the agent outside hi are untouched.

HIPS keeps no prompt or answer; your agent keeps its own history on your computer, as it does without HIPS.

Requests are billed to the workspace you chose at sign-in and appear in the Console as that sign-in's usage. The workspace's budget covers them: past it, the agent shows the gateway's 402 budget_exceeded as it shows any error. If the same model fails on one server, the gateway tries another before the first byte; hi never switches models for you.