hyai changelog

[0.6.0] - 2026-10-07

hyai works with a router of your own, not only with hostyourai.com. hyai login --router https://your-router.example (or -r) checks your key on that router and stores the address next to it. You can paste the root, the /v1 address or the /ui page of the router; hyai finds the API.
hyai login recognises a key for your own router (it starts with hyr-). When no router address is known yet it asks for one, or, in a script, stops with the command to run, instead of sending the key to hostyourai.com and failing with "not accepted".
hyai doctor says which router you use (HostYourAI or your own router, and its address) and fails with the fix when the key and the router do not belong together.

When a router does not accept a key at login, hyai says which router it asked and, when the key looks like it is meant for the other kind of router, how to log in there.
A key is never sent to the other kind of router. When a hyr- key meets the hostyourai.com address, or a hyai-rt- key meets the address of your own router, hyai login stops before any request and nothing is stored, and hyai doctor reports a FAIL and skips the router check.
HYAI_BASE_URL still wins over the stored address, and nothing changes for existing hostyourai.com users.

Every link to the API keys page pointed at https://hostyourai.com/api-keys, which does not exist. The install script, hyai login, the key validation error, the router 401 error and the README now all point at https://hostyourai.com/app/api-keys.

Images work with Autopilot. Paste or attach a screenshot and Autopilot sends it to a model that can see images. Before, OpenCode quietly dropped the image, because hyai did not tell it that Autopilot accepts images.
Models that see images accept them when you pick them by name too. hyai reads this from the router's model list (supports_images).
The model picker shows how fast each model writes, measured over the last three days, for example "Kimi K3 · 40 tok/s".
hyai warns when you start on a slow model and names the fastest one.
hyai doctor checks whether a newer version of hyai is available.

Long answers from thinking models were cut off. A model such as Kimi K3 thinks several thousand tokens before it writes, so writing a file could end halfway and the agent said "I'll build it now" without doing anything. The output budget now follows the highest value the router measured for each model (at most a quarter of the context), instead of a fixed 16,384.

Models on your own provider key show up in hyai models and in every agent's model picker, labelled as yours. hyai never picks one for you.

The model pickers of Claude Code and Codex list the router's models. The settings live in ~/.cache/hyai; nothing is written to ~/.claude or ~/.codex.

hyai run claude starts Claude Code on the router. Address, key, model and context window go through the environment only.
A model on your own dedicated instance is usable and chosen before the shared list.

First release: hyai login, hyai code, hyai models, hyai doctor.
Reads the live model list, keeps your key in the system keychain and only starts on models that are warm, billed per token and able to call tools. It never picks a cold model, because that can start a rented GPU.
OpenCode by default, hyai run codex for Codex.