What's new in Htvag Agent
Every release, in plain language — what changed and what it means for the machine you run it on. The current release is v0.9.35 (2026-10-04).
v0.9.35
Long runs are no longer cut off by a clock
- Scheduled and headless runs have no wall-clock limit. A PC-control job that needs an hour now gets the hour — the ten-minute ceiling that stopped long runs mid-task is gone, instead of the agent appearing to give up for no reason.
- A local model's turn has no request deadline either, so a long task on a CPU-only machine no longer dies mid-stream at the 45-minute mark.
- Nothing new to wait out: the run's own turn limits, the Stop button and the kill switch still end a run the moment you ask, and a finished run never blocks the next one.
v0.9.34
Click coordinates land where the screenshot says, on every monitor layout
- Clicks picked off a screenshot now account for where the capture started. On desks where a second monitor sits to the left of or above the main screen, the cursor lands on the target instead of being offset by that monitor's size.
- A click coordinate that falls outside the screenshot the model read is refused with a prompt to take a fresh one — instead of the cursor being snapped to a screen edge the model never chose.
v0.9.33
Scheduled runs never pause on a token budget
- The per-day token cap on scheduled runs is gone, along with its setting. A schedule that is due simply runs.
- Runs still respect quiet hours, one at a time, and never start in a folder a session is already working in.
v0.9.32
Window focus works again, and PC runs read their full playbook
- Window commands are baked into their script instead of passed as flags PowerShell mis-parses, so listing and focusing windows works reliably again.
- The PC-control lane reads its full doctrine: a launch is a step and not the task, focus is confirmed before a click, and a message counts as sent only when a screenshot shows it sent.
- PC workflows are pickable — WhatsApp reply, WhatsApp send, browser task, files, YouTube — and only the chosen playbooks enter the prompt.
v0.9.31
A quick PC command gets a short turn
- A small PC-control request runs as a short turn instead of a full planning cycle, so simple jobs finish sooner.
- The PC lane carries one less tool to choose between.
v0.9.30
PC control gets its own prompt; local models get a prompt they can hold
- PC-control runs use a prompt and tool set built for driving a desktop, rather than the coding agent's.
- The same lane runs on a prompt small enough for a local model's context window, so a machine without credits can still drive its own desktop.
v0.9.29
Local runs continue into action, and a stale gateway token heals itself
- A local model's reasoning now continues into tool calls instead of stopping at the thought.
- An expired gateway token is refreshed in place instead of failing the turn.
Prefer the complete feature list rather than the delta? It lives on the features page.
Run the current release
Download the installer for Windows or Linux, or start with the CLI. The free plan is the complete agent.