Venom Computers on a daemon
Turn on persistent agent browsers on your own daemon, and see how many seats it holds, how many run at once, what happens beyond that, how idle seats hibernate, and what the daemon tells the cloud.
A Venom computer is a browser an agent owns: a Chromium profile on disk (cookies, logins, history) that survives between runs. It lives on your own daemon, not on a third-party sandbox. The daemon runs the browsers in host containers (the venom-computer image), up to 25 seats each, and starts, queues and hibernates them for you.
Turn it on
In the cloud console, open the Computers app (or Daemons → the daemon) and click Enable Computers on this daemon. On the machine itself, Local models → Computers has the same switch. The daemon checks Docker, disk and memory, gets the venom-computer image, saves the setting and starts, with no restart. The steps and what each error means are in Enable Computers on a daemon.
The requirements, sizing, image and Kubernetes notes for a fresh machine are in Run Computers on your own daemon.
Advanced: the environment
make computer-image # builds venom-computer:dev
export VENOM_COMPUTERS=onVENOM_COMPUTERS=on keeps working and wins over the saved setting. It needs Docker on the PATH and an orchestrator that runs on the host (not inside a container). Without either, it stays off and logs why. Seat profiles live under $VENOM_DATA_DIR/computers/hosts/, the seat table in $VENOM_DATA_DIR/computers.db. All settings are environment variables.
On Linux the daemon reaches a host at its address on the venom-computers Docker network. Docker Desktop on macOS does not route to container addresses yet; run the daemon on Linux for now.
How many run at once
You can hold 100 seats (VENOM_COMPUTERS_MAX_SEATS) but only so many browsers run at once:
max_active = min(VENOM_COMPUTERS_MAX_ACTIVE, floor(RAM budget / memory per active seat))The RAM budget is VENOM_COMPUTERS_MEMORY (half the machine by default). The memory per seat starts at 450 MiB and follows what the hosts measure.
A start beyond max_active first hibernates the least recently used seat that is idle, has nobody watching and no tool call in flight. If every seat is busy, the start waits in a first-in-first-out queue (VENOM_COMPUTERS_QUEUE_MAX, 200) and the cloud shows "Queued (n ahead)". Nothing is dropped; past the queue limit a start is refused with 429.
A seat that nobody watches and the agent has not used for VENOM_COMPUTERS_IDLE_AFTER (10 minutes) hibernates on its own. A hibernated seat costs disk only and resumes in a second or two.
A shared seat (several agents on one browser) is admitted once whatever its member count, and counts as busy while any member has a call in flight.
What the cloud sees
The daemon adds capabilities.computers to its heartbeat (at most every 10 seconds, when it changed): max_seats, max_active, active, queued, hosts and mem_per_active_seat_mib. The cloud places seats on the daemon with the most free capacity and sends start, stop and destroy jobs. The contract is in cloud mesh §6.9.
An agent session on a seat runs in a venom serve the daemon starts for that seat. It gets the seat's address and token in its environment only, and needs the browser tools enabled. The desktop daemon runs the venom it ships (bin/venom in its install); otherwise the orchestrator runs venom on PATH, or the binary VENOM_COMPUTERS_SERVE_BIN names.
Metrics
GET /metrics adds venom_computers_seats{state}, venom_computers_queue_length and venom_computers_rss_bytes.