Skip to content
Docs menu

Where a daemon chat's model runs

Run a code or computer chat's model on the daemon's own keys or on your organisation's cloud keys, see which one each gear uses, and set or lock the default for your organisation.

View .md

A code chat or a computer chat runs in a session on one of your daemons. Its tools (files, commands, the browser) always run there. Its model can run in one of two places:

  • This machine's keys: the daemon's own orchestrator and whatever it serves (local models through Ollama or vLLM, and the provider keys set on that machine).
  • Cloud keys: your organisation's provider keys in Venom Cloud (the Provider keys page), through the cloud's /v1, the same way a quick chat runs.

Before, a daemon chat could only use models the daemon served, and a gear that needed a key the daemon didn't have failed with "No API key found … Use /login". Now the picker says where each gear would run, and you choose.

The choice

In the new-chat box, next to Where the agent works and the gear, Model runs on has three values:

Value What happens
Auto (the default) The daemon's own model when the daemon serves the gear's model, else the cloud's keys.
This machine's keys Only the daemon's own models and keys. Nothing the chat does goes to the cloud's providers. A gear the daemon can't run is unavailable.
Cloud keys Always the cloud's keys.

Your choice is remembered for your next chat. An admin can set the organisation's default and lock it (below); then the picker shows the organisation's value and you can't change it.

What the gear picker shows

For a daemon chat each gear says its model and where it runs:

  • ollama/qwen3:8b · this machine
  • claude-opus-5-5 · cloud keys

A gear that can't run is listed unavailable with the reason, for example:

  • "JC computer serves no model for qwen; the cloud: there is no provider key for its model qwen (Qwen): an admin adds one on the Provider keys page"
  • "JC computer serves no model for anthropic, and this chat runs only on the machine's keys"
  • "your organisation keeps its models on its own machines"
  • "update the daemon on JC computer to use this gear here"

Default (no gear) is the daemon's own default under Auto and This machine's keys; under Cloud keys it is the organisation's default gear on the cloud's keys, else the cloud's venom-default.

A running chat shows where its model runs next to where the agent works. A chat's model stays where it started: shifting to another gear only offers gears that run in the same place. To move a chat to the other side, start a new chat.

What goes through the cloud

With cloud keys, the chat's prompts, the files the agent reads, screenshots and tool results are sent through Venom Cloud to the model's provider, as a quick chat's are. The daemon never receives the organisation's provider keys: the session gets the chat's own key (the same personal key of yours its Venom MCP servers already use, revoked when the chat is archived or ends, or when you leave the organisation). The requests are pinned to the cloud's providers (x-venom-daemon: cloud), so the cloud never relays them back to a daemon.

Spend is yours, as for a quick chat: it is booked to you under the chat's session, counts toward your limits and budgets, the session guard and the organisation's model allow and deny lists.

With This machine's keys nothing new leaves the daemon: the chat works as before.

For admins: the default and the lock

On the Gears page, Where a daemon chat's model runs sets the organisation's default (Auto, This machine's keys only, Cloud keys always) and Members can't change it per chat. The same is daemon_inference and daemon_inference_locked on PUT /api/admin/v1/gears and the admin_gear_set MCP tool. Changes are audited as gears.updated.

Locking This machine's keys applies at once: chats and computer runs already running on cloud keys end (their key is revoked, so their next model call stops), and each chat says why. Start a new chat to carry on on the machine.

A sovereign organisation (it uses only its own daemons) is always This machine's keys, locked, whatever this says.

A daemon whose organisation policy pins its orchestrator (orchestrator.baseUrl in policy.json) refuses cloud keys for its sessions too.

Bots and routines on a computer

A bot's or a routine's run on a computer follows the organisation's setting (there is no member choice for an unattended run). Under Auto or This machine's keys a run with no gear uses the daemon's default as before; a bot's gear runs where Auto places it, and falls back to the daemon's default when it can run nowhere. Under Cloud keys a run uses the bot's gear, else the organisation's default gear, on the cloud's keys.

Spend and the daily cap

An agent's or bot's run on a computer has a daily cap. A run whose model runs on the machine's keys counts what the daemon reports, as before. A run on cloud keys counts what the cloud itself booked for it (its requests under the run's own key), and the daemon's own estimate for that run is left out, so nothing counts twice. A day with both kinds adds each from its own source. A cloud request for a model nothing prices is booked at $0; for the cap it counts at the dearest model your organisation serves, as the Evals estimate counts an unpriced model. The AI spend page shows the same figure the cap checks.

Older daemons

Cloud keys need a daemon released with this change or later (its venom serve says cloudInference: true in GET /v1/gears). On an older daemon the gears it runs itself keep working, and a gear that needs cloud keys says to update the daemon.

For agents

agents_gear_list with daemon_id returns each gear's runs_on (daemon or cloud), why and reason, and inference (the setting in force, the org's default, whether it is locked, and why cloud keys can't be used there). Pass inference (auto, machine, cloud) to agents_gear_list and agents_chat_create. A chat's inference says where its model runs. REST: GET /api/workspace/agents/gears?daemon=<id>&inference=…, POST /api/workspace/agents/chats with inference.