Guide 11 of 61 · Getting started
See the AI runtimes and desktop AI apps on a machine
List the local model runtimes and desktop AI apps on a machine, see which ones no governed session used, and route local model calls through the proxy.
Built, not yet claimed preview: this guide describes a package that has not passed its release gate.
See which local model runtimes (Ollama, vLLM, LM Studio, the llama.cpp server) and desktop AI apps are on a machine, which of them no governed session used in the last 24 hours, and how to bring an agent's calls to a local runtime under the proxy. The inventory observes and blocks nothing; on a fleet, Central Lite shows the count for every machine.
1. Know what the inventory looks for
agentgarm-session looks for local model runtimes 30 seconds after it starts and then every 10 minutes. It sends a plain GET, with no credentials, only to 127.0.0.1 (or ::1) on the ports AgentGarm's catalog lists for Ollama, vLLM, LM Studio and the llama.cpp server, and identifies each by its answer, not by the port alone; it never scans ports. An LM Studio that requires a token is not listed, since its answer without one does not show it is LM Studio.
It also looks for the desktop AI apps the catalog lists: ChatGPT, Claude, Cursor, LM Studio, Ollama and Google Antigravity. On macOS it reads each app's Info.plist in /Applications and ~/Applications, with its version and code signature; on Linux it matches the app's .desktop file by name, with no version; on Windows it looks for each app's program under C:\Program Files and %LOCALAPPDATA%\Programs, with its version and signer. Nothing is run or launched, and prompts, model output and browsing data are never read.
2. List them from the command line
agentgarm status has an unmanaged AI line: how many runtimes and apps were seen in the last 24 hours that no governed session used, and how many were used by one. Add --inventory to sweep now and list each one: a runtime with its address, port, version and the program listening on it, an app with its path and code signature, and its state, UNMANAGED or governed. --json has every entry under inventory.
The sweep is refused from inside a governed task, runs once a minute at most, and may wait up to two minutes for a periodic sweep to finish first. On a root install agentgarmd names the listening program itself, for every user; on a no-root install it sees only your own.
agentgarm statusagentgarm status --inventoryagentgarm status --json | jq '.inventory.runtimes[]? | select(.governed | not)'Checked by the walk: this command ran as written on a no-root Linux install.
3. See them on the dashboard
The dashboard's Overview has one line, “Other AI on this machine”, with the count and a link to the list. The Agents screen has the list in a card of the same name: each runtime or app with its kind, name, version, address or path, listener or code signature, when it was last recorded and its state, unmanaged first. It says “Inventory, not enforcement”: nothing in it is a denial or an alert, and it is kept apart from both. When agentgarmd cannot read its store, the card says “not reported” rather than zero.
agentgarm ui --open4. Read unmanaged and governed
A runtime counts as governed when a governed agent's model call went to it through its proxy route in the last 24 hours. The Cursor app counts as governed when a Cursor task ran in that time, and the Ollama and LM Studio apps when their runtime is governed. The ChatGPT, Claude and Antigravity apps are always unmanaged: their chats pass no AgentGarm control point. Everything here is inventory with no assurance level, and a process running as you can stop the session or pad its reports, so treat the count as a prompt to look, not as evidence.
5. Route an agent's local model calls through the proxy
When an agent is set up for a local runtime, agentgarm run points its base URL at the runtime's route on the proxy instead: OPENAI_BASE_URL=http://localhost:11434/v1 becomes http://127.0.0.1:4747/local-ollama/v1. The routes are local-ollama, local-vllm, local-lm-studio and local-llama-cpp. Only the chat completions, responses and model-list paths are forwarded; a runtime's own API, such as pulling or deleting a model, gets 403. The calls are counted per task and priced at zero, and the runtime then shows as governed.
export OPENAI_BASE_URL=http://localhost:11434/v1
agentgarm run -- <agent>See also
Route other model providers through the proxySend Gemini, Bedrock, Vertex, Azure, OpenRouter, DeepSeek and local-runtime calls through the proxy, accounted, budgeted and checked for secrets.
6. Choose whether model ids are recorded
A list of models can reveal what someone works on, so model ids are not recorded by default. To record them, set inventory.record_models: true in default.yaml, the only policy file this section is read from, then restart agentgarmd; the first report after that still has none. What is always recorded is each runtime's kind, version, port and listening program, and each app's id, version, path and code signature. These records stay on the machine for 30 days.
inventory:
record_models: falseagentgarm policy lint ~/.config/agentgarm/agents.d/default.yamlSee also
Watch your agent for a week, then enforceStart a policy in audit mode, see what enforcing it would have blocked in your own history, fix the rules that get in the way, then switch it to enforce.
7. See the fleet's count in Central Lite
An enrolled machine sends Central Lite its counts and the catalog ids of its unmanaged runtimes and apps, never paths, listeners, code signatures, versions or model ids. The console's Inventory shows each machine's count in its Unmanaged AI column and, in the “Unmanaged AI detected” card, the fleet's total, how many machines reported it and each machine's ids. A machine that sent no inventory is shown as not reported, never as zero.
See also
Enrol developer machinesEnrol a machine in Central Lite with a one-time token that pins Central Lite's key, see it report in the inventory, and let its users enrol presence keys.