AI inventory on developer machines Cloud

The AI Inventory page (Activity → AI Inventory) lists every AI coding harness, MCP server and skill bundle on your developers' connected machines, and marks the ones that run outside governance. It covers machines that run intutic connect and nothing else: it does not read an EDR, an identity provider or a cloud account, so an agent on a machine that never ran intutic connect does not appear.

Who sees what

Role	Sees
Owner, Admin, Engineering manager	Every machine in the workspace
Developer	The machines whose latest report was sent with their own credentials
Viewer	Nothing; the page is hidden and the API answers 403

What is collected

intutic connect collects the inventory on the machine and sends it to the control plane with the agent facets it already reports.

Item	Fields
Machine	A fingerprint (a hash of the hostname, platform and machine id, the same one intutic enforce reports), the hostname, the platform, the CLI version, the intutic connect workspace folder relative to your home directory, and the member whose credentials sent the report
Harnesses	Every harness whose detection rule matches the machine or the connected workspace, plus every harness intutic connect is configured for. For each: its type; its version where the install records it in a folder name (the Cline, Roo Code and Continue VS Code extensions); whether intutic connect is configured for it; how its gate works (a hook file, an SDK gate inside the harness, another harness's gate, an external bridge, or no gate); whether the gate file is on disk and where, relative to your home directory; the time of the newest event the gate wrote on this machine; and the time of the newest event it wrote with guards disabled (INTUTIC_GUARD_DISABLE=1)
MCP servers	Every server in the MCP configs Intutic reads for Claude Code, Claude Desktop, Cursor, Cline, Windsurf, OpenHands, Continue, Goose, Muse Code, Grok Build, OpenCode, Gemini CLI and Google Antigravity: its name, the harness whose config declares it, its transport, whether the MCP governance proxy fronts it, and why not when the daemon can tell. A remote server's URL is sent without its user name, password, query string and fragment, and a path segment that looks like a token (16 or more characters mixing letters and digits) is replaced with [redacted]
Skill bundles	Each folder under .agents/skills and .claude/skills in the connected workspace and under ~/.claude/skills: its name, the folder it was found in, the SHA-256 of its SKILL.md, whether the content scan ran and how many findings it had, and how many files are bundled with it
Guard probes	The local proxy's last scheduled guard-probe run: how many probes ran and how many failed, when the proxy answers

What is never collected

• The contents of any file: rules files, SKILL.md, bundled scripts, MCP configs, hook logs.
• Environment variables and their values, including those an MCP config sets for a server.
• An MCP server's command line or arguments, and a remote server's headers.
• Paths outside your home directory; one is sent as [redacted].

The control plane removes credentials from MCP server URLs a second time when it stores them.

How often

Step	When
A machine sends its inventory	When intutic connect starts, then every five minutes (every tenth poll at the default 30-second interval)
A gate's latest event is recorded	When intutic connect drains the hook-event log: when the log changes, and at least once a minute. Cline and n8n gates keep their own log, which is read when the inventory is collected
Status is decided	On every page load and API read, from the latest report and the current time
A machine is shown as not reporting	72 hours after its last report, the same threshold device posture uses, unless it ran intutic disconnect: then it is shown as disconnected

How status is decided

A harness is:

Status	When
Ungoverned	It has no tool-call hook a gate can attach to; or it was detected but is not one of the machine's intutic connect harnesses and no gate file is present; or it is connected and its gate file is missing; or its gate ran with guards disabled in the last 48 hours; or it, or the whole machine, was disconnected with intutic disconnect (reason: disconnected)
Gate stale	Its gate file is present, it has been on the machine for more than an hour, and the gate has written no event in the last 48 hours, the same window as the Gate Stopped Reporting alert
Unverified	Its gate was installed less than an hour ago and has not written an event yet; or its gate runs inside the harness's own process, in another harness or in an external service (LangChain, CrewAI, the OpenAI Agents SDK and the other SDK-gated frameworks, Xirp, Agentic Orchestrator) and no event from it has reached the machine
Governed	Its gate is installed, guards are on, and it wrote an event in the last 48 hours

A gate writes an event for every tool call it sees, allowed calls included, so a harness nobody used for two days shows Gate stale until it is used again.

Google Antigravity and Gemini CLI share one integration but have a gate each, so each product found on a machine is its own row: antigravity and gemini-cli, with its own gate file and last event. A machine where neither is found but the harness is connected shows one antigravity row.

An MCP server is Ungoverned when the MCP governance proxy does not front it, and Governed when it does.

Disconnected machines

A machine that ran intutic disconnect is marked Disconnected in the Machines view until its next report, and its harnesses are ungoverned with the reason disconnected. The Disconnects view lists every disconnect the machines reported, newest first: the machine, what was disconnected (every harness and the login, or the harnesses named with --harness), the member whose credentials reported it, and when. It follows the same roles as the rest of the page: owners, admins and engineering managers see every machine's disconnects, a developer sees the ones their own credentials reported. Each disconnect also streams to SIEM export as device_disconnects.

Notifications

The first time a machine reports an ungoverned harness or MCP server, the control plane sends an Ungoverned AI Tool Found notification (inventory.ungoverned.detected, severity MEDIUM), unless the harness was disconnected on purpose. It is sent once per machine and item: not again when the item disappears and comes back, and not when only the time-based statuses change (a stale gate already raises Gate Stopped Reporting). Add the event to a notification rule to receive it, and SIEM export carries it under governance_alerts.

Export and API

The Harnesses and MCP servers views download as CSV with the filters on screen. A cell that starts with =, +, -, @, a tab or a carriage return is prefixed with ', so a spreadsheet opens it as text.

The CLI reads the same views: intutic inventory summary, intutic inventory devices, intutic inventory harnesses and intutic inventory mcp-servers (both with the page's filters, and --csv for the same download), intutic inventory skills and intutic inventory disconnects. See the CLI reference.

Endpoint	Returns
GET /api/v1/inventory/summary	Machine, harness, MCP server and skill counts, and the governed percentage
GET /api/v1/inventory/devices	One row per machine
GET /api/v1/inventory/harnesses	Harnesses by machine, filtered by status, harness, device and q; format=csv for a download
GET /api/v1/inventory/mcp-servers	MCP servers by machine, with the same filters and CSV
GET /api/v1/inventory/skills	Skill bundles by machine, filtered by device and q
GET /api/v1/inventory/disconnects	The disconnects machines reported, newest first; limit (default 100, at most 500)

Limits

• Inventory is self-reported by the machine, like the device posture under Settings → Security. A member could send a false report for a machine they control.
• Harness detection runs against the connected workspace and the home directory. A framework detected from a project's dependencies (LangChain, Mastra and the like) is found only in the workspace intutic connect runs in.
• A machine that stops running intutic connect keeps its last report, marked as not reporting after 72 hours.