Every AI tool, agent and connection, across devices, cloud and applications, in one accountable system.
Discover what runs across your company, connect it to people, access and spend, and enforce supported policies where AI is used.
Start with devices. Add cloud services and your own applications to the same inventory and policy model.
Local evidence and supported endpoint controls.
Read-only AI discovery across connected accounts.
Evidence from the software your team builds.
Discovery support does not imply identical enforcement support. Confirm the tool, operating system and interaction—especially application launch prevention or removal. Removing an installed application requires an administrator.
Three capabilities of one platform, built on the same evidence: Surface, Monitor and Govern. Explore the views your teams use to discover, investigate and act.
| AI asset | Account context | Connection | Review |
|---|---|---|---|
| CursorCoding agent | Work account | Corporate repository | Approved |
| Claude DesktopDesktop AI | Personal account | MCP configured | ! Needs review |
| OllamaLocal runtime | Not applicable | Local models | Undecided |
No named approval is recorded for the configured operation.
Find the unregistered. Discover supported tools without a self-reported inventory.
Connect the context. Join tools, devices, accounts, projects and usage.
Keep the reason. Every verdict stays tied to its evidence and policy.
Discover supported tools, agents and models alongside the accounts, credentials and device posture that change their risk.
Connect a tool to its account type, declared MCP access, repository context and security posture. Discovery reports that a key exists; the secret value is discarded.
Installed AI tools, CLIs and coding agents.
ai_toolingConfigured autonomy and pre-approved operation.
ai_autonomyKnown AI-domain visits, by category and count.
ai_web_usageThe wider application inventory on the device.
installed_softwareDeclared server transport, host and capability.
mcp_serversProvider-key presence, variable name and git-ignore context; no secret value.
ai_credentialsCorporate, personal or unknown; organisation and email-domain context.
ai_accountsToken and usage records already written by supported tools.
ai_usageRepository owner/name and a bounded inventory of data files; no source-code content.
ai_projectsDevice make, model, operating system and version.
device_factsWhether the device is enrolled in device management.
mdm_enrollmentWhether OS updates are configured to install automatically.
auto_updatesHow far the operating system is behind on patches.
os_patch_levelWhether the volume is encrypted.
disk_encryptionWhether the screen locks and the configured timeout.
screen_lockWhether the host firewall is enabled.
firewallWhether an endpoint detection and response agent is present.
edr_presentDevice network configuration.
network_postureVerify configured policy without modifying it.
browser_policyWhether password policy is enforced.
password_policyWhich device accounts have administrator rights.
admin_accountsClaude Code · Claude Desktop · Codex · ChatGPT Desktop · Cursor · Windsurf · GitHub Copilot · Aider · Goose · Gemini CLI · Amp · Cline · Continue · OpenCode · OpenClaw · Ollama · LM Studio
Signal availability varies by operating system and tool. Declared connections describe configuration, not proof that every resource was accessed.
Follow available usage and cost back to tools, models, employees, projects and devices. Investigate a change before deciding what it means.
See cost by tool, model, person, project and machine where source records support the attribution.
Unpriced models and unavailable records should remain visible, rather than disappearing into a misleading total.
Investigate configured unattended operation, personal versus work accounts and declared repository or MCP access.
Use the evidence to examine possible waste. Assess ROI alongside time saved, work completed and other outcomes.
Token-cost attribution is source-dependent. It is not automatically a full subscription bill, employee productivity score or automatic ROI calculation.
Apply versioned company policy. Use local controls on supported interactions, and give the resulting findings an owner and a traceable record.
The supported request is stopped according to policy. The decision remains connected to the rule.
Version the policy. An edit publishes a new version and re-evaluates the fleet without rewriting old verdicts.
Own the finding. Assign an owner, due date and status; track it through remediation and review.
Choose the right action. Blocking an interaction is different from preventing launch or uninstalling an application.
Framework mappings support assessment and evidence work. They do not certify your organisation or guarantee compliance.
Evidence, policy version, verdict and owner remain connected. Review what was true at the time, even after a policy changes.
Device and repository context enter the evidence record.
The finding points to the exact policy version used.
Assignment, state and resolution remain reviewable.
Export evidence with its decision history and policy reference.
Store verdict transitions and immutable policy versions. Changing a rule does not rewrite the past.
Nightly evidence export to object storage, configurable retention and deletion workflows support longer-term records.
Carry verdicts and policy versions into chat, ticketing, SIEM and warehouse destinations.
Missing evidence remains unknown. Missing policy remains ungraded. Neither is silently treated as a pass.
Choose your connection and data boundary. Roll out through existing fleet processes, with reporting status visible along the way.
Install the agent, claim the device in your console and review the first inventory. Use the pilot to confirm the signals and controls your team needs.
Start with the hosted platform and connect the sources you need.
Run the console, API and worker with your database and object storage. Discuss residency, controlled updates and air-gapped deployment with the team.
Discuss self-hosting ↗Understand what discovery collects, what it excludes and which additional collection settings require an explicit choice.
Credential values are discarded on the device before discovery evidence is stored.
Identify repository owner/name and bounded file context without collecting source-code contents.
Prompts, responses and transcripts require explicit administrator enablement; review the scope before rollout.
Inventory does not execute discovered tools. Collection excludes precise location and personal account IDs.
Tenant-scoped access connects devices, accounts, tools and findings within your organisation.
Review the published signals, privacy boundaries and deployment documentation with your security team.
Surface is a separate open-source local inspection tool. The enterprise platform adds connected inventory, policy, workflows and evidence.
Everything the Guardian Agent Platform does, grouped by capability. Open a group or search for the feature your team needs.
Supported AI tools, coding agents, desktop apps, browser AI, local models and Model Context Protocol (MCP) servers.
Account type, credential presence, declared repository access and project context are connected to devices.
Inspect devices, tools, MCP servers, accounts, credentials and connected cloud APIs. Filter and export inventory.
Seven signal groups combine AI use with device facts and security posture. Individual collection support varies by environment.
Distinguish sanctioned, undecided and unsanctioned tools; inspect configured unattended operation and pre-approved access.
Connect available token and cost records to tool, model, project, person and machine; unpriced models remain visible.
Reporting, deliberately paused and silent devices are distinct. Track changes, events, findings and items needing attention.
Use consumption evidence in waste investigations and ROI reviews alongside business outcomes. Usage alone does not establish waste or productivity.
Immutable policy versions and fleet re-evaluation keep each verdict attached to the rule that produced it. Rules include allowed MCP hosts and reporting cadence.
Allow, warn, redact, block or require approval on supported AI interactions. Guardrails act locally; an administrator approves app removal.
Published rule packages support SOC 2, the EU AI Act, ISO/IEC 42001, NIST AI RMF and GDPR reviews. Mappings do not confer certification or guarantee compliance.
Assign an owner, due date and state; retain the decisions and resolution history behind each finding.
Preserve policy versions and verdict transitions. Missing evidence stays unknown, and a missing policy stays ungraded.
Nightly evidence projection to object storage, configurable retention and deletion workflows keep records manageable.
Device, tool, agent, MCP server, account, credential, repository, usage, policy, finding, control and event context share one model.
Authenticated, idempotent ingestion, deduplication and redaction before storage. Device evidence is saved locally before delivery and queued through interruptions.
Tenant-scoped data access and live console updates keep the organisation’s evidence connected and current.
Read-only discovery of supported cloud AI inventory, exposure, safety configuration, usage and cost.
Python, TypeScript, Rust and Go ingest model calls, agent runs, tool calls and MCP calls. Prompt/response content is excluded unless explicitly enabled.
Send verdicts and their policy versions into chat, ticketing, SIEM and warehouse workflows. Confirm the destination you need during evaluation.
Home, Surfaces, Monitoring, Events and Governance, with Instruments, Integrations and Settings for inputs, outputs and administration.
Claim-code enrolment, native packages, fleet secrets, managed configuration and tools such as Intune, Jamf and SCCM.
Per-device credentials, rotation, controlled release versions and a deliberate pause for rollout management.
Use the hosted console or arrange deployment on your infrastructure. Self-hosted options include controlled updates and air-gapped environments.
Secret values are discarded on the endpoint, source-code contents are not collected, and discovered tools are not executed for inventory.
Prompt, response and transcript collection is off by default and requires an explicit administrator choice. Collection excludes precise location and personal account IDs.
Surface is a separate open-source tool for inspecting local AI signals before evaluating the enterprise platform.
No matching capabilities. Try a broader term or ask the team.
See the Guardian Agent Platform on your own fleet. Find every AI tool, understand what it reaches and costs, and apply the controls your team needs.