The Platform
One policy, in every tool's own configuration, on every machine
Sielum writes your rules into the managed configuration each AI coding tool already supports, on every endpoint, then reads the machine back and shows you where reality has drifted from what you declared. Including on the tools no admin console reaches.
Platform Capabilities
Everything you need to govern AI agents
See every AI agent on every machine
Sielum's lightweight endpoint agent scans running processes and config files to build a live inventory of all AI tools in your environment, without proxying traffic or touching file contents.
Agent Inventory
Detects Claude Code, Cursor, Copilot, Codex, Amazon Q, ChatGPT Desktop, Gemini CLI, Windsurf, and OpenCode, automatically, across every enrolled endpoint.
MCP Server Inventory
Lists all configured MCP servers per endpoint: name, command, enabled/disabled state. Know exactly which extensions your agents have access to.
API Connection Tracking
Network-level monitoring of which external domains each agent contacts. No proxy, no MITM. Tracking happens at the OS network layer.
Define and enforce how agents operate
Set boundaries for which agents are allowed, which APIs they may contact, and which MCP servers are permitted. Sielum writes them into each tool's own configuration and checks every machine, so a boundary you set is the boundary in force, not a suggestion left to the agent.
Policy Engine
Five built-in policy types covering unknown agents, unauthorized API connections, MCP filesystem access, new MCP servers, and endpoint offline. Every breach fires an alert automatically.
Managed policy enforcement
Full config enforcement on Claude Code, partial on Cursor, and policy alerts on every other tool from day one, with enforcement depth expanding across the fleet.
Network Layer Blocking
Blocking is an OS level rule and applies to the whole machine, so use it for domains you want no traffic to at all, and use the configuration layer for everything you want scoped to the agent.
Export audit evidence for every compliance framework
Sielum writes every policy event, enrollment, and config change to an append-only audit log. Export structured evidence for SOC 2, EU AI Act, and GDPR on demand, with no manual data assembly.
Append-Only Audit Log
Every policy event, enrollment, and configuration change is written to an append-only log, exportable as JSON/CSV, queryable by time range or endpoint.
Compliance Reports
Pre-built report templates for SOC 2 and EU AI Act, generated from live audit data, ready for auditor review without manual compilation.
GDPR Art. 17 Controls
Process data deletion requests from the dashboard. Full endpoint data purge on demand, covering agent events, config snapshots, and enrollment records.
Detect & Intervene
Catch unsafe AI usage before it becomes an incident
Visibility is the start. Sielum also detects when AI safety guardrails get switched off, and lets you require a human decision before risky agent actions run.
Detect when guardrails get switched off
A single flag disables every permission check in an AI coding tool. Sielum logs every attempt across your fleet, and remediates the dangerous ones automatically.
- High-severity alert with endpoint, process, PID, and the exact flag used
- Cursor configs auto-remediated to safe values, atomically, no manual fix
- Config Guard blocks bypass flags on Claude Code, even when passed on the CLI
- Live 24-hour unsafe-event counter on the AI Safety dashboard
Require sign-off before risky actions run
Route sensitive agent actions through a human. The agent pauses and waits for an admin to approve or reject, right from the dashboard, with a full audit trail.
Agent requests approval
MCP server access · production
Admin approves with a note
audit-logged decision
- Agent pauses on a policy match and waits for a human decision
- Approve or reject from the dashboard, optionally with a note
- Every decision is written to the append-only audit log
- Auto-rejects after a configurable timeout, no agents stuck waiting
Delimitation
Where Sielum fits, and where it does not compete
EDR and DLP were not built to understand AI agent sessions. The admin consoles from Anthropic, GitHub and Cursor now govern their own tool well. Sielum sits in the space neither reaches: every tool a developer actually runs, in one policy, checked independently, with the data staying in the EU.
Why not just EDR or DLP?
Endpoint protection and data loss prevention operate on files and network flows. Neither has a concept of an AI agent session, an MCP server registration, or a managed settings file. They can see that a process opened a connection, not that a coding agent just requested access to a new tool. Governing AI activity needs a lens built for AI activity.
Why not just the vendor's own admin console?
Anthropic, GitHub and Cursor each ship real controls for their own tool now, cloud pushed, no agent required, refreshed automatically, and for that one tool they are often stronger than anything built on top of them. Use them. The gap opens the moment a developer runs a second tool, signs in with a personal account, or uses something no console was ever told about, because every vendor console only governs its own tenant.
| Capability | EDR / DLP | Vendor console | Sielum |
|---|---|---|---|
| Sees personal accounts, browser AI and unmanaged installs | No | Nosees only its own enrolled org | Yes |
| One policy across every AI tool your teams actually run | No | Noeach console covers only its own tool | Yesauthored once across every tool your teams run, written directly into the tool's own configuration where we support it, and enforced by policy everywhere else |
| Independent check that a guardrail is actually active on the machine | No | Noreports on itself only | Yes |
| Independent, append-only history of every endpoint's AI tool configuration | No | Nothe record lives with the vendor being governed | Yesconfig snapshots and history, held independently of every vendor |
| Governance data hosted in the EU | Partialdepends on the vendor you buy | NoAnthropic, OpenAI and GitHub consoles run on US cloud | YesHetzner, Germany |
Scope
What Sielum's agent does not see
Sielum's own endpoint agent watches which AI coding tools run and where they connect. It never reads what a developer writes or what an AI tool returns.
Does Sielum's agent read my code or AI prompts?
No. Sielum's agent monitors processes and network connections only. It records which AI tools are running and which APIs they connect to. It never reads file contents, prompt text, or AI responses.
What does Sielum's agent monitor? Is it privacy-invasive?
Sielum's agent monitors which AI tools are running (process names, PIDs), what network connections they open, and their configuration files, for example whether unsafe modes are enabled. It does not read your code, record keystrokes, take screenshots, or access file contents.
Which operating systems are supported?
Linux (kernel 4.15+), macOS (12 Monterey and later), and Windows (10/11 and Server 2019+). All three platforms are supported on x86-64 and ARM64.
Architecture
Secure by design
Your data stays in the EU. End-to-end encryption on every connection, mutual TLS for every agent. Security built in at every layer.
Lightweight endpoint agent
A small Go binary deployed per host. Monitors AI tool processes at the OS level, with no code changes required in agent applications. Direct process inspection, not a proxy.
Mutual TLS enrollment
Each agent authenticates via mTLS during enrollment. Device certificates are issued automatically, with no shared secrets and no manual key distribution. Revocation is immediate.
Encrypted event stream
Events flow over gRPC with end-to-end encryption. The server receives structured event data and applies policy evaluations as each event arrives.
Central control plane
The Sielum server aggregates events, evaluates policies, and stores audit logs. Deployed on Hetzner Cloud DE by default, with data residency in the region you require.
Who Uses Sielum
Built for the teams responsible for AI
Enforce guardrails across every AI deployment
- Keep agents inside the rules you set, enforced in each tool's own configuration on every machine
- Alert on policy violations: unknown agents, new MCP servers, unauthorized APIs
- Maintain append-only audit trails for incident response
Operationalize AI agents at scale
- One policy across every AI tool, with enforcement coverage growing tool by tool
- Monitor which agents and MCP servers are active across all machines
- Manage enrollment, certificates, and policy assignments centrally
Make AI decisions explainable and auditable
- Export structured audit logs for SOC 2 and EU AI Act reviews
- Prove which policy rules were in force on each endpoint, and when
- Process GDPR Art. 17 deletion requests from the dashboard
Documentation
Technical Documentation
Full documentation is hosted in our Docusaurus portal: installation guides, API references, architecture diagrams, and operations runbooks.
Open Full DocumentationPowered by Docusaurus 3: full-text search, versioned docs, MDX support
Need implementation support?
Our team provides hands-on onboarding and PoC support.