Purpose-built endpoint security and governance for AI.
Much of the AI risk in a company is homegrown. Developers and business teams are building their own agents and AI tooling, and those agents run locally. When blocking isn't enough, Ospiri adds control.
Each layer sees something different. Only one sees the agent act.
Model guardrails
AI gateways and API security
Ospiri
EDR / XDR
Operating system
Monitor and observe first. Then enforce. Get alerted on anomalies.
Most teams watch first, learn what normal looks like for each agent, then switch on enforcement one agent and one path at a time.
Know what is running and what it did.
Inventory
AI agents, AI apps, local models, note-takers, skills and MCP servers, running or dormant, per endpoint, per identity group and per account. Personal accounts on managed devices are flagged.
Attribution
Actions attributed at the process level to the agent that took them, with a behavioral baseline for each agent so anomalies stand out.
Skills review
Skill files and downloaded agent code checked for malware, embedded secrets and hidden instructions before they execute.
Limit what each agent can reach.
The safe copy
When an agent tries to change something outside its limits, the change goes to a safe copy instead of the real file. The agent keeps working and the original is untouched.
MCP control
MCP connections inventoried and monitored, scored for reputation, with unapproved servers blocked.
Network limits
Limit which hosts each agent can reach. Unapproved tunnels and destinations are held or blocked.
Decide, in line, what happens next.
Block
Unapproved AI apps and unreviewed skills stopped before they execute.
Hold for approval
Risky actions pause until a person approves them, and the record shows who said yes.
Protect data
Label what matters (payroll, source code, customer records) and keep agents out of it at the endpoint.
Reconstruct
A timeline of an agent's reads, writes, spawned processes and connections, built from what happened rather than inferred.
An agent writes outside its repository.
09:18:39.912 agent claude-code session 7f2 cwd ~/src/payments/app 09:18:40.104 agent claude-code write ~/src/payments/config.ts outside allowed path 09:18:40.105 policy payments-dev → writes outside app/ go to a safe copy 09:18:40.106 safe copy created config.ts → safe/7f2/config.ts 09:18:40.131 agent claude-code write completed in safe copy session continues 09:18:40.132 notify #sec-eng · original file untouched
The questions your endpoint team asks first, answered here.
Start with a one-week assessment that shows your AI inventory.
- Footprint
- A lightweight agent. Measured CPU, memory and boot-time figures are shared during evaluation.
- Deployment
- Direct install, or one token for Intune, Group Policy or your RMM tool. Silent install.
- EDR coexistence
- Designed to run alongside your EDR.
- Uninstall
- Removed automatically on a preset timer, or manually with one command.
- What leaves the device
- Telemetry and logs, with metadata only where policy needs it. No raw company data.
Built to collect what policy needs, and nothing more.
Deployment options for regulated industries are available on request.
Collected
- Agent and AI app process events
- Files, paths and hosts an agent touched
- Skills and MCP servers loaded
- Policy decisions and their outcome
Not collected
- Keystrokes
- Screen content
- Prompt or response text
- Personal-account credentials
What each capability gives your auditor.
| Capability | The control it evidences | Framework reference |
|---|---|---|
| Inventory | A current inventory of AI systems and agents in the environment | NIST AI RMF (Map) |
| Attribution | Actions taken by an AI system are distinguishable from the user it runs as | Not yet in any framework |
| Activity log | Automatic recording of events over the system's lifetime | EU AI Act Art. 12, applying from 2 Dec 20271 |
| Safe copy and limits | AI systems operate under least privilege, with enforced boundaries | Mapping available on request |
| Skills review | Third-party AI components are reviewed before execution | Mapping available on request |
| Reconstruction | AI-involved incidents can be investigated and reconstructed | Mapping available on request |
The questions we hear first.
What is an agent firewall?
A runtime control point for AI agents. Like a network firewall, it enforces policy in line and blocks or contains anything unauthorized, without breaking the agent's workflow. Ospiri sits on the endpoint, where the agents run.
What problem does it solve?
AI agents now run on employee laptops with the employee's access: files, credentials and internal systems. Ospiri lets you approve them with limits, so you can say yes to more agents without giving each one the run of the machine.
What does it do?
It finds the AI agents, apps, skills and MCP servers on your endpoints, sets what each one may read, write and reach, enforces that before the action completes, and flags behavior that's out of pattern.
How does it deploy?
Most teams start with a one-week assessment: a lightweight scanner lists the agents, skills and MCP servers on the endpoints they choose, and we review the findings together. The endpoint component then rolls out through your MDM or RMM tool, one group at a time.
We already run CrowdStrike. Why add this?
Keep it. Ospiri adds limits per agent, the safe copy for out-of-bounds changes, and a policy that stays the same if you ever change EDR.
Does it monitor employees?
Ospiri watches AI agents and AI apps, not people. It does not collect keystrokes, screen content, or prompt and response text.
See it on one of your own agents.
Bring a coding agent your developers already use. We'll take it from monitor and observe to enforce.
- EU Digital Omnibus on AI, Council approval 29 Jun 2026: Annex III high-risk obligations, including Article 12 logging, deferred from Aug 2026 to 2 Dec 2027.