Skip to main content
Oso sits between AI agents and the models they talk to. By intercepting that traffic - through a combination of hooks, a browser extension, and a network proxy - Oso can see what agents are doing, who is running them, and what data is flowing in and out.

​
Agents and sessions

An agent is a piece of software that sends requests to an LLM on a user’s behalf - a coding assistant, a chat interface, or an automated workflow tool. Oso tracks agents at two levels:
  • Discovered: Oso knows the agent is installed or in use on a device. You can see which agents are present across your organization, who is using them, and whether they are sanctioned.
  • Monitored: Oso captures the full content of the agent’s activity - prompts, completions, tool calls, and data flow. This requires agent traffic to flow through Oso.
A session is a single continuous interaction between a user and an agent. For a coding agent, this might be a single task. For a chat interface, it maps to a conversation thread. Sessions are the primary unit of investigation in Oso.

​
How Oso captures traffic

Oso uses three integration points to capture agent traffic, each suited to different agent types.

​
Hooks

For terminal and desktop coding agents (Claude Code, Cursor, Github Copilot), Oso installs hooks that intercept requests at the agent process level. These are distributed to endpoints through your MDM.

​
Edge proxy

Agents that support custom LLM base URLs can be configured to route traffic through agents.osohq.cloud. The proxy forwards requests to the upstream model and records everything that passes through. This is the setup path covered in the Quickstart.

​
Browser extension

For web-based agents (Claude.ai, ChatGPT, Gemini), the Oso browser extension captures session content directly from the browser. It also serves as a discovery signal for browser-based agents that are not yet monitored.

​
Identity

Oso ties sessions to users and devices. When an agent session is captured, Oso records:
  • User: the identity of the person running the agent, resolved from device context or authentication
  • Device: the machine the agent is running on, correlated with your EDR inventory
  • Agent: which agent software generated the session
This means every session in Oso is attributable - you can answer not just “what happened” but “who ran it, on which machine.”

​
What Oso records

For each monitored session, Oso captures:
  • Every prompt sent to the model and every completion returned
  • Each tool call the agent made, including the tool name, parameters, and response - this covers both MCP server calls and CLI commands, so nothing slips through regardless of how the agent is invoking tools
  • The MCP servers and integrations the agent used during the session
  • Timestamps for the full sequence of exchanges
This data is retained according to your data privacy settings.