Agent
An AI agent is software that uses AI to carry out tasks on a user’s behalf, often using tools and making decisions about how to complete the work. In Switch Trust, an agent is a tracked record discovered by a scan or reported by a connected data source. Each record may represent an individual agent or a group, depending on the source. For example, the Claude Code plugin creates a separate agent for each person who installs it, while Switch Rooms groups all rooms on a Switch server into a single agent named Switch Rooms Agents.Agent evaluation
An agent evaluation tests your agent by sending prompts to its endpoint and scoring the responses. The results combine into the agent’s overall evaluation health. Evaluations can test the agent with adversarial prompts, check the quality of its answers, or score it with a judge prompt you write. Each result is either a Pass or Fail, or a score from 0 to 100. An agent evaluation is different from a model evaluation, which tests the model rather than your agent. See Evaluate an agent.API key
An API key lets a tool or script connect to your Switch Trust workspace. Each key has a role that limits what it can do. An Ingest only key can send data and read guardrail policies, but can’t read your data. That’s enough for the SDK, a data source, or the Claude Code plugin. Some setup steps call an API key a token. See API keys.Asset
An asset is something an agent uses: a model, an MCP server, or a tool. Agents aren’t assets. Some labels in the product, such as Assets analyzed, count agents as well. See Assets.Burn rate
Burn rate is an agent’s average daily spend this month: its spend so far divided by the day of the month, in UTC. For an agent that started reporting mid-month, burn rate reads low for the rest of that first month. See Spending and caps.Cap
A cap is a monthly spending limit you set on an agent. Switch Trust forecasts when the agent will reach it, and marks the agent Over cap once spending passes it. A cap doesn’t stop the agent from spending, and it doesn’t send an alert. See Spending and caps.Claude Code plugin
The Claude Code plugin reports each person’s Claude Code sessions and cost to Switch Trust, with no code changes. The plugin only reports activity. It doesn’t check or block anything. See Claude Code plugin.Data source
A data source is a way to bring agents into Switch Trust. Data sources include code repositories and cloud accounts to scan, OpenTelemetry, the Claude Code plugin, and Switch servers. See Data sources.Detector
A detector is a single check in a guardrail policy, such as a check for jailbreaks, toxicity, or credit card numbers. Each detector checks the side of the conversation its detector role sets, and takes its guardrail action when it finds a match. See Guardrail policies.Detector role
A detector role sets which side of a conversation a detector checks:- User: Checks what goes into the agent, such as prompts.
- Assistant: Checks what comes out, such as the model’s replies.
Discovery
Discovery scans your code repositories and cloud accounts for AI agents and the assets they use, and adds what it finds to your inventory. A repository scan finds agents in Python code. In JavaScript and TypeScript, it recognizes only the Anthropic, OpenAI, Claude Agent, and MCP SDKs, and finds agents only from the Claude Agent SDK. Each repository needs the scan workflow added, and an API key from your own LLM provider. Scans don’t count toward managed agents. See Find your agents.Guardrail
A guardrail checks an agent’s messages or requests against a guardrail policy and takes action when it detects a violation. Guardrails apply to traffic routed through the Switch Trust guardrails gateway, including traffic from agents using the SDK, and to Switch Rooms messages. OpenTelemetry and the Claude Code plugin only report activity to Switch Trust; guardrails do not apply to traffic reported through these sources.Guardrail action
A guardrail action determines what happens when a detector finds content that matches its criteria:- Alert: Allows the content through and records the match.
- Redact: Replaces the matched text before allowing the content through.
- Block: Prevents the content from passing through.

