Deployment options
In the standalone and cloud options, AgentShield runs on the machine where the agent runs. The options differ in where the decision is made and where the record goes.
| Option | Best for | What runs where |
|---|---|---|
| AgentShield standalone | An individual, or a team that only needs local protection | The gate runs locally with the embedded community rules and your own YAML. The audit log is a local file at ~/.agentshield/audit.jsonl. No account, no network. |
| AI Agent Lens Cloud (recommended) | Teams and organisations that need one policy, one audit trail and compliance evidence | The gate still runs locally. Decisions stream to your organisation at app.aiagentlens.com. Policy, premium rule packs and enforcement mode are pushed down to every agent. |
| Shield server | Production agents, CI runners and MCP servers you do not control machine-by-machine | A long-running AgentShield process fronts the agent or MCP server and applies the same policy over the network. See Shield server and Agents in CI. |
| Enterprise gateway (pilot) | Organisations that want a control point they own, whatever is installed on each laptop | Shield sits between your coding agents and their model provider, either as a guardrail in your LLM gateway or behind your VPN with TLS inspection. Each option has a reference deployment we lab-tested with Claude Code against a mock of the Anthropic API (LiteLLM; WireGuard with mitmproxy). Other components, and other agents and providers, fit the same design once they meet the published requirements, but are not yet tested, and neither is the live API. Each tool call the model proposes is checked before the agent receives it, and in enforce mode a blocked call never reaches the agent. Observe mode records instead of blocking, for a gentle start. Records stay on the gateway in the pilot, written best-effort. See Enterprise gateway. |
What never leaves the machine
With AgentShield standalone or connected to the cloud, evaluation is local: AgentShield makes no model calls and no network calls to decide. With a Shield server or the enterprise gateway, evaluation happens on a host you run; the gateway also sees each request and response between the agent and the provider, prompts and files included, because it sits on that path. It inspects only the tool calls in responses; requests pass through uninspected. In every option, AgentShield never calls a model to make a decision. When connected to the cloud, the audit event carries the decision, the rule, the reason and the command; commands are redacted in the dashboard by default. See Security.
Precedence of settings
Local configuration always wins over anything pushed from the cloud, so an engineer can never be put into a mode they did not set on their own machine without seeing it in agentshield.yaml:
CLI flag → ~/.agentshield/agentshield.yaml → policy.yaml → organisation policy → built-in defaults