Try a gateway for your agents on your own machine.
AgentGuard is a sandbox and MCP gateway for AI agents. Your agents reach tools only through one gateway, which decides each call — allow, hold for a reviewer, or deny — and records every decision in a hash-chained audit log that never leaves your machine.
Everything to evaluate the gateway, nothing to send us.
The trial edition is a complete local setup. You run it, you hold the data.
One MCP gateway
A single endpoint for your agents, with pinned tool definitions, per-caller capabilities and a deny-by-default egress proxy.
A stand-in gate
Deterministic decision logic that allows, holds for a reviewer, or denies each tool call, so you can see the flow work end to end.
A local sandbox
Docker Compose files with a demo agent and demo MCP servers, so you can watch the gateway work before you connect your own agents and tools.
Review for held calls
Calls that change or send something wait for a person. You list, approve or deny them from the command line, and the reviewer’s name goes into the audit log.
Traces in a bundled Grafana
Every gateway decision is a trace span you can browse locally. Spans carry tool names and outcomes, never argument values.
Offline licence check
The trial runs on a signed licence file that expires. Verification is local and nothing phones home.
Every call goes through the gateway.
Agents never talk to your tools directly. The tools sit on a network only the gateway joins.
- Step 1
Identify the caller
Each agent presents a workload credential and a capability. A call outside its capability is refused before it reaches the gate, and the refusal is recorded.
- Step 2
Check the tool
Only tools pinned in your sandbox file are exposed. A tool whose definition has changed since it was pinned is quarantined.
- Step 3
Decide
The gate allows the call, holds it for a reviewer, or denies it. Tiers follow the tool’s name: read-only, mutating, egress or destructive.
- Step 4
Record
Each step is written to a hash-chained audit log on your machine. Traces are a view; the audit log is the record.
What the trial does not do.
Know the limits before you evaluate.
- Interocept’s detection. The trial uses a deterministic stand-in, not a model-based detector. Interocept, the production gate, is licensed separately.
- Production hardening. The sandbox is for evaluation. Containment is best-effort, because an internal Docker network stops routing but is not a container-escape boundary.
- Protection against whoever runs it. Anyone with the source can rebuild the image, so treat the trial as a way to evaluate AgentGuard, not as a control against its operator.
Run it on your own machine.
Tell us who you are and we will issue a signed trial licence for your organisation. You need Docker with Compose v2 and uv.